The Voice-Acting Actor: Instrument, Tools, Ears, and the Road There

A self-contained technical report on the MOSS voice-acting line, from the first prototype on MOSS-TTS-Local-Transformer v1.5 to the directing agent of September 2026 — told along the project leader's actor metaphor: the Instrument (the speech model), the Tools (adapters, guidance, prompt forms, steering), the Ears (learned scorers) and the Mind (the directing LLM) — with every failure, retraction and measurement error kept.
LAION · Christoph Schuhmann and collaborators · rendered 2026-09-09 19:20 by $SC/code/vareport/en_build.py. No external resources.

How to read this report

Every number below was measured, and the measurement is named — the source file, the protocol section, or the earlier report it is transcribed from, with its n, its effect size and its uncertainty wherever the source recorded them; where a source did not record one, the text says so rather than inventing it. Numbers in tables that come from files on disk are computed when this page is rendered. Where a result is negative it is reported as a result. Where two sources disagree, the disagreement is recorded (Part III.5), not smoothed over. Numbers use English notation: a point for decimals, a comma for thousands. German protocol quotations are translated.

The metaphor. The project leader describes the system as an actor. The Actor's Mind is the directing language model that reads the scene and decides what to say and how. The Instrument is the speech model it plays. The Tools are what it can do to the instrument at the moment of speaking — adapters, guidance, prompt forms, steering vectors. The Ears are the learned scorers it listens with, and Selection is how it picks a take. Part I is the biography of the instrument, the tools and the ears before the actor existed. Part II is the actor. Part III is the journey — everything that went wrong, in order, and what it changed. Part IV is what is public and what is open, with the human evaluation first, because it is the one measurement that has never been made.

Part I — The instrument and how it was built

In the actor metaphor that organises this report, the Instrument is the speech model itself — the thing that produces sound when the actor's mind decides what to say and how. Part I is the biography of that instrument: how a general-purpose text-to-speech base model became, over two full fine-tunes, three supervised rounds on a timed-script format, seven preference runs, five reinforcement-learning runs and roughly a thousand adapters, the model the actor of Part II plays. Part I also builds the Tools (adapters, doses, prompt forms) and the Ears (the learned scorers), because both were made before the actor existed and both were shaped by the failures recorded here. Sources: [TR-0829] §1–§17 and §24 are the backbone; [TR-0826] and [REF-V1] supply what later versions dropped; the protocol supplies the details no report carried.

I.1 What the system is

MOSS Voice-Acting is a text-to-speech model that is directed rather than configured. Instead of exposing a handful of knobs — speed, pitch, an “emotion” enum — it takes a written stage direction in natural language, of the kind a director gives a voice actor, and performs the line accordingly. It speaks English and German, outputs 48 kHz audio, can clone a voice from a reference recording, and can produce non-verbal vocal bursts — sighs, gasps, chuckles, groans — inside a spoken line. Source: [TR-0829] §1.

The production checkpoint at the time of the 29 August report was laion/moss-tts-local-transformer-4.55b-voice-acting-v2, 4.55 billion parameters; the checkpoint the actor plays today is its third supervised successor, SFT3 (…-voice-acting-v2-sft3, Part II). Neither appeared in one step. Section I.3 reconstructs the road: an off-the-shelf OpenMOSS base model, a full fine-tune on distilled voice-acting data plus a rank-256 adapter merged into it (“v1”), then a second six-dataset full fine-tune selected on generated audio rather than on loss (“v2”), then the timed-script rounds of I.9.

Around that model sits an ecosystem that this report describes end to end:

• a family of LoRA adapters (small trainable weight deltas) for emotions, character archetypes, voice identities, timbre dimensions, vocal bursts and register — roughly a thousand of them across the project (I.4);
• a set of learned scoring models used as measurement instruments — they decide which of many generated takes is kept, and they are how every claim in this report is checked (I.6, and the Ears of Part II);
• a corpus of about 3.1 million annotated utterances, roughly 38 % of them generated by the model itself and filtered, the rest real recorded speech (I.8.4, I.15);
• the timed-script prompt format plus supervised fine-tuning, preference tuning and reinforcement learning aimed at one specific capability — making the model perform an emotion at a requested intensity (I.8–I.13);
• and, since early September, the directing agent of Part II, which plays this instrument with the tools and listens with the ears.

The short version of the result, as of 29 August ([TR-0829] §1). Timing control works: ask for a sentence that takes 3.4 seconds followed by a 0.6-second pause and a 0.3-second sigh, and you get it — the median error on the total clip length is 0.08 s and essentially every clip lands within 0.5 s. Intensity control does not work yet: ask for an emotion in the 90th-to-98th percentile of how strongly that emotion appears in the corpus, and the model delivers something around the 34th percentile. Four training objectives were aimed at that gap. Three things moved it, all by less than a hundredth: data — a small adapter trained on the most intense and most genuine 8 % of the corpus (I.12.3); overdriving that adapter past its trained strength (I.12.4); and a better-constructed contrast in the preference corpus, which as of 28 August produced the best model in the project and was the first preference tuning here that raised intensity without costing intelligibility (I.13.4). What happened after 29 August — the agent, the shipped adapter stack, Best-of-N with new ears, the low-α steering result that superseded the negative one — is Part II.

I.2 Terms, defined once

I.3 The base model and the two full fine-tunes: v1.5 → v1 → v2

The production model was not trained from scratch and did not arrive in one step. It is the end of a four-stage chain, and every stage constrains what the later ones could do. The emotional-intensity ceiling that the later sections keep running into is partly inherited from it. Source: [TR-0829] §3.

1. OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 — a general-purpose autoregressive TTS base model, 12-codebook RVQ at 48 kHz.
2. Voice-acting v1 — a full-parameter fine-tune on distilled voice-acting data, plus a rank-256 LoRA trained on 1 M samples and merged in. Public as laion/moss-tts-local-transformer-4.55b-voice-acting.
3. Voice-acting v2 — a second full fine-tune over six datasets (185,503 rows), the checkpoint selected on generated audio quality, not on loss. Public as …-voice-acting-v2. This is the model every adapter in I.4 is attached to.
4. The timed-script round (I.8–I.13) — three more supervised rounds on a new prompt format, then preference tuning and RL on top of that; SFT3 is the instrument of Part II.

I.3.1 The base checkpoint and its representation

MOSS-TTS-Local-Transformer-v1.5 is an autoregressive local-transformer TTS model: a large semantic backbone predicts one hidden state per audio frame, and a small second transformer expands that hidden state into the codec tokens for the frame.

4.13 B of the 4.55 B parameters are trainable; the remainder is the frozen codec-side embedding. The audio tokenizer is MOSS-Audio-Tokenizer-v2: a residual vector-quantised codec with 12 codebooks of 1,024 entries at 12.5 frames per second, decoding to native 48 kHz audio. Two consequences run through everything below:

• One frame is exactly 80 ms. A clip of duration d seconds is floor(d × 12.5) frames. This is the unit in which length is requested and in which length error is measured — and the unit in which the layer forensics of Part II tap the residual stream.
• Each generated position carries 13 channels (12 codebooks plus a binary stop head), so the model's negative log-likelihood sits at roughly 54 nats per position. Any loss weight that has to be commensurate with it — a DPO anchor, for instance — must be scaled accordingly; getting that wrong once turned a preference run into a plain supervised pass (I.10.1).

The base model's prompt surface, which v1 and v2 both inherit, is two fields: instruction carrying a GENERAL: voice description and a SCRIPT: block with inline parenthesised delivery cues and [pause] markers, and text carrying the same spoken words plain. Recommended sampling with a reference clip: temperature 0.8–0.9, top-p 0.9, top-k 40–50.

I.3.2 Stage one — voice-acting v1

Two things happened here, and they are often conflated because the public artefact is their merge.

(a) A full-parameter fine-tune of v1.5 on DramaBox-style voice-acting data. “DramaBox-style” means the data is formatted as a structured director's instruction plus the spoken line, and it was filtered for vocal bursts and strong emotion. The checkpoint, with its checkpoint-cont-* continuations, lives in the private repository TTS-AGI/moss-dramabox-ft.

(b) A rank-256 voice-acting LoRA on top of that, α = 256, over all attention and MLP projections, trained on 1 million samples (the adapter repository's root is the 100 k checkpoint; checkpoint-1m/ is the full run), then merged into the weights. Training: 8×GPU DDP, bf16, gradient checkpointing, global batch 128, warmup 0.03 with linear decay.

The public release laion/moss-tts-local-transformer-4.55b-voice-acting is that merged model — “the voice-acting LoRA is already folded into the base fine-tune”, in the card's own words. Its card names four data sources, and this mix is the first place where the balance between synthetic and real material is set:

(c) A vocal-burst LoRA. Rank 256 again, trained on a 100 k inline-burst dataset and resumed from the 1 M voice-acting adapter rather than started fresh. Its training data (TTS-AGI/moss-inline-vocal-bursts, private) writes the bursts into both the instruction and the text, which is what teaches the model that a burst tag corresponds to an event in the audio. The merged result was called merged_inline, and a further all-parameter fine-tune was run from it. A separate library of per-class burst adapters — chuckle, breathy giggle, growl, fearful gasp, contented sigh, and so on — is published as laion/vocal-burst-lora-adapters.

I.3.3 Stage two — voice-acting v2, the six-dataset full fine-tune

This is the production model of the pre-agent era. It started on 2026-07-22 and ran roughly 14–15 hours on 8×A100-80GB — not on JUPITER. Its training scripts lived on an ephemeral RAM disk that has since been wiped; the surviving record is the run notes, the published repositories and the live results pages, which is why this section is reconstructed rather than quoted from a script ([TR-0829] §3.3).

Where the six datasets come from. The prep_* half is prepared external material: prep_top3_neutral is the top-3 reward-selected takes of a neutral-delivery set, prep_emolia_elise is a cut of the Emolia German emotional-speech corpus, and prep_a1plus is an additional voice-acting collection (published tokenised, without per-clip scores, as TTS-AGI/additional-data-a1-plus, 31,306 rows). The our_* half is the model family's own generations, kept by a best-of-N reward: our_top3 the top-3 takes per prompt, our_emolia the Emolia-conditioned generations, our_adult the adult-register set. The Gemini-generated named-character material (TTS-AGI/gemini-adult-voices, 30 named characters × 1,000 clips) sits in the same family of synthetic sources.

I.3.4 Determinism and serving

Randomness lives entirely in the autoregressive token sampling (the torch RNG, fixed by a seed); the codec decoder is deterministic. The same seed therefore gives byte-identical audio despite a non-zero sampling temperature — verified at MAE 0.0 across two seeded generations. That is what makes the paired sweeps of I.12.4 and the paired Best-of-N experiments of Part II possible at all.

Serving throughput was characterised with SGLang-Omni on a JUPITER GH200 (aarch64): 27.7× realtime at concurrency 32 and 31.2× at concurrency 128 on a 5-second sweep, against 18.8× for the plain transformers baseline — a 1.47–1.66× speedup, saturating around concurrency 96–128 at roughly 6.4 clips/s. On 30-second clips it peaks at 85.3× realtime at concurrency 64. The lesson in that pair of numbers is to always quote a realtime factor together with the clip length it was measured on; the advertised ~35×-per-A100 figure was not reproduced locally. Source: [TR-0829] §3.4.

I.3.5 The numerics that cost hours, recorded because they generalise

• v1.5 has no loss in forward(). It is a base model that returns only last_hidden_state (architecture MossTTSLocalModel). The channel-wise loss has to be computed externally, by re-projecting the hidden states through the local transformer and the text_lm_head / audio_lm_heads — the reference implementation is compute_supervised_loss_from_hidden in moss_tts_local_v1.5/finetuning/sft.py. Labels must be packed with the 13-channel MossTTSLocalV15SFTDataset, not the delay-pattern dataset of the sibling architecture.
• bf16 full fine-tuning NaNs unless AdamW's eps is raised to 1e-4. At the default 1e-8 the term underflows in bf16 and the loss is NaN after step 1.
• Gradient checkpointing must use the module-wrapper path — configure_gradient_checkpointing(scope="global", use_reentrant=False, impl="module_wrapper"), not model.gradient_checkpointing_enable(), which yielded NaN gradients.
• With a manual DataLoader the batch must be moved to the device explicitly. The global forward tolerates CPU input; the manual loss uses the labels directly and does not.
• Checkpoints are about 27 GB each (8 GB model + 16.5 GB bf16 optimizer state + 9 GB of save_pretrained for the best copy), so they were written to a 504 GB RAM-backed /tmp, keeping the last three and the best two.

Two of these — bf16 eps and the checkpointing path — are the same class of failure: a numerical default that is correct in fp32 and silently wrong in bf16. The sibling diffusion-model project in the same group (I.16) hit the identical class from the other side, where bf16's unit in the last place exceeded a typical Adam update and training silently froze; the fix there was fp32 master weights. The transferable rule is that bf16 needs its optimizer numerics checked explicitly, not inherited. Source: [TR-0829] §3.5; the diffusion case is [REF-V1] Part II.

I.4 The adapter ecosystem built on v1 and v2 — the first Tools

Before the timed-script round and long before the agent, the project's main mechanism for control was adapters. In the metaphor these are the actor's first tools: things you hand the instrument to change what it does without rebuilding it. Every one is a PEFT LoRA attached to the same frozen base model, and — this is what makes them stackable — they all declare the same 23 target-module patterns: the global Qwen3 stack's q, k, v, o, gate, up, down_proj; the local GPT-2 decoder's c_attn, c_proj, fc_in, fc_out; and all twelve audio_lm_heads.0…11. They differ only in rank and in what they were trained on. Source: [TR-0829] §4.

I.4.1 Emotion LoRAs: three generations, and what each one taught

v1 — the rescue result (2026-07-26). Evolutionary search over natural-language prompts had failed outright to elicit Anger and Amusement from the base model. Training a per-emotion rank-64 LoRA on emotion-heavy audio rescued them. The ablation used the best checkpoint and the same emotional prompt in both arms, so the LoRA is the only variable (EmoNet z; higher = more of the emotion). No n per cell is recorded in the source for this first ablation.

Rank 64 became the default (rank 128 overfits the ≈1,000-clip buckets), and the best adapter was consistently an intermediate checkpoint — 20 epochs drives the loss to 0.03–0.4 and overfits. Speaker similarity was preserved at 0.41–0.58.

Overnight scale-up and the mixture study (2026-07-27). 34 emotions were trained on natural EmoLia top-1000 subsets. Two things came out. First, there is no single best rank: Anger and Fear peak at rank 128, Sadness at 64, Amusement only moves at all at rank 32. Second, and more consequential, a seven-way mixture study over three sources — D cross-folder DACVAE, E natural EmoLia, G LAION's Got Talent (high-emotion synthetic voice acting) — found that Got-Talent wins: pure Got-Talent is best for Fear (z = +2.62), Distress, Sadness and Anger; only Amusement preferred a three-way mix. High-emotion synthetic data became the primary emotion-LoRA training source. Clear intensity lifts at this stage: Amusement +1.29, Sadness +0.59, Anger +0.52, Fear +0.47 — with many emotions near zero, because natural speech is subtler.

v3 — 40 of 40, and the synthetic character dialled back. The shipped set, TTS-AGI/moss-emotion-loras-v3, is 40 adapters at rank 32, α 64, single-phase, trained against v2. Its distinguishing change is a cap: Got-Talent material is limited to ≤ 25 % per bucket, so at least ≈60 % of every mix is natural or expressive non-Got-Talent material — EmoLia, DACVAE, the Gemini-generated adult-voice set, and the mitermix collection. The card states the reason plainly: this “removes the synthetic Got-Talent character that v2 had”. So the trajectory of this line is a full round trip — synthetic data was found to be the strongest single source, and then had to be capped because it stamped its own character on the output.

I.4.2 The full 40-emotion sweep: 11 responders and 29 that barely move

One adapter per emotion, best of ranks {16, 32, 64}, about 1,000–1,400 clips per emotion (natural EmoLia topped up with Got-Talent, and about 25 % of rows carrying a same-voice contrasting-emotion reference). The best adapter per emotion was chosen by a gated composite, z + 0.15·blend + 0.5·spk + 0.1·genu — the emotion term is gated and the quality terms are already there, which is the ancestor of every reward used later in this report, including the Best-of-N reward of the agent.

Strong responders — z-lift ≥ 1.0 (11 of 40): Amusement +1.99 · Helplessness +1.79 · Intoxication +1.58 · Fatigue +1.47 · Sexual Lust +1.38 · Distress +1.35 · Anger +1.34 · Fear +1.26 · Sadness and Teasing +1.16 · Malevolence +1.15.

Weak or inverted (29 of 40): high-valence and subtle states — Awe, Elation, Pride, Contentment, Pleasure, Hope, Thankfulness, Astonishment, Infatuation, Contempt … Two are self-reducers and are usable as minimising directions: Contemplation (+0.06 → −0.56) and Concentration (−0.30 → −0.40).

• Rank can collapse an emotion. Emotional Numbness scores −1.68 at rank 32 while ranks 16 and 64 give +0.8 and +0.7. There is no universal rank; low rank (16) wins for narrow or data-scarce emotions.
• Half-merge often beats full merge on total reward — Fatigue, Sexual Lust, Jealousy, Helplessness and Bitterness all peak at λ ≈ 0.5. Only Sadness and Pleasure keep gaining to 100–150 %. Base-native emotions (Interest, Concentration, Impatience) are best prompt-only, with no adapter at all.
• An adapter buys less over a good prompt than it looks. Against the best prompt found by evolutionary search, the mean LoRA lift is only +0.30.

A reference-audio ablation on the same sweep is worth one line because it is so emotion-specific: attaching a same-voice, contrasting-emotion reference to ≈50 % of rows helps Amusement enormously (+2.93 z, beating even the best reference-free mixture) and is mixed or small for everything else, with no improvement in speaker similarity. It is a per-emotion option at around 25 % of rows, not a default.

I.4.3 Character LoRAs, and the geometry of voice space

Voice space is a continuum, not a set of types. The 567 character reference voices, z-scored over their 57 VoiceNet dimensions and clustered at k = 12 and k = 18, give a silhouette score below 0.2 everywhere; k = 18 adds only about 5 % of variance and is less separated than k = 12. Twelve is a workable practical taxonomy; eighteen just subdivides the same families. The poles of the space are bright / high / feminine at one end and dark / low / gravelly / masculine at the other. A caveat travels with it: the 120 creature adapters skew masculine and monstrous, so the feminine and clean clusters get loose matches — the “High-Bright-Feminine” centroid maps to a guttural imp, which is a conceptual mismatch. Good matches: Playful-Feminine → elf, Full-Chesty → ogre, Silken → titan, Gravel → skeleton.

A cross-lingual reference does not clone a voice; it invents one. Over 1,280 generations conditioning German and English output on Japanese anime references, mean speaker similarity to the reference was 0.188 against an unrelated-speaker floor of 0.105 and a self-similarity of 1.000. Pitch tracks (r = 0.591) but compresses toward the model's own register (median F0 182 → 157 Hz). English transfers better than German (0.217 vs 0.158), and best-of-8 lifts it to 0.294. The recommendation is to use cross-lingual references to mine new character voices, never to clone — a limit that returns in III.4 as a constraint on dubbing.

I.4.4 VoiceNet-dimension LoRAs

One rank-32 adapter per pole of each VoiceNet dimension, trained on v2. The published evolution study covers 86 dimensions across high and low poles. Δ100 is the measured shift at full dose. Most responsive: Ranting style S_RANT high +0.78 · Tension TENS high +0.67 · Dynamic arc DARC high +0.65 · Respiration RESP high +0.65. Least responsive: Mixed resonance R_MIXD low +0.04 · Roughness ROUG low +0.06 · Mixed resonance high +0.06.

Averaged over all of them, λ = 0.5 delivers about 36 % of the full-strength shift (mean in-direction shift +0.350 at λ = 1.0 against +0.157 at λ = 0.5) — so a half-strength merge buys a large fraction of the effect at much lower risk to speaker identity. That matches the search agent's independent canary finding ([REF-V1] Part III) that tempered doses of 0.4–0.5 preserve identity where 1.0 and above maximise the target and destroy it.

A practical trick from the manual that has no obvious analogue elsewhere: use VoiceNet adapters as stabilisers. Adding vn_CLRT_high @ 0.4 and vn_STRU_high @ 0.3 (plus vn_S_NARR_high for storytelling) protects intelligibility when a strong emotion or style adapter would otherwise push word error rate up; adding +0.5·ESTH to the reward biases toward “nice to listen to”. This is the direct ancestor of the esthetics_high adapter in the agent's shipped stack (Part II).

Bucket feasibility, measured on game-voice data. Per-dimension buckets at three poles — high (≥ 5), mid (3–4), low (≤ 1), capped at 1,500 each — give 154 trainable buckets (≥ 200 clips) across 57 dimensions. Sixteen dimensions have no high bucket at all in game voices (S_TECH, VALN, S_NEWS, STRU, SMTH, R_ORAL, R_MASK, R_MIXD, R_HEAD, EXPL, DFLU, ESTH, COGL, CHNK, BRGT, BKGN) and need podcast or empathic-speech fill; EXPL has essentially no variation in that data at all.

I.4.5 Explicitness LoRAs

Six adapters, published both publicly and in a gated form requiring manual approval, an 18+ confirmation and name/email collection: TTS-AGI/moss-explicitness-loras. The set is aesthetic_mix_r32 (the recommended default), adult_r32, a1_r32, and raw_r16 / raw_r32 / raw_r64, of which raw_r32 gives the largest Sexual-Lust lift. The recommended recipe is the aesthetic adapter at 0.2–0.8 with the emotion adapter held at 0.35–0.75. Source: [TR-0829] §4.5.

I.4.6 A negative result worth keeping: conditioning on 99 numeric scores

An attempt to bypass natural language entirely. Ninety-nine numbers — the 40 EmoNet emotions, the 57 VoiceNet dimensions, genuineness and vocal-burst blend — were each projected into the model's 2,560-dimensional latent space by a tiny per-score projector, and the 99 resulting rows were spliced into the prompt between the backbone's <|vision_start|> and <|vision_end|> tokens, using the otherwise-unused <|image_pad|> token as a placeholder. A rank-128 LoRA and the 99 projectors were trained jointly, warm-started from an existing adapter. [REF-V1] adds the slot layout: slots 0–39 the emotions, 40–96 the VoiceNet dimensions, 97 genuineness, 98 blend.

The first thing measured was the ceiling: how much of a conditioning signal survives a round trip through the codec at all. Encode a clip, decode it, re-score it, and correlate:

The outcome of the experiment itself was negative: the projected-score conditioning path came out slightly below plain v2 on the two proxies in a head-to-head evaluation (genuineness −0.12, blend −0.44; n not recorded in the source). Numeric conditioning is promising and did not beat text in this iteration. The correlation ceilings, though, are permanently useful, and the 0.823 for blend explains a recurring difficulty later — including why the blend term of the agent's Best-of-N reward correlates negatively with the human-proxy target (Part II).

I.4.7 Merging: what a dose actually does

Merge weight is an intensity dial, and it is roughly linear. Averaged over adapters and multiple seeds, λ = 0.5 recovers about 40 % of the full-strength gain, with no snap-on, no collapse and naturalness held. (With a fixed seed the same curve looks like a staircase rather than a smooth line — an artefact of sampling, not of the adapter.)

And it destroys speaker identity, monotonically. Measured as ECAPA speaker-embedding similarity to the reference clip, on cross-lingual Japanese → German/English generations:

Anchors from the same encoder: a reference against itself scores 1.000, two different speakers score 0.105. Published guidance: keep λ ≤ 0.5 if reference fidelity matters, 0.75–1.0 if only the emotion matters. Overdriving also truncates: in a controlled stacking run, one sentence and one seed produced 3.84 s with the voice adapter alone and 1.68 s — under half — with an emotion adapter at dose 1.9 stacked on top. An overdriven stack does not merely sound wrong; it stops early. This is the earliest observation of the mechanism behind the agent's babble bug (Part II.7).

Difference vectors. Taking Effect = mean z at λ = 1.0 minus mean z at λ = 0, and recording for each measured dimension which adapter raises it most and which lowers it most, turns the adapter library into a set of directions rather than a set of labels. Highlights: the Amusement adapter raises Intoxication by +1.19 as well as Amusement by +1.80; the Sadness adapter is a broad negative lever, raising Embarrassment +1.08, Disappointment +0.85 and Fear +0.66; and subtracting Teasing (−1.70 on Concentration) is a “more concentrated” direction. Hope at −0.83 on Contentment behaves the same way.

I.4.8 Vocal-burst merging: the dose knee and a blind evaluator

An in-sentence merge-dose curve over 1,536 generations, 8 families × 64 classes:

λ = 0.75 gives 91 % of the full-dose burst presence at the peak of genuineness and a better blend; presence is bought against blend roughly one for one. Two exceptions: Sexual Lust peaks at 0.75 and then drops, Fear is best at 1.0. Five classes never fire inside a sentence at all — hiss, kissing, lip-smack, playful whistle, slurping. Rank barely matters (composite 0.416 / 0.424 / 0.427 for ranks 16 / 32 / 64). The bare-burst prior is the opposite: in isolated-burst sweeps λ = 1.0 won every class. Do not carry that setting into sentence generation.

I.4.9 Contained emotion — a mechanism that works, and where it reverses

“Contained” means the emotion is present but held in, which is what most adult dramatic performance actually is. The mechanism that works is specific: keep the emotion and drop vulnerability — a light vn_VULN_low at 0.06–0.18 plus a parenthetical restraint cue, over a character-adapter scaffold with the emotion at about 0.4. What does not work: turning the emotion down, adding tension, or applying heavy Emotional-Numbness or vn_TENS_high — all of those armour the delivery into a news-anchor flatness.

A deep paired re-run over 1,878 clips measured ΔVULN as the suppressed cohort's mean minus the free cohort's; negative means the masking worked:

The rule that falls out: masking works when the audible signal is effort, arousal or social surface and vulnerability is incidental; it fails or reverses when vulnerability is the emotion's core — for Sadness, Pride, Distress and Affection, ship the free take. This result overturned an earlier “moderate contains, intense flattens” story that turned out to be an artefact of four anchors and sparse unpaired data — the first of the retracted claims collected in Part III.

I.4.10 “Text carries the condition”

A large ablation — 106 growing to 124 arms, up to 3,968 candidates — found that several apparently “vocal” dimensions are in fact properties of the written line, and no adapter moves them. Source: [TR-0829] §4.10.

• Disfluency is a text property. Adapter-alone levels are indistinguishable (0.27 / 0.44 / 0.71 / 0.71). Writing hesitation into the line — ellipsis, filled pauses, a false start, and a cue naming it — ladders DFLU from 0.44 to 2.07, at a cost of +0.176 WER. Ellipses drive the dimension harder (span 1.54 vs 1.07) while [pause] tags are gentler on intelligibility (0.122 vs 0.238 WER) — so ship both. [pause] and [long pause] were verified not to be spoken aloud.
• Intoxication headline: writing the state into the text scales measured emotion strength by ×8.5 (0.068 → 0.582) and genuineness by +40 %, while doubling WER.
• Explicitness has two conflated targets. The EXPL register dimension is essentially not adapter-drivable (ladder span −0.006), and vulgar text gives only a modest span (0.26–0.43). The explicitness LoRA in the adult-intimacy sense is real: adult text plus inline burst tags — without a burst adapter — lifts blend from 1.72 to 6.77; adding a burst adapter lowered it.
• Chunking is a clean negative. Not text-drivable this way; every recipe scored at or below baseline at about 2.4× the WER.
• Language dependence is real. Heavy broken-speech treatment on the most-broken emotions gives +0.703 emotion and +0.637 genuineness; the moderate version is a negative result in English (−0.224) and a positive one in German (+0.113). Decide per language.

I.4.11 Notation, ranking and the reward form

Three further ranking decisions were each measured on 28,698 stored candidates:

• Rank on raw scores, not z-scored within the group. Z-scoring normalises blend inside each group, so a group whose takes are all badly blended still hands its best one a good score. Going raw lifted the median winner's blend from 3.65 to 4.67 (+28 %). (The agent's Best-of-N of Part II goes the other way — min-max within the candidate set — and Part II records why and what it cost.)
• Multiply by (1 − WER); never divide by (1 + WER). The division form is too lenient and produced more runaway generations (18 vs 14). Worse, since most candidates have a negative core score, dividing a negative number by a larger denominator makes it larger — so the division form actively rewards transcription errors. The real fix for runaways is a length-fit ceiling: expected duration = words / 2.8, free up to 1.6×, quadratic falloff after; runaways 14 → 7 at a ≈15 % genuineness cost.
• Character-matching groups zero the emotion term entirely: reward = (genu + blend) × (1 − WER) × identity × length-fit.

The production matrix at the end of this line was 169 distinct adapter keys — 13 character, 40 emotion, 114 VoiceNet, plus explicitness and sports.

I.4.12 Best-of-N economics, and why ASR is the swappable cost

The 16 → 32 step is worth +4.46 % reward, −8.95 % duration error and +0.254 blend; 32 → 64 costs 3.4× for the same increment. Crucially, generation time is flat from batch 16 to 32 (20.34 → 20.54 s on a GH200) because the decode loop is latency-bound, so a batch of 16 wastes half the GPU. N = 32 is the operating point of the corpus-building era, and peak VRAM there is about 23 GB. (The agent of Part II runs BON_N = 8 because it is interactive and must answer within one turn; the cost argument is the same, the budget differs.)

The dominant cost is then the speech recogniser, and it is swappable: Parakeet-v3 runs at 52× realtime and is 12 % of a 32-candidate group's cost; Whisper-turbo 12.7× and 32 %; CrisperWhisper 3.1× and 66 %. Rank on Parakeet and re-transcribe only the top three with CrisperWhisper — on a 40 × 3,000 production run that took the bill from 2,952 to 1,148 GPU-hours. Parakeet-TDT 0.6b v3 is the ear the agent uses for WER today.

Two more rules from the same body of work, both of which recur later: rank whole assemblies, not parts — across 108 rounds the top-ranked assembly used the best candidate of every part in only 11 rounds (10.2 %) — and n = 8 per cell is noise; use n ≥ 30 with the base model in the same batch.

I.4.13 What building all of them taught about rank, dose and stacking

Rank is not where the quality is. The voice-profile pilot trained each of ten voices at ranks 16, 8 and 4 in a single process against one shared frozen base, stepping on the same micro-batches with the same seed — so rank is the only difference between the three adapters. On 1,920 held-out clips per arm:

The base → rank-4 effect is about ninety times the rank-16 → rank-4 gap, and the inter-rank gaps are indistinguishable from zero on every axis measured. Nine of the ten voices ship at rank 4 — 8.6 M trainable parameters. (Speaker similarity here is cosine similarity between ECAPA embeddings from speechbrain/spkrec-ecapa-voxceleb.)

Dose is not what carries the result. Across the acting corpus, emotion-adapter dose correlates −0.226 with the measured emotion peak and +0.084 with genuineness. In the manual's words: “you cannot buy a turn by raising λ.” 98.6 % of the 492 doses actually used across the corpus sit in the 0.35–0.75 band. What carries the result is the written line (I.4.10).

Only 22 of 40 emotion adapters beat a good prompt. The manual's own emotion index reports that only 22 of the 40 emotion LoRAs produce a positive lift over a well-written prompt with no adapter. Biggest wins: Sexual Lust +0.25, Fatigue +0.21, Helplessness +0.18, Jealousy and Envy +0.17. Biggest losses: Distress −0.18, Concentration −0.12, Anger −0.12, Relief −0.11, Disappointment −0.10. Eight emotions — Interest, Concentration, Impatience and Irritability, Amusement, Hope, Doubt, Disgust, Pride — are best served by a neutral prompt and no adapter at all.

Stacking is a real engineering problem. Because every family targets the same 23 modules, a voice adapter and an emotion adapter are always fighting over the same weights, and at equal λ the larger-rank adapter dominates. Voice adapters are rank 4; emotion, burst, VoiceNet and character adapters are rank 32. That asymmetry — not any target-module difference — is why every non-voice λ is below 1 unless that adapter is the condition. Two library-level traps are documented with their symptoms, and both killed multi-hour runs: PeftModel.active_adapter is a plain attribute that base_model.set_adapter() does not update, so it stays pinned to the first adapter ever loaded and raises KeyError forever once that adapter is evicted; and a snapshot of the trained scalings goes stale the moment another adapter is loaded, so a dose computation must consult both dictionaries and record the trained value on first sight. A third surfaced on 28 August and is recorded with the published cards (IV.1).

I.5 Five hundred synthetic voice profiles

A large part of the training material for the timed-script round is not recorded speech but voice profiles: synthetic speakers, each generated by the v2 model with the adapters of I.4, each exercised across the whole expressive range — 40 emotions, 57 VoiceNet dimensions, acting edge cases, character clusters, vocal bursts, English and German — and each filtered by the scoring models before anything was kept. Source: [TR-0829] §5.

Scale, measured on 12 verified voices: 40,143 non-empty clips per voice · 108.4 h of playtime per voice · 842 condition groups · 30,869 unique utterances · mean clip 9.72 s. Projected to 500 voices: ≈20.1 M clips, ≈54,200 h — about 6.2 years of audio. Cost: 31.2 GPU-hours per voice billed → ≈16,150 GPU-h ≈ 1.16 M core-hours for 500 voices. Throughput held at scale: 1,354–1,451 generations per GPU-hour against a pilot mean of 1,405.

The selection rule that produced them is the direct ancestor of the reward used in the timed-script round and of the agent's Best-of-N reward:

reward = (genuineness + w_blend·blend + 1.25·target) × (1 − WER) · score = reward × durationFloor × lengthCeiling × identityRank × floorPenalty

Three of its design choices were each measured on 28,698 stored candidates, and each is a small lesson (the first two are I.4.11 seen from the data side):

• Raw scores, not z-scored within a group — median winner's blend 3.65 → 4.74 (+30 %), at a cost of −15 % genuineness. (Note the two sources disagree on the second number: I.4.11 says 4.67 / +28 %, I.5 says 4.74 / +30 %. Both are transcribed from [TR-0829]; the discrepancy is recorded in III.5.)
• Multiply by (1 − WER), never divide by (1 + WER). On a 1,000-clip control set a literal WER × quality filter put 602 of 1,000 candidates at exactly 0 and then selected the half with worse word error rate.
• Identity is ranked at 19 % weight, never gated high. Among listener-confirmed same-speaker pairs the median ECAPA similarity was 0.632 and 55 % fell below a 0.68 gate — gating there would have discarded most genuinely correct takes.

The largest published artefact from this stage is laion/laion-voice-profiles-annotated — 28,212,933 rows, 5.8 TB.

[REF-V1] adds two details that later versions dropped: the voice-profile pipeline used 108 rounds of assembly ranking (the 10.2 % figure of I.4.12 comes from there), and the per-voice adapters of the pilot were trained in a single process against one shared base so that rank comparisons would be paired — the design that made the p = 0.22 / 0.35 result of I.4.13 possible.

I.6 The measurement instruments, and how they were trained — the Ears before the actor

Nothing in this project can be evaluated by listening at the scale at which it is generated — a single GRPO run produces 1,400–1,700 fully scored clips per hour. So the project runs on a set of learned scoring models. They do three different jobs, and it is worth keeping those apart: selection (which of 32 or 64 candidate takes to keep), reward (what an RL objective maximises) and evaluation (the numbers in this report). The same models do all three, which is a known circularity and is discussed in I.6.8. In the metaphor these are the actor's Ears; Part II.5 describes how the agent uses them, this section how they were built. Source: [TR-0829] §6.

I.6.1 VoiceCLAP — the shared audio embedding

VoiceCLAP is a CLAP-style (Contrastive Language–Audio Pretraining) dual-tower model: an audio encoder and a text encoder trained so that a clip and a sentence describing how it sounds land close together in one shared vector space. The workhorse is laion/voiceclap-commercial:

A controlled ablation reported on the card shows that removing the non-commercial data costs nothing — the commercial-only model wins 4 of 5 benchmarks against the full-mix model, trailing only on fine-grained intensity ranking of synthetic audio. Two recovery experiments were tried and discarded: adding VoxCeleb1/2 hurt, and upweighting the safe corpora was flat.

A larger sibling exists, laion/voiceclap-large-v2 — a rank-16 LoRA (α 32) on the 7 B LCO-Embedding-Omni-7B (Qwen2.5-Omni thinker), producing a 3,584-d embedding and setting the better numbers on the human-annotated VoiceNet benchmark (emotion balanced accuracy 0.7069 against voiceclap-small's 0.6754). It is not what the pipeline runs on, because the 768-d commercial model comes within a hair of it on the two heads that matter at 1/4.7 the embedding width. In the agent era the two were compared directly as burst-detector backbones (Part II.5); the large model wins the ranked-vs-measured comparison there.

The architectural fact that matters most: one VoiceCLAP encode feeds three separate instruments. In the production scorer (pp_scores_fast.FastScorer) a clip is encoded exactly twice in total — once by VoiceCLAP-commercial and once by BUD-E-Whisper — and every downstream head reads one of those two vectors.

I.6.2 Genuineness: does it sound like a person or like a reading?

There are two generations of this predictor, and conflating them is easy, so both are described.

Generation 1 — the Empathic-Insight “Authenticity” head. The original genuineness signal is one of the expert heads in the Empathic-Insight-Voice suite: the repository ships 55 checkpoints, of which 40 are emotions and 15 are demographic/vocal/psychological attributes — and Authenticity and Arousal are two of those 15. It is trained exactly like an emotion expert: a BUD-E-Whisper encoder is fine-tuned once and then frozen, and a single MLP is trained on its flattened sequence embedding. This is the head occupying slot 97 of the project's 99-dimensional scorer (va_rescore.NinetyNineScorer), and it is what filtered the data for the v1/v2-era emotion buckets. Its published private siblings are laion/genuineness and laion/vocal-burst-blend. One number from that generation is worth carrying forward: inter-rater agreement on authenticity is very low — Cronbach's α ≈ 0.185 on n = 300, far weaker than the emotion labels. The head is modelling a genuinely subjective target. That is precisely why genuineness is down-weighted (≈0.1 in the selection score of I.4.2) and used as a gate rather than as a primary objective.

Generation 2 — the VoiceCLAP head the pipeline uses. laion/voiceclap-commercial-genuineness predicts a single continuous score in [0, 6] for how much a clip sounds like a real, lived-in spoken moment rather than a rehearsed or synthetic read. Anchors from its own card: 0–2 reads as rehearsed, flat or synthetic; 3–4 plausibly natural; 5–6 sounds like a genuine, unscripted moment. It explicitly is not about audio fidelity — it is about natural timing, breath, micro-imperfections and emotional grounding.

I.6.3 Vocal-burst blend: does the sigh belong to the sentence?

laion/voiceclap-commercial-vocalburst-blend predicts, on a 0–10 scale, how naturally a non-verbal burst blends into the speech around it. Anchors from its card: 0 = disconnected — spliced or pasted in, robotic, or the wrong emotion for the context, and also assigned to clips with no genuine burst at all; 5 = fits the context but sounds performed or stagey; 10 = totally organic, indistinguishable from a natural human reaction. Intermediate values interpolate; output is clamped to [0, 10]. Architecturally it is the same shape as the genuineness head — frozen VoiceCLAP-commercial 768-d → L2-normalise → standardise → Linear(768,50) → GELU → Dropout(0.2) → Linear(50,1), about 38 k parameters — which is why one encode yields both scores. Its card publishes a comparison at equal head architecture across three embedders:

These two heads are why the project's data selection is not naive. As I.12.1 shows, the most intense recordings in the corpus are on average the least genuine, so selecting on intensity alone selects for overacting. Genuineness and blend are the gate that stops that — and they are two of the four terms of the agent's Best-of-N reward.

I.6.4 Empathic-Insight-Voice-Plus: 40 emotion heads and 4 quality heads

laion/Empathic-Insight-Voice-Plus is not one model but a suite of independent single-output MLP regression experts on top of a frozen BUD-E-Whisper encoder (a fine-tuned Whisper-Small producing a 1,500 × 768 sequence).

• Emotion / attribute experts (from Empathic-Insight-Voice-Small, 55 checkpoints = 40 emotions + 15 attributes): flatten the whole sequence and project it down — Flatten(1,152,000) → Linear(…,64) → ReLU → Dropout → Linear(64,64) → … → Linear(16,1), about 73.7 M parameters each, dominated by that first projection. Trained on LAION's Got Talent (≈5,000 h of multilingual synthetic voice acting) plus an in-the-wild voice-snippet collection of similar size. The taxonomy and benchmark come from the EmoNet-Voice work.
• Quality experts (the 4 that -Plus adds): pooled features instead of the full sequence — mean + min + max + std over the sequence dimension, 3,072-d — and therefore only ≈200 k parameters each. Three are distilled from Microsoft's DNSMOS (overall quality, speech quality, background quality) and one from Meta's AudioBox content-enjoyment model. Validation on 200 samples each: overall MAE 0.26 / r 0.899; speech 0.30 / 0.517; background 0.35 / 0.865; content enjoyment 0.34 / 0.691.

The production scorer loads 44 heads — the 40 emotions and the 4 quality experts — and leaves the attribute heads unloaded, because the VoiceNet regressor (I.6.5) covers those dimensions. Emotion outputs are unipolar, nominally 0 → 4 (0 = the emotion is absent, 4+ = extreme manifestation, card maximum 4.5), but the heads are unbounded regressors and real corpus values overshoot — the project's own normalisation statistics record a maximum of 8.20 on the Anger head.

The 40 emotions, verified against the repository's file listing, are the axis along which everything in this report is requested and measured: Affection · Amusement · Anger · Astonishment/Surprise · Awe · Bitterness · Concentration · Confusion · Contemplation · Contempt · Contentment · Disappointment · Disgust · Distress · Doubt · Elation · Embarrassment · Emotional Numbness · Fatigue/Exhaustion · Fear · Helplessness · Hope/Enthusiasm/Optimism · Impatience and Irritability · Infatuation · Interest · Intoxication/Altered States of Consciousness · Jealousy & Envy · Longing · Malevolence/Malice · Pain · Pleasure/Ecstasy · Pride · Relief · Sadness · Sexual Lust · Shame · Sourness · Teasing · Thankfulness/Gratitude · Triumph.

I.6.5 VoiceNet: 57 dimensions of how a voice sounds

laion/voicenet-dimension-predictors-commercial is a bank of 114 small MLP heads — 57 dimensions × 2 (a regression head and a classification head) — on the same frozen 768-d VoiceCLAP-commercial embedding. One encode yields all 57 dimensions. Outputs are ordinal levels, 0–6 for most dimensions, 0–4 for background noise and 0–2 for content appropriateness.

How it was trained. Labels are laion/emolia-voicenet-gemini-annotations — Emolia clips each re-scored on a single VoiceNet dimension by Gemini-3.5-flash, combining three annotation rounds (round 3 topped up the most data-starved score levels of the hardest dimensions). Each unique clip is encoded once. For every dimension three class-balancing variants are trained (buckets capped at 3× / 4× / 5× the second-smallest usable bucket), and the shipped default is chosen per dimension on a frozen validation split — lowest MAE for the regression head, highest accuracy for the classification head, chosen independently. Regression uses Huber (δ 1.5), H = 64, dropout 0.33; classification uses cross-entropy, H = 48, dropout 0.33; Adam, 50 epochs, best-val checkpoint. Best-per-dimension results: mean regression MAE 0.744, mean Pearson 0.79, mean within-one-level accuracy 0.870, mean classification accuracy 0.617. Per-dimension quality varies a great deal — head resonance reaches MAE 0.478 with r = 0.88, while cognitive load sits at 1.052.

I.6.6 The burst locator and the burst detector

Bursts are handled by two models in sequence, because “where is it” and “what is it” are different problems. laion/vocalburst-locator does frame-level binary segmentation at 20 ms resolution: a Whisper-small encoder adapted with a rank-8 (α 16) LoRA on q_proj/v_proj and then merged, followed by Linear(768→384) → GELU → Conv1d(384, kernel 7) → Linear(384→1) → sigmoid, giving 1,500 per-frame probabilities over a 30-second window, post-processed into (start, end, confidence) events. Event F1 at IoU 0.5 on 992 real in-the-wild expressive-speech clips: 0.607 for v2 against 0.152 for v1 — and v1 fails in a specific way, reaching usable precision only at threshold 0.80 where recall collapses to 0.21.

laion/vocal-burst-detector-v2 then classifies each located span: an MLP (Linear(768,256) → BatchNorm → GELU → Dropout(0.3) → Linear(256,83), ≈219 k parameters) on the VoiceCLAP-commercial embedding of the cut, over an 83-class taxonomy — 82 vocalisations plus no_burst at index 82. Argmax accuracy on the official 1,940-clip validation set is 58.1 % overall: 70.4 % on clips with a single pure burst, 39.4 % on composites. Five classes that are hand or body impacts rather than vocalisations — blowing a kiss, finger snaps, hand scratching head, hand slaps, slap face — are folded into no_burst and masked at inference, so the recognised label space is 77 burst classes plus no_burst. A burst is additionally vetoed when P(no_burst) ≥ 0.50. The agent era replaced this detector with a 17-class successor, laion/vocal-burst-detector-x2, because 83 classes at 58.1 % could not be trusted for a reward term; Part II.5 has its numbers.

The pre-agent reward pipeline ran the locator at threshold 0.5, merge gap 0.2 s, minimum duration 0.08 s, padded each span by 0.05 s before classification, and dropped anything under 0.10 s (LOC_THR, LOC_GAP, LOC_MINDUR = 0.5, 0.2, 0.08; BURST_PAD, MIN_BURST, NOBURST_GATE = 0.05, 0.10, 0.50). The card's own recommendation is slightly tighter — merge gap 0.10 and minimum duration 0.10 — and the difference has not been ablated.

I.6.7 Turning raw scores into a scale you can ask for

Everything above produces numbers on incomparable scales. Measured on the project corpus, the Interest head has a median of 2.082 and never outputs zero, while the Infatuation head has a median of −0.017 and outputs zero on 87.7 % of clips. A reward built on raw values would have been an Interest detector. The answer is the corpus ECDF: every raw score is mapped to its percentile within the full 3,147,802-row corpus, per head. Four intensity bands are defined on that percentile scale, and the same four cutoffs are used by the prompt builder, by the reward function and by the evaluation harness, so what the prompt asks for is exactly what gets measured.

The zero-point defect, measured 25 August 2026. A plain ECDF percentile turned out to be still not comparable across the 40 heads, for a reason that is obvious in retrospect. A clip with a raw score of exactly zero — no measurable emotion at all — lands at a percentile that depends entirely on how often that head outputs zero:

Asking for band 0.90–0.98 was therefore a different task for every emotion: on Awe the model starts 0.08 below the target having produced nothing, while on Affection it starts at zero. Averaged over 40 heads, the reported figure was measuring where each head's zero point happens to fall, not how well the emotion was performed. The fix is a per-head affine rescale of the tail above the zero point, p′ = max(0, (p − p₀) / (1 − p₀)) where p₀ is the percentile of a raw score of zero. It was verified to map raw 0 to exactly 0 on all 40 heads while remaining monotone within each head — so every comparison between two clips of the same emotion, which is what DPO pairs and GRPO group rankings consist of, is unchanged. Only absolute targets and cross-head averages move.

I.6.8 How far these instruments can be trusted

• No human listening study had been run as of 29 August — not on the corpus, not on the adapters, not on any model comparison. As of 9 September the situation is: 2,000 stimulus pairs are published and a listening Space exists, but no votes have been collected (Part IV.1). Every quality figure in Parts I–III is the output of a learned scorer.
• The project's own estimate of agreement between these learned scorers and human listeners on spoken material is about ρ = +0.21 (I.7.3). That is a weak positive correlation: enough to rank two arms of the same experiment, not enough to certify quality. On non-verbal material the manual reports that the sign is wrong.
• The findings protocol records seven separate occasions on which an automatic metric or a job status reported a success that did not exist, each found only by checking something the metric did not cover. Two examples: an aesthetics head scored an arm at 4.12 while human-facing judges scored the same audio 0 out of 30; a page reporting 89 groups as “improved” was serving audio from a different run entirely, because the audio key was not unique across runs.
• Every head above is trained on labels produced by another model — Gemini-3.1-Pro for genuineness, Gemini-3.5-flash for the VoiceNet dimensions, DNSMOS and AudioBox for the quality experts, forced alignment for burst timing. The genuineness card says it plainly: treat scores as a soft ranking signal, not ground truth.

Two mitigations have actually worked and are recommended rather than merely stated: always run a control arm that isolates the claimed cause, and verify identity from an independent namespace — never infer “this file is that record” from a shared key. The agent-era Ears add a third: a human-proxy target (the burst detector's own AUC against a held-out human-labelled set) that the reward terms are correlated against, so that a term's usefulness is a measured number rather than an assumption (Part II.6).

I.7 The prompting model, the manual, and the v2 casting agent — the actor's first draft

The shipped v2 model is documented in a public manual: *MOSS Voice-Acting — Manual & Studies Hub*. It is 25 chapters: a hub, an emotion-conditioning manual with one page per emotion, twelve recipe chapters, and an eight-page tree recording an autonomous casting agent's four generations of runs. It is unusually forthright — a large fraction of it is a record of things that did not work, with the measurements attached. Source: [TR-0829] §7–§8.

I.7.1 The prompt surface the model ships with

Two fields, exactly:

instruction = GENERAL: <the standing description of the voice and the situation> / SCRIPT: (delivery cue) "the spoken line, (burst) with tags inline, [pause] and pauses in squares" — and text = <exactly the same spoken words, plain — no cues, no brackets>.

The words appear in both fields — plain in text, cue-annotated inside SCRIPT: — and that redundancy is deliberate and matches the training data. The manual's notation rules, verbatim:

• Round brackets carry delivery cues and vocal bursts: (quietly, with lethal control), (screams). This is the only inline notation the model was trained on.
• Square brackets carry pauses: [pause], [long pause]. Never put one directly after a burst — that is the truncation shape.
• Never capitalise a tag. (SCREAM) is read out as “ess see arr ee ay em”. Lower-case everything inside brackets.
• Give the line enough words. At ≈2.8 words/s a line under 8 words cannot fill three seconds. Ten words is the working minimum.

A worked example, copied verbatim from the manual's casting-agent prompt archive (adapters Scream @ 0.9, Anger @ 0.4, Distress @ 0.35): GENERAL: Start wounded and breathless, then rebuild into nearly screaming but controlled rage. The final words should land cold and absolute, with the voice hardened by pain. Place the vocal burst inline and continue speaking immediately afterward. SCRIPT: (a broken breath turning into a controlled eruption) "Don't you dare touch me." (Scream) "You don't get to cry now. You don't get to make this about your guilt. Pack your things, and get out before I say something neither of us can take back."

There are no duration tags in the shipped prompt format. Length is controlled only indirectly, through the tokens field (target frames, ≈ words × 6) and max_new_frames. The manual gives a specific warning for non-verbal parts: tokens defaults to the number of words, and a scream's text has almost none, so the model is told to make ≈5 tokens and stops after 0.3 s — for non-verbal material you must set about 12.5 tokens per intended second by hand. This is exactly the gap the timed-script round (I.8) closes.

I.7.2 How intensity is requested, and why the manual says the prose does the work

Intensity in the shipped format is not a tag and not a number. It is requested three ways, and the manual is emphatic about their ranking:

• The written line and the parenthetical cue. The manual's own chapter title is “Genuineness comes from the writing, not from the dose”. The rule for a cue is: make it an action the character performs, not a description of a tone. “The sentence collapses halfway and they have to start it again” is a direction; “sad, resigned” is a label, and a label is an invitation to narrate.
• The adapter dose λ, which as I.4.13 showed correlates negatively with the measured emotion peak.
• A declared 2 × 2 condition — {moderate, intense} × {free, suppressed} — used in the contained-emotion chapter (I.4.9).

One finding here anticipates the whole of the GRPO story (I.11): a maximise objective cannot produce a moderate emotion. The manual states it directly — “a maximize objective always drifts to intense; a believable moderate/subtle emotion scores low against ‘maximize’ and dies out of the population.” The fix used there is a target-band fitness function, which is precisely the shape the timed-script reward uses. The manual also independently reports the cross-head incomparability of I.6.7, in its own units: “Anger/Sadness/Fear reach ≈2.4–3.7 on the detector; Amusement tops out ≈1.7 even when strong — do not compare raw emotion scores across emotions; compare each against its own free take.”

I.7.3 Best-of-N is not optional, and it is nearly free

Because emotion strength, burst placement and prosody are seed-dependent, the recommended way to use this model is to generate many candidates and rank them. The economics are in I.4.12; the manual adds: for a mid-sentence burst specifically, a burst appears in roughly 50 % of takes, so five candidates give about a 97 % chance that at least one has it; rank whole assemblies, not parts (10.2 % agreement between the best assembly and the part-wise best); at scale, transcribe for ranking with a fast ASR model and re-transcribe only the top three with the accurate one. Best-of-N is the one element of the v2-era practice that survives unchanged into the agent (Part II.6).

I.7.4 The v2 casting agent and its reliability limits, stated plainly

The manual documents an agent built on the v2 model: it writes a ≈30-second performance as 2–4 parts, generates each best-of-16, assembles the best combination across parts, ranks with the local scorer stack, and has a supervisor listen afterwards and give feedback for the next round. Seven generations of it were run — four complete, three aborted — over nine challenges, roughly 2,000 generations. A separate autonomous LoRA-plus-prompt search agent (LAION-AI/voice-acting-search-agent) ran an overnight programme of 163 agent generations and ≈10 k scored takes on a single A100; its notable finding is about the brain, not the speech model — a Gemma-4-31B QAT policy won decisively on single-dimension missions using only 22 tool calls, while a Gemini supervisor was an effective active director (tracking the objective at r = 0.72) and a local audio judge was usable only as a lenient score-only gate, because active steering by it was harmful. That finding — a capable local Mind, a supervisor that directs, an audio judge that only gates — is the design of the Part II agent in embryo.

[REF-V1] Part III preserves more of the search agent than any later version: the policy comparison ran Gemma-4-31B against Gemma-4-12B and a Gemini policy; the winning runs used 22 tool calls against a budget of 40; the canary that caught dose overdrive was a fixed neutral line regenerated after every merge; and the “gotchas” list of that report (adapter eviction, the stale-scaling snapshot, the active_adapter attribute, capitalised tags, empty decodes) is the same list that I.4.13 and I.7.5 carry. The autonomous runs cost about 12 GPU-hours in total on one A100.

That agent can merge the LoRAs, and the audio it produces is more emotional than what the SFT/DPO line of I.9–I.10 produces. It is also, on the project's own measurements, not reliable on character consistency.

I.7.5 The voice drifts

• Without an explicit continuity mechanism, every part is a different actor. Verbatim: “If you omit it, every part is an independent draw of an unspecified speaker. The model has no reason to pick the same voice twice, and it does not. A three-part assembly is then three different actors, and no amount of crossfading hides it.”
• Reference conditioning alone is not enough. Later parts conditioned only on a reference clip reached mean similarity 0.625, with 69 % below 0.75 and 23 % below 0.55. It keeps timbre roughly right but starts a fresh utterance with its own level and attack, heard as a cut.
• Naive chaining collapses identity. Continuing from the full previous part gave a minimum similarity of 0.280 across an assembly; continuing from an anchor plus a 4-second tail gave 0.835. Two causes are named: drift compounds, and a long prefix drags the old emotion forward (peak emotion fell 2.29 → 2.05 under full chaining).
• Ranking does not fix it; only excluding does. “Ranking the best of one batch is not enough: if a whole batch drifts, the least-bad take still ships.” And: “A penalty is not a floor. Multiplying a sub-floor take's score by 0.25 still let an excellent-sounding take from below the floor win. Exclude them.” The sequence score used is ½ mean + ½ minimum, so one bad part cannot be averaged away.
• Every remaining identity failure in the corpus is a non-verbal clip. Spoken clips: mean similarity 0.816, 0 % below 0.55. Non-verbal clips: 0.610, 38 % below 0.55.
• Cross-lingual reference conditioning does not clone at all (I.4.3): mean 0.188 against a 0.105 unrelated-speaker floor.

And the thresholds themselves are wrong when set high. The casting playbook originally used similarity gates of 0.82 / 0.75 / 0.68. The voice-consistency chapter overrides them: regenerate below 0.58, repair with voice conversion below 0.45, reject below 0.40 — and calls that correction “the single most important thing on this page”. The evidence: among 170 human/model-confirmed same-speaker pairs the median ECAPA similarity is 0.632, 55 % fall below 0.68 and 84 % below 0.82, while the highest score reached by a genuinely different speaker was 0.816. Four embedding models were tried (ECAPA AUC 0.847, VoiceNet 0.814, WavLM timbre 0.794, VoiceCLAP 0.770) and none gives a clean threshold. The conclusion: ranking on identity is fine, gating on it is not — and never raise the floor “to be safe”, because the failure that creates (flat, over-clean delivery) is less visible than a voice change but affects every take rather than a few. The agent's speaker-similarity ear (Part II.5) inherits the 0.40 floor and the rank-don't-gate rule.

I.7.6 Under strong emotion, output goes strange

• High emotion dose degrades everything else. For Anger at a high dose: “blend craters from 0.446 to 0.129, quality from 0.552 to 0.261, and WER more than doubles (0.202 to 0.570) … it drags in Malevolence and Contempt while stripping Warmth (−0.48) and articulation clarity, so the anger sounds cruel and slurred rather than heated.”
• Pushing one dimension to its extreme garbles words. “Round 1 failed because ‘maximum volume’ made the model garble.” The mitigation is a safety clause in the prompt: “roaring but the words still land”, “whispered but every word clear”.
• The truncation wall. A named failure mode: the model produces the burst and then emits end-of-sequence, regardless of frame budget or prompt. A silence beat straight after a burst triggers it reliably. Six prompting strategies were tried across 1,152 generations to instruct the model to continue; every one scored lower than the plain inline tag. The conclusion: “you cannot prompt your way out of the truncation. Lower the dose and select over candidates.” And burst-presence alone cannot see this failure — a take that produces a perfect gasp and then stops scores presence 1.0.
• Capitalised tags are spelled out. (SCREAM) becomes “ess-see-arr-ee-ay-em”. The corpus's single clip whose entire purpose was to be a scream transcribed as “Age.” / “Atch! Itch!” / “Akish.”
• The narration failure. The commonest continuation failure is subtle: the voice stays correct but the model stops acting and simply reads the remaining words out. It is measured as narration = ½(S_NARR + S_NEWS) − ½(S_DRAM + S_CONV).
• The dials saturate. Asked for tempo levels 1/2/3/4, the model produced 2.97 / 2.05 / 2.79 / 2.82 — the real range is about 2 to 3, and asking for 4 buys nothing over asking for 3.
• One take holds one emotion. A single generation conditioned on one caption does not turn from composure to collapse; the turn has to happen between generations.
• Whole batches sometimes decode to empty audio, and continuing from an empty clip poisons the next part (0.367 similarity downstream). Empty decodes are a normal silent failure of this model, not an exception — production code must assert on the decoded output.

I.7.7 And the local reward does not predict a listener

“Short answer: on spoken scenes, a little — ρ ≈ 0.21, about 4 % of the variance. On non-verbal scenes, nothing, and the sign is wrong. There is currently no configuration in which the local stack can replace the listener.” Two of the reward's own components point the wrong way against listener preference — genuineness at −0.143 and blend at −0.148; only emotion peak (+0.179) is useful. An optimally weighted linear combination of everything the local stack measures explains under 3 % of a listener's within-round preference. The manual's conclusion is that re-weighting cannot fix this, because the information is not in the feature set. The n behind ρ ≈ 0.21 is the listener's within-round rankings across the casting agent's generations; the source does not state it as a single number, and neither does this report.

Two further findings sharpen it. The casting score in use was measured to correlate +0.91 to +0.98 with its own (1 − WER) multiplier and +0.00 to +0.04 with genuineness — it ranks assemblies by intelligibility and speaker continuity and is close to blind to everything else it claims to measure. And three of its eight terms were dead: the emotion-peak term was saturated at its cap in 92.4 % of assemblies, and the arc term sat at exactly 1.0 in 84 % and took three distinct values ever. Compare the agent-era Best-of-N reward of Part II, where the same question — which term carries the signal — was answered by correlating each term against a human-proxy target on 15,517 candidates: genuineness −0.0085, blend −0.0839, CLAP +0.1482, 1−WER +0.0388, burst +0.3053. The two reward generations agree on the sign of genuineness and blend.

I.8 The timed script

Listening to the round-1 supervised-plus-preference model, two faults dominated: uncontrolled vocal bursts — a sigh that should last 0.5 s runs 5 s — and hallucinated content — text appears that was never prompted. Both are the same missing signal: the prompt never stated how long anything may take. The timed-script round exists to supply it. All of it runs on JUPITER: GH200, 4 GPUs per node, 102 GB each, SLURM account reformo, partition booster. Source: [TR-0829] §9.

I.8.1 The format

The instruction block is unchanged in shape from the shipped format — the same Reference(s) / Instruction / Tokens / Quality / Sound Event / Ambient Sound / Language / Text template — but the SCRIPT: block is now a rendered timed script: [0.9 seconds pause] [3.4 seconds duration] Erster Satz. [0.6 seconds pause] (contented sigh, 0.4 seconds) [4.1 seconds duration] Zweiter Satz. There are exactly four kinds of tag, told apart by their brackets:

The disambiguation rule in one line: square bracket = a number of seconds; round bracket with a number = a vocal burst; round bracket without a number = a delivery direction. That is the only thing separating a burst from a direction, which is why directions are always emitted with no number at all. There is one soft spot: in the untimed 30 % of samples nothing carries a number, so a burst renders as a bare (chuckle) — the same shape as a direction. What still separates them there is length: a burst label is one or two words, a direction is a phrase.

Segmentation rules, all measured against the corpus: split at sentence ends; split again at a vocal burst, at the nearest word gap; split any remaining segment over 12 s at its largest internal gap provided both halves stay above 2 s. A duration is measured from the first word onset to the last word offset of that segment, so two sentences of 12 s and 8 s produce [12.0 seconds duration] and [8.0 seconds duration], never a single 20-second tag. Gaps under 0.20 s are folded into the neighbouring speech rather than printed, so the printed numbers still sum to the clip length. Bursts under 0.05 s are dropped, and a burst that overlaps speech prints only the part that does not overlap.

Verification, done the right way. The emitted string is re-parsed by a separate function with no access to the source data, and the recovered numbers compared against the true clip length. Residual median 0.010 s on 8,614 real-speech rows and 0.020 s on 14,359 voice-profile rows; within rounding tolerance on 99.9 % / 98.7 %; the transcript comes back byte-identical on 100 % of rows. This is the pattern I.6.8 recommends: make at least one leg of a check independent of the pipeline that produced the thing being checked.

I.8.2 The inline delivery direction

The round-2 format dropped the parenthesised delivery direction that round 1 had carried — by accident, and the accident became the round's central negative result (I.9.2). Round 3 puts it back, inside the script, before each duration tag. A direction is composed of five pieces:

Only the first segment gets the full direction; later segments get a short reminder ((still clearly amused)), and segments under 1.6 s often get none — repeating a thirty-word note in front of a 0.6-second line buries the line it is meant to shape. When a row has no emotion label at all — a VoiceNet-axis row, for instance — the opening segment gets the corpus's own most common direction instead.

A limitation kept deliberately rather than silently repaired. All 40 emotions resolve to exactly two harvested phrasings each — one expressed, one suppressed — drawn from only 10 distinct family-level texts. Some pairings read oddly: intoxication, pain and shame all inherit the fear family, so an intoxication direction can say “the fear plain in the voice”. This is not a bug in the generator — it is exactly what the round-1 corpus contains (425 of the intoxication rows use that phrasing verbatim), and the model was already trained on those pairings. It is kept for comparability and flagged as the first thing to revisit if intensity control still lags for those emotions.

Two bugs in this generator were found and fixed on 25 August, seventeen minutes before the round-3 run launched — one that would have made every row render the same intensity word, and one that left the untimed 30 % of samples with no direction at all. Both would have made the round meaningless without changing any training-time number. I.9.4 gives them in full.

I.8.3 What varies between training samples

Nothing is fixed; each sample draws from a seeded generator, so the model sees the whole range and never requires any single element to be present. Measured end to end on 62,672 real corpus rows through the training loader:

The Instruction field itself takes one of four shapes: gen_both (a GENERAL: line and a SCRIPT: block) at 45 % for voice profiles and 75 % for real speech; caption_tpl (a natural-language instruction instead of the GENERAL: line) at 30 % for voice profiles only; gen_general at 15 %; gen_script at 10 %. These four shapes are the prompt forms the agent of Part II tests as a Tool (II.3.4). Realised band mix over those 62,672 rows: 58.8 % extreme, 23.3 % intense, 10.1 % moderate, 7.8 % faint. That skew is not an accident — it comes from training on the “extreme subset” — and I.9.11 argues it is a prime suspect for the failure that follows.

I.8.4 The corpus

The extreme subset is worth one paragraph because of how its definition had to change. The original rule was “the top 10 % of each of the 40 emotions plus the top and bottom 10 % of each VoiceNet dimension”. Measured, that rule selects essentially the whole corpus: the emotion top-10 % union alone is 89.01 %, the VoiceNet top-plus-bottom is 99.37 %, and the combination is 99.84 % of the full corpus. With 154 overlapping axes, a per-axis decile is not a filter. It was re-implemented as a fixed quota per axis; at 6,000 rows per emotion each pick is the top 0.19 % of its axis. Band composition of the subset — 59.5 % extreme, 23.8 % intense, 9.9 % moderate, 6.8 % faint — against the full corpus's 45.6 / 32.4 / 12.8 / 9.3 %.

I.8.5 The evaluation harness

Every model comparison in I.9–I.13 uses the same harness, so the rows are comparable to each other and to nothing else:

• 80 prompts — one per emotion × {English, German} — built by the same function the GRPO trainer samples from, so the evaluation cannot drift away from the trained format.
• 4 completions per prompt = 320 clips per model. One completion would make every per-emotion number a sample of size one, and the within-group spread is itself a quantity GRPO needs to be non-zero, so it is reported.
• Band requested: intense (percentile 0.90–0.98), inline directions always on.
• Per clip: emotion band achievement, quality (blend and genuineness as corpus percentiles), burst realisation, word error rate against the script's own plain text, and duration error against the length the script's own numbers sum to.

The evaluation deliberately uses a stricter yardstick than training. Training needs a gradient, so its band score ramps; evaluation asks a different question — did the clip land in the band that was requested — and answers it with a plateau that pays nothing outside the band. Keeping the two apart is what makes numbers published before and after a reward change comparable. Burst realisation is an F1 over prompted-versus-detected bursts in which each match is weighted by how well its duration matches — a matched class scores 1.0 and a mismatched one 0.35, times exp(−|Δduration| / 0.5), with a 1.5 s tolerance on position. Recall alone would reward emitting a burst everywhere and precision alone would reward emitting none; the original listening complaint was bursts that are too long, so the duration term is what carries the signal, and a matched pair with the wrong length scores near zero. Note that 320 clips per model means the standard error on a between-model reward difference of the size seen in I.12.3 (≈0.01) is of the same order as the difference; I.12.3 says so where it bites.

I.9 Supervised rounds: timing solved, emotion lost, format regained

Three supervised rounds have run. Round 1 established the line: all 4.13 B parameters, 3 epochs over the 3,147,802-row corpus, global batch 4,096, 2,232 steps, 64 nodes × 4 GH200 in 1 h 53 min, held-out loss improving monotonically 4.7076 → 4.6314 across all twelve evaluations. Rounds 2 and 3 are much shorter runs on top, whose purpose is the prompt format rather than more data: each is 2 epochs over the 398,282-row extreme subset, 712 steps on 32 nodes. The sequence is the most instructive thing in Part I, because round 2 was a success and a failure in the same run, the failure was caused by a line that was deleted by accident, and round 3 exists only to put it back. Source: [TR-0829] §10; the round-1 configuration and the loss series are in [P§3–§5].

A comparability note that the protocol records rather than hides: round 2 ran at learning rate 5e-6 and round 3 at 7e-6, so their validation losses are not comparable to each other; nor are they comparable across the round-2 → round-3 prompt change, which altered the objective. I.9.7 argues that in this regime the validation loss is barely comparable to itself. The round-1 learning rate is not stated in [TR-0829]; it is not reproduced here.

I.9.1 Round 2 — timing is solved, and emotion control disappears

320 generated clips per model, band intense requested:

(These three rows are on the pre-correction percentile scale of I.6.7 and are therefore comparable to each other but not to the tables from I.10 onward.) Timing is solved. Median duration error 0.08 s — one audio frame — and every one of 320 clips within half a second of the length its script's own numbers add up to. That is the capability the timed script was built to deliver, it was delivered in a single 29-minute run, and nothing since has degraded it. Emotion is not. Asked for percentile 0.90–0.98, every round-2 model lands at 0.49–0.51 — the corpus median, which is what a model produces when it is not being told anything at all. Three models with three different amounts of preference tuning all sitting on the median is not a training-strength problem; it is the signature of a missing input.

I.9.2 The diagnosis: what round 2 deleted

The cause was found by reading the diff between the round-1 and round-2 prompt renderers. Round 2 had dropped the parenthesised delivery direction entirely. Round 1 had put a direction in front of the spoken text on 55 % of voice-profile samples and 85 % of real-speech samples — (intensely amused: letting it out / not hiding it, warm and open, unguarded) and the like. Round 2 replaced that slot with the timed script, and the direction did not survive the replacement. Nobody removed it deliberately; the new renderer simply did not carry it forward. What was left was description, not instruction. The emotion still appeared, in the GENERAL: line, as a clause saying what the source recording sounded like — “reads as contentment”. That tells the model what kind of clip this is. It does not tell it what to do. The model was trained for two epochs on a format in which no one ever asked it to perform anything, and it learned exactly that: produce the corpus median.

Word error rate 0.48–0.51 for every round-2 model including the bare supervised checkpoint, against 0.11 for the same checkpoint on direction-free prompts. Because the bare model fails as badly as the tuned ones, this is not a property of any adapter. It is the format. Verbatim from the evaluation log — clip k550_age5_bg1__E__Relief__C__en.c036, on a script asking for “November was the only month the headline appeared: The last sporting taboo”: "wer": 4.545454545454546, "hyp": " Ah, ah, ah, ah, ah, ah, ah, ah, ah, ah, …". This matters well beyond the supervised round. Two GRPO runs had been spending hours optimising against exactly that collapse (I.11.1), which is how a reward computed on out-of-distribution prompts ends up measuring the prompt instead of the policy. It is also the earliest appearance of the “ah, ah, ah” babble that the agent's burst-adapter bug reproduced in September (Part II.7).

I.9.3 What round 3 changed: the direction goes back, inside the script

Round 3's entire purpose is to put the delivery direction back and train on it, in the position the timed script leaves for it: in round brackets, without a number, standing immediately before the duration tag of the segment it applies to. Nothing else about the format changed. Two consequences bound what round 3 could possibly have achieved: the direction carries the intensity word, and nothing else does — if the direction is absent, or if the band is computed wrongly, the model has no intensity signal in the prompt at all; and the direction is emitted from the row's own measured percentile, using the same cutoffs as the reward and the evaluation harness, so what the prompt asks for is exactly what gets measured.

I.9.4 Two prompt bugs, found and fixed seventeen minutes before the run started

Both were found by rendering real corpus rows through the training loader and reading the output, which is a cheap check the project now runs before every prompt-format change. Both would have made the round meaningless in a way that no training-time metric would have revealed. The backup files the fixes left behind are timestamped 09:54:21 and 09:55:19 on 25 August; the SFT-3 job started at 10:12:48 the same morning.

The prompt-format hash moved from 64d956b8bb97f0d9 to f90a7b16b4cea093 — the hash is written into every checkpoint's state file, so a checkpoint can always be matched to the exact prompt renderer that produced its training data. That is the mechanism by which the round-2 accident was findable at all.

I.9.5 The round-3 training configuration

One correction to a figure that has circulated internally: round 3 was not trained on 247,049 rows. That number is the emotion-LoRA corpus of I.12.2, built the same evening by a different selection script, and the coincidence of dates is the likely source of the conflation. Round 3 trained on corpus_x. This checkpoint — SFT3 — is the Instrument of Part II.

I.9.6 The validation series, in full

I.9.7 Why that series is nearly useless, and what it is good for anyway

The whole run moves the held-out loss by 0.0032 nats, and it moves the wrong way for the first three evaluations. Against a model whose negative log-likelihood is about 54 nats per position — each supervised position carries 13 channels (I.3.1) — that is a relative change of six parts in a hundred thousand. The reason is structural and it applies to every short fine-tune in this project. The loss is dominated by predicting audio codes, and a change to the prompt format changes almost nothing about how hard the audio is to predict. The whole intervention — putting a parenthesised phrase back in front of each segment — is a few dozen text tokens against several hundred audio frames per example. A metric that averages over both cannot see it. So the honest reading of the validation series is: it confirms the run did not break, and it says nothing whatever about whether the run achieved its purpose. It is a smoke test, not a result. Every claim in I.9.9 comes instead from generating 320 clips and scoring them.

I.9.8 What actually moved in the weights

The export step compares every tensor against the base checkpoint, which is a cheap check that a run did what it claimed and touched what it should have. Round 3 moved more of the model than round 2 did, both in count and in magnitude — 378 tensors against 375, and a peak relative change 31 % larger (2.740e−02 vs 2.098e−02) — which is consistent with the higher learning rate and with the fact that round 3 introduced a new token pattern into every sample rather than merely reshuffling an existing one. Roughly 47 tensors were untouched by both runs, which is the expected shape for a fine-tune that leaves embeddings and certain normalisation parameters alone. This is a sanity check, not a result; it is recorded because a later run did report a suspicious tensor count, and it was a bug (III.2.9).

I.9.9 Round 3 — training the directions in fixes the collapse

Both models on the same 320 prompts, all carrying inline directions:

The collapse is gone. Word error rate falls from 0.447 to 0.099 — a factor of 4.5 — and it falls on both languages, from 0.574 to 0.070 on English and from 0.320 to 0.127 on German. Timing improves at the same time (median duration error 0.100 → 0.080 s, and 100 % of clips inside half a second against 92.8 %), and so does burst realisation (0.224 → 0.356) and burst hit rate (0.516 → 0.666). One 31-minute run on 32 nodes bought all of that. It is worth being precise about what was and was not demonstrated. Round 3 did not teach the model a new capability. It taught the model a format — and the size of the effect is a measure of how completely a transformer will fail on a prompt shape it has not seen, even when every individual element of that shape is familiar. The directions round 3 trains on are the same directions round 1 trained on. Two epochs of not seeing them was enough to make them poison.

Emotion control did not improve. SFT-2's higher emotion number (0.5496 against 0.5095) is not usable as a comparison: it was producing broken audio at word error rate 0.45, and broken audio scores on emotion heads — a screaming-noise clip reads as high arousal. Rescored on the zero-point-corrected percentile scale of I.6.7, with no regeneration, the picture is worse and more honest:

4.4 % of clips land in the band that was requested. That number, not the composite reward, is the one that states the problem: asked 320 times for a specific intensity, the model delivers it fourteen times.

I.9.10 A check that had to be done before considering a retrain

If the zero-point defect had also corrupted the training labels — the emo_strength percentile from which each direction's intensity adverb is derived — the whole round would have needed redoing. It did not. If the labels were a floor artefact, heads with a high zero point would win almost always. They do not: the winning head has a real signal (raw > 0.01) on 99.9 % of rows, including 99.9 % of the rows labelled at percentile ≥ 0.90, and the 40 heads are evenly represented at 3–4 % each. Repeating the supervised round on corrected labels would produce the same labels.

I.9.11 The standing hypothesis for why intensity does not move

The training subset is 59.5 % extreme and 6.8 % faint. The model sees “overwhelmingly” almost always and “faintly” almost never — so there is little contrast from which to learn that the word controls anything. This is the same failure as bug 1 of I.9.4, one order of magnitude weaker: bug 1 would have given the model one intensity word on every row; the corpus as shipped gives it four words in a 59/24/10/7 mix. The full corpus holds roughly 290,000 faint and 400,000 moderate rows, so a band-balanced selection of the same size is feasible. This remains untested as of 9 September and is the most obvious unexplored lever in the project (IV.2).

I.10 Preference tuning: seven runs, four corpora, and one metric that lies

Preference tuning is the objective this project has spent the most compute on, and it is the one that produced the DPO adapter in the agent's shipped stack (weight 1.0, Part II.3). It has also produced a completely consistent negative result on the capability it was aimed at. This section covers all of it: how the objective is configured, every run that was launched, how each of the four preference corpora was constructed and what each was meant to fix, which checkpoint is best and why, and what the training telemetry does and does not say. Source: [TR-0829] §11; corpus builds in [P§6–§9].

I.10.1 How DPO is set up here

DPO trains on pairs: a chosen audio sequence and a rejected one for the same prompt, pushing up the log-probability of the chosen relative to the rejected. Two configuration choices carry most of the weight, and both were arrived at by getting them wrong first.

I.10.2 Every preference run

Seven runs exist on disk. Two of them are not results — one trained on the wrong base and is kept deliberately (III.2.9), one was superseded before it was evaluated — and they are listed anyway, because a table of five successes is a different document from a table of seven runs.

None of the three long runs finished an epoch. All three hit a six-hour wall clock somewhere between a third and two thirds of the way through, and in all three cases the best checkpoint was the last one taken — which means nobody knows whether the remaining half helps, and it is an open task rather than a finding (IV.2).

I.10.3 Corpus v1, and 88.9 % of it going missing

The preference corpus has five pair families. The rejected side is constructed, not sampled — the design decision that makes the whole approach cheap, since no second generation pass is needed:

Version 1 of that corpus came out at only 236,166 pairs across 128 shards (1.7 GB), and the reason is a good example of a cost that gets stated rather than buried: the builder looked word timings up in the supervised corpus, which contains only the rows selected for supervised training. Every voice-profile pair whose chosen clip was not in that selection was dropped — 946,266 pairs, 88.9 % of the emotion arm — and because the length arm is capped to the emotion arm's size, the whole corpus shrank with it.

A correction to an earlier figure that propagated through the protocol and through the 26 August report: v1 is 236,166 pairs, not 236,656. The parquet shards and the build log both give 236,166, and the family counts force it arithmetically — 118,328 on the emotion arm against 117,838 on the length arm. It appears to be a digit transposition that was then copied forward.

I.10.4 Corpus v2 — the timings were never missing

They were in the annotation product's tar files all along: <uid>.json members carrying MMS-FA word alignments in exactly the {w, s, e} form the renderer expects. A sidecar keyed by clip id was built in three stages on a single node — plan (3,451,531 references → 2,146,797 distinct ids over 2,000 shards, 85 s), bursts (1 s per task), words (84 s per task). Coverage after the sidecar: word timings 2,137,175 of 2,146,797 (99.55 %); burst spans 2,146,797 of 2,146,797 (100 %).

The v2 drop count splits as 2,981 rows with no annotation at all and 803 with an annotation but no usable duration; the protocol labels the 803 as the no-annotation count, which is the wrong one of the two. The 11.9 % pre-sidecar coverage figure quoted elsewhere could not be reproduced — the v1 auxiliary table was overwritten by the rebuild — so it is stated here as a protocol claim rather than a measurement. The 57 : 43 ratio is a deliberate departure from the original brief's strict equality, kept because the length arm has already done its work — timing is solved — and the emotion arm is what remains unsolved.

I.10.5 The classifier-free-guidance corpus

Every contrast built so far pits an intense clip against a mild one — which a model can win by being generically expressive, without ever reading the instruction. The classifier-free-guidance (CFG) construction tries to make that impossible. (This CFG is a *data* construction; the inference-time classifier-free guidance the agent uses as a Tool is a different thing, Part II.3.2. The name collision is unfortunate and is kept because both artefacts carry it.)

• Per voice profile and per dimension (40 emotion heads plus the VoiceNet axes), rank that voice's clips by the measured value — not the prompt, not the intended bucket.
• Pair a top-1 % clip with a bottom-1 % clip of the same voice, with durations within ≈10 %. (A DPO pair shares one prompt and therefore one length budget; a length mismatch would teach length instead of intensity.)
• Emit each pair twice with the roles flipped. Under a “high” instruction the intense clip is chosen; under a “low” instruction the mild one is. Both clips appear as chosen and as rejected, so the model cannot win by a global preference for intense audio and must read the instruction.
• Remove the words. The prompt keeps the direction, the duration and pause tags, the burst tags and the frame budget; the spoken text becomes ..., so the preference cannot be decided from the words.

A rendered CFG prompt, verbatim: <user_inst> - Reference(s): <|audio|> - Instruction: GENERAL: A voice; timbre is very dark - Tokens: 101 - Quality: None ... - Text: (very dark, and hold it there; otherwise exactly as this voice normally speaks) [5.5 seconds duration] ... [0.5 seconds pause] [2.0 seconds duration] ... </user_inst>

The two CFG halves are not merely balanced to within a tenth of a percentage point, as the 26 August write-up of this build reported — they are exactly equal, 237,209 each, which is what a correct flipped-role duplication must produce and is therefore a stronger check than the approximate one. The render-time recount confirms it (Appendix A.6: German 247,350 / English 227,068 pairs). The prompt-format hash moved to 073aeb09dc923376.

I.10.6 The contrastive-families corpus

Two further families were added on top of the CFG corpus, aimed at two measured defects rather than at a general idea. I.13 gives their construction, their selection rules and the per-family learning curves. In summary: p2_emox (309,128 rows) pits intense against intense — same speaker, matched length, emotion A against emotion B, with the instruction naming A — so that turning intensity up is worth nothing and only reading the instruction pays; p2_len (59,389 rows) keeps speaker, text and emotion fixed and varies only the duration. Mixed onto the CFG corpus: 2,327,904 + 368,517 = 2,696,421 pairs, the new families 13.7 %.

I.10.7 Which checkpoint is best, and why

All rows on the same 80 prompts, 4 completions each, inline delivery directions, zero-point-corrected percentile scale, plateau yardstick — so they are comparable to each other and to nothing else. (The full 18-run ranking table with every arm of I.11–I.12 is Appendix A.1.)

Step 5022 of the contrastive-families run is the best model this line produced, and the case for it rests on three things rather than on the composite reward alone, since the top four rows span 0.0059 in reward and this evaluation cannot resolve ±0.01:

• It is the only preference-tuned model whose emotion percentile beats plain SFT-3. 0.3541 against 0.3494. Every other DPO checkpoint on this table is below the supervised baseline on the metric all of them were built to raise — v2 at 0.3401, CFG at 0.3373. The two new families were built for exactly this and it is the one thing they changed.
• Its word error rate beats the supervised baseline too. 0.0977 against 0.0987. Before the CFG corpus, every preference run paid for its gains in intelligibility; two corpora in a row have now stopped doing that, and the contrastive-families run does it while also raising emotion.
• It does not give up the structural gains. Burst realisation 0.4180 and hit rate 0.762, against the supervised 0.3564 and 0.666. It is slightly below the CFG checkpoint on both (0.4271 / 0.772) — that is the one axis on which the previous best is still ahead, and it is the honest cost.

Its neighbour at step 3906 is a statistical tie on reward, is better on word error rate (0.0916, the lowest of any model in the project), and is worse on emotion (0.3478) and on burst realisation. Step 5022 was chosen because emotion is the unsolved capability and 3906's WER advantage of 0.006 is inside noise. Both are published. What none of them did. The target band is 0.90–0.98. The best model reaches 0.354. Four preference corpora, seven runs and something over a hundred node-hours have moved that number from 0.3494 to 0.3541 — a gain of 0.005 on a scale where the goal is 0.55 away. The corpora are not what is wrong, and I.10.9 argues that the objective may be.

I.10.8 rew_chosen: a metric that is uninformative across runs and anti-informative within one

The DPO trainer logs two reward terms per evaluation: rew_chosen, the implicit reward on the preferred sequence, and rew_rejected, the same on the rejected one. A healthy preference run should raise the first while lowering the second. Every run, without exception, ends with a negative rew_chosen:

The trainer emits an explicit <-- DAMAGE: policy suppresses the CHOSEN audio line and writes "healthy": false into eval.jsonl whenever this happens. It has happened on every evaluation row of every preference run in this project. Within a run the metric moves the wrong way as the model improves. From dpo_sft3:

Preference accuracy climbs from 0.865 to 0.981 while rew_chosen gets three times worse — and step 864 is the checkpoint that generates best. The policy is suppressing the chosen audio while learning the ranking: it makes the bad audio catastrophically less likely and the good audio slightly less likely. Both go down; one goes down less. That satisfies the objective exactly and is not what anyone wanted.

I.10.9 Why none of it moved emotion — one explanation refuted, three standing

Refuted: “the high side was only locally intense.” The obvious explanation is that pairing within a voice makes “high” a relative target — a voice profile that never gets angry has a top-1 % anger clip that is barely angry. Measured over 224,750 voice-profile rows in 500 voices, each voice's own top 1 % per emotion head scores a median 0.995 on the global zero-point-corrected scale, and 99.7 % reach the 0.90 band. The chosen side was genuinely extreme, and the training signal pointed at real extremes.

• Recognition is not production, and the telemetry supports it. DPO raises the chosen sequence's likelihood only relative to the rejected one. At the CFG run's step 4476, rew_chosen is −1.12 against rew_rejected −46.6. The policy is not making intense audio more likely; it is making mild audio catastrophically less likely under a “high” instruction and intense audio slightly less likely. Pushing everything down, one thing less, does not move the sampling mode. Ranking two fixed recordings is a discrimination task and the model solved it (0.5625 → 0.9750). Sampling draws from wherever the probability mass actually sits. *Test:* measure the absolute log-probability of held-out intense audio before and after a preference run. If it fell, the fix is an absolute anchor — a larger NLL weight than 0.013, or a β below 30. This costs one evaluation pass and is the cheapest decisive experiment in the project. It is still not instrumented (IV.2).
• The text-free regime never occurs at inference. The CFG and p2_emox rows are the only ones where the spoken words are absent; at evaluation every prompt has words. What generalises out of that regime is structural — the duration, pause and burst tags mean the same thing with or without words — and that is exactly the profile of the gains: burst hit rate 0.709 → 0.772, burst realisation 0.3929 → 0.4271, WER 0.1094 → 0.0950, all structural. The semantic mapping from an intensity word to an acoustic target has no lexical content to attach to when the direction sits next to silence. *Test:* the same pairs with the words present, only the requested level flipped.
• The symmetry that prevents drift also prevents push. Whatever gradient pushes towards intensity under “high” is mirrored by an equal gradient away from it under “low”. The net movement of the unconditional distribution is zero by construction. That is correct behaviour for a conditioning signal and useless for raising a ceiling. If the goal is a model able to go further than it currently can, symmetric contrast cannot supply it — only data at a higher ceiling, or an objective with an absolute target, can. *Test:* an asymmetric 70/30 mixture.

A fourth possibility has since arrived from a different direction and is recorded with the layer forensics (Part II.4): measured inside the model, the direction that means “more of this emotion” and the direction that means “higher measured quality” have a cosine of −0.746 at h20 and −0.954 at the last state before audio codes are chosen. If those two are close to opposite instructions, then any objective that rewards emotion and quality together is asking for something the representation makes difficult, and every reward in this project — including the agent's Best-of-N reward — does exactly that.

I.11 GRPO: five runs, and one structural negative result

GRPO generates G completions for the same prompt with sampling, scores each with the measurement models, normalises the rewards within the group (A = (r − mean) / std), and pushes the policy towards the completions with positive advantage. With one inner update per rollout the PPO ratio is exactly 1, the clipping term is inert, and the loss reduces to −(A · logp).mean(). The whole learning signal is therefore the spread of scores inside a group. Source: [TR-0829] §12; run logs in [P§10–§14].

reward = (0.40·r_emo + 0.40·r_qual + 0.20·r_burst) × wer_factor(WER) — r_emo the band score of the percentile reached on the requested emotion head against the requested band; r_qual = 0.5·percentile(vocal-burst blend) + 0.5·percentile(genuineness); r_burst the F1 over prompted vs detected bursts, weighted by duration match; WER from Whisper-large-v3-turbo against the script's own plain text.

I.11.1 Runs 1 and 3 were not a GRPO result

Both fell, and the cause was outside the optimiser. Run 3 was given prompts with inline delivery directions — the right idea — but the model had never been trained on that format, because round 2 had dropped directions by accident (I.9.2). It collapsed, and GRPO spent 2.7 hours optimising against a reward dominated by that collapse. The conclusion is not that GRPO needs tuning; it is that a reward computed on out-of-distribution prompts measures the prompt, not the policy.

I.11.2 Run 4 — the first run that neither deadlocks nor collapses

88 steps, 176 rollout groups, 2.00 h on one 4-GPU node, from SFT-3, KL β 0.02, corrected percentile scale. skipped = 0 for the entire run: the rank-divergence fix (III.2.1) held.

Burst realisation rises 0.561 → 0.669 and WER falls 0.128 → 0.098. Emotion is flat within noise. (An earlier reading of single steps had suggested an emotion collapse; the aggregate over 176 groups does not support it, and per-step numbers of two groups each are far too noisy to read as a trend — a small methodological correction that is recorded rather than quietly dropped.)

I.11.3 The decisive measurement

Splitting run 4 by the band the prompt asked for, where emotion is the band score of the percentile achieved against the percentile band requested:

I.11.4 Run 5 — four fixes for that, and why they made it worse

Four levers were built and all four default off, so run 4 stays reproducible:

• A monotone ramp. Replace the flat floor of the band score with a straight ramp from 0 at percentile 0 to 1.0 at the band's lower edge, so the slope is non-zero everywhere below the target. Verified before the run: for band extreme, percentiles 0.20 / 0.35 / 0.50 score 0.000 / 0.000 / 0.000 under the old shape and 0.204 / 0.357 / 0.510 under the new one.
• A WER-free allowance. WER below 0.15 costs nothing; the old slope applies to the excess only. The old form taxed every gain in emotion that cost a hair of intelligibility, and the ASR hypothesis of an intensely delivered line genuinely is a little worse than that of a flat one. (The agent's Best-of-N carries the same idea as BON_WER_KNEE, set to 0 in production — Part II.6.)
• A band curriculum, starting at 5/30/40/25 (extreme/intense/moderate/faint) and interpolating to the briefed 25/25/25/25.
• Dump the best completion of each group (most emotional with WER ≤ 0.20) with its prompt and codes.

On the strict evaluation yardstick, run 5 is the worst of the four models on emotion (0.3275 against SFT-3's 0.3494 — see I.12.3). The training-time gain was the mechanical artefact it had been predicted to be: lever 1 inflates the number by paying for partial progress, lever 3 inflates it by asking easier questions. The realised band mix confirms the second: v4 ran 0.28 moderate / 0.28 intense / 0.22 extreme / 0.21 faint, v5 ran 0.36 / 0.30 / 0.15 / 0.20 — the curriculum moved the run onto the easier half and the run ended before it came back. Hypothesis, stated as a hypothesis: a ramp that pays for partial progress supplies a gradient but removes the incentive to arrive. A completion that moves slightly toward the target now collects reward without reaching the band, and the strict yardstick only counts arrival. Levers 1 and 3 push in the same direction and it is the wrong one. A ramp that is steeper near the band, or a rank-within-group reward that keeps the target absolute, would test this.

I.11.5 What GRPO did produce: data

Lever 4 is the one that paid. Run 5 saved the most emotional usable completion of every group: 204 clips at median emotion percentile 0.641 and median WER 0.049, median 148 frames — 73 moderate, 61 intense, 41 faint, 29 extreme. High emotion and clean speech in the same clip is the combination that has been missing from every evaluation in this project. The model can do it; it does not do it on command. Two hours of rollout on one 4-GPU node therefore yields supervised training material at roughly 100 usable clips per GPU-hour — and best-of-N supervised training on that harvest was, on 29 August, judged a more promising route than further reward shaping. The agent of Part II is the inference-time version of the same conclusion: select, do not push.

Hyperparameters, for reproduction: group size 16 (32 in run 3), 2 groups per optimizer step, micro-batch 2 completions, LoRA rank 32 / α 64 (68.7 M trainable of 4.27 B), learning rate 2e-6 constant after 20 warmup steps, KL β 0.02 with the k3 estimator against the same model with the adapter disabled (so no second copy in memory — measured KL stayed at 2–3e−4 and the constraint was never binding), sampling temperature 1.0 / top-p 0.95 / top-k 50, max 340 new frames (27.2 s), prompt pool 4,000 rows, checkpoint every 25 steps, 2-hour stopping rule. Prompt mix: 50 % carry vocal bursts, 35 % carry two emotions, 50 % carry a reference clip, 85 % carry a delivery direction. Throughput: one 4-GPU node produces 45–52 optimizer steps per hour at G = 16, i.e. 1,440–1,680 generated and fully scored clips per hour. Loading the three measurement models costs about 40 s per rank.

I.12 Specialist adapters: the only thing that moved intensity before the agent

I.12.1 Why a quality gate is not optional

Measured on a 360-clip listening sample decoded from the corpus's own audio codes — one clip per emotion per tier ([TR-0829] §13.1):

Genuineness falls as intensity rises, and both quality axes sit below the corpus median of 0.50 throughout. The most intense recordings are on average the least genuine. An adapter trained on intensity alone would be taught to overact.

I.12.2 The general emotion LoRA: two arms and one gate

• Emotion arm — the top 1 % of any of the 40 heads, on the zero-point-corrected percentile scale.
• VoiceNet arm — several VoiceNet axes imply strong emotion without naming one, and the emotion heads do not reliably fire on them. Two-sided (top and bottom 1 %) for AROU, VALN, ARSH, VALS — high valence is elation, low valence is despair, and both are extremes. Top only for TENS, VOLT, VULN, EXPL, RANG, EMPH, VFLX, and the ranting / ASMR / whisper / cartoon / drama style axes.
• Gate — normalised vocal-burst blend AND genuineness both above the corpus median.

Two things fall out of this. First, the VoiceNet arm added about 40 % more material than the emotion heads alone would have found, and the axes that contributed are the affective ones — high valence (16,726 rows qualifying on that axis alone), explosiveness (14,437), arousal shift (9,890), valence shift (8,164), velocity flux (4,680), low arousal (3,323), low valence (2,383), high arousal (1,972). The stylistic axes — ranting, ASMR, whispering, cartoonish — contributed nothing in the top eight: their top 1 % was either already in the emotion arm or failed the quality gate. Second, the source mix flipped. The gated selection is 65.4 % real speech, 34.6 % voice profiles — the reverse of the ungated listening sample (273 synthetic clips to 87). The gate removes synthetic extremes disproportionately, which supports the reading that the synthetic “top 1 %” is often overacted. [TR-0826] §12.6 adds the bucket caveat: the largest VoiceNet buckets are S_ASMR bottom (23,817) and S_CART bottom (18,934) — “not ASMR” and “not cartoonish”, the unremarkable default, whose high genuineness (0.87) reflects that. Only the four two-sided affective axes have a meaningful lower tail.

I.12.3 Result: rank 16 wins, and it is the first thing to move the number

Three adapters on the same 247,049-row corpus, 5 epochs, cosine 1e-4 → 5e-6, α fixed at 2 × rank so the effective scaling is identical and rank is the only variable, global batch 256 (4 × 2 × 32 on 8 nodes), 4,395 steps, 2 h 35 to 2 h 40 each. Base SFT-3, evaluated alone on it.

The complete ranking of every evaluation on the corrected percentile scale — eighteen runs, all on the same 80 prompts, 4 completions per prompt, the strict plateau yardstick — is Appendix A.1. It replaces every partial comparison published earlier, including the eight-row table of [TR-0826] §12.3. Three readings of it:

• Two things raise the emotion percentile above the supervised baseline, and they do it in opposite ways. The rank-16 emotion adapter reaches 0.3572 against 0.3494 — the largest lift on the table — but buys it with word error rate, 0.0987 → 0.1239. The contrastive-families preference adapter reaches 0.3541 and lowers word error rate to 0.0977. One is more extreme data; the other is a better-constructed contrast. The target band remains 0.90–0.98 and neither is close. With 320 clips per arm, a difference of 0.008 in a percentile average is not distinguishable from seed noise; [TR-0829] itself says the evaluation cannot resolve ±0.01.
• The bare SFT-3 base ranks eleventh of eighteen. Seven of the ten arms that beat it are preference-tuned variants. Since the CFG corpus, preference tuning has stopped costing intelligibility: the top three rows all have a lower word error rate than the supervised baseline, which nothing before them managed.
• The ordering r16 > r64 > r32 is not monotone in rank, so this is not yet a capacity story; with one seed per rank the r32/r64 difference is within what seed noise could plausibly produce. What the ranking does support is that more capacity is not the lever here — the largest adapter is not the best one.

Six further evaluations exist and are deliberately excluded from that ranking, because they predate the percentile correction (I.6.7): eval_sft2_dir, eval_sft2_dirv3, eval_dpo2_dir, eval_grpo1_dir, eval_grpo3_dir and eval_sft3_dir_UNCORRECTED_ECDF. Their word error rates (0.45–0.51 for the SFT-2-based ones) are comparable and genuinely bad; their reward and emotion columns are not.

I.12.4 Merge weight: two sweeps, two answers, and why both are right

An obvious next question is whether merging an adapter more strongly pushes emotion further. Two sweeps have been run, and they came out differently. Reported together, because the difference between them is the finding — and because the agent's shipped stack (Part II.3) sets adapter weights above 1.0 on the strength of sweep B.

Sweep A — one general adapter, band intense, 80 prompts: no trend. The single pooled emotion LoRA (rank 16, trained on all 247,049 gated rows) evaluated at seven merge weights on the standard harness, 320 clips each:

Emotion percentile spans 0.3315–0.3572 with no monotone structure and a peak at the trained value. Reward spans 0.015 — inside what these evaluations show between arms that ought to be equivalent. On this axis, with this adapter, at this band, merge weight is not a dial for intensity. ([TR-0826] §12.4 reported this sweep first, with six weights and the note that the numbers were “measured today and reported here first”.)

Sweep B — 31 per-emotion adapters, band extreme, with bursts: a clear monotone rise. The second sweep takes the per-emotion bucket adapters of I.12.5, each on prompts naming its own emotion at the extreme band with vocal bursts requested, and scales each adapter's contribution over six weights. Design details that make it a strong measurement: the same eight prompts are used at every weight for a given adapter, the torch seed is reset to a fixed value before each weight, and weight 0 is the same loaded model with the LoRA scaling zeroed rather than a separately loaded baseline — so a difference cannot come from anything but the adapter. 31 adapters × 6 weights × 8 clips = 1,488 scored generations.

Emotion percentile rises monotonically from weight 0.5 onward and is +0.084 higher at weight 2.0 than at weight 0 (paired over 248 clips, bootstrap 95 % CI [+0.048, +0.123]); at weight 1.5 the paired gain is +0.064, CI [+0.030, +0.098]. 27 of 31 adapters are higher at the top weight than at zero — the three that are not are Helplessness (−0.185), Bitterness (−0.104) and Contempt (−0.024) — and the fourth, Jealousy and Envy, is the dead metric of I.6.4 and reads 0.000 everywhere. At weight 1.5 the split is 25 up, 5 down, 1 tied. Per-adapter argmax weight: 2.0 for 13 adapters, 1.5 for 7, 1.0 for 6, 0.5 for 2, 0.25 for 2 and 0.0 for 1. Timing is untouched at every weight — mean absolute duration error stays inside 0.115–0.133 s across the whole sweep, so overdriving an adapter does not cost the capability that was actually solved.

• Genuineness and blend are ceiling-bound. Their medians are 1.0 at every weight; blend percentile hits 1.0 in 225–239 of 248 clips and genuineness in 190–208. The rise from 0.817 to 0.880 and 0.925 to 0.969 is real but is happening in the top few percent of a saturated scale, and should not be read as a large quality gain.
• Burst realisation does not rise — it peaks at weight 0.5 (0.3035) and falls to 0.2227 at 2.0, and is exactly zero on 104–129 of 248 clips at every weight. Overdriving helps emotion and does not help bursts.
• The whole word-error story is an English-side story. Split by language: German WER is 0.059 at weight 0 and 0.059 at weight 2.0 — flat — while English runs 0.270 to 0.304. The German half of the sweep shows the emotion gain with no intelligibility cost at all. (In the agent era the sign flipped: with German burst cues the agent's WER was 0.267 against 0.000 for English — Part II.7. The two measurements are on different prompts, different adapters and different ASR models, and the contradiction is recorded in III.5 rather than resolved.)

Why the two sweeps disagree, and which to believe for what. They are not the same experiment. Sweep A scales a single general adapter trained on all 154 axes at once, on the intense band, against a prompt set that mostly does not request bursts; sweep B scales per-emotion specialists on the extreme band, on prompts that name the adapter's own emotion and request a burst. The reconciling reading is that a general adapter has no single direction to overdrive — scaling it amplifies an average of 154 directions, which is close to noise — while a specialist adapter has exactly one, and scaling that one works. Both results stand. Coverage note: sweep B was launched when 31 of the 40 per-emotion adapters existed. Amusement, Embarrassment, Intoxication, Relief, Sadness, Sexual Lust, Shame, Thankfulness and Triumph are absent from it. Re-running it would cover all 40; as of 9 September that has not been done.

I.12.5 One adapter per emotion — and the selectivity problem

The natural next step is 40 adapters, one per emotion, rather than one general one. All 40 were trained (rank 16, α 32, 5 epochs, cosine 1e-4 → 5e-6, 34.4 M trainable parameters, each on the gated top 1 % of its own emotion, all finishing with zero non-finite batches). Training set sizes run from Sourness at 2,663 rows to Intoxication at 15,725; step counts from 3,330 to 19,660; wall time from 42 to 224 minutes per adapter. These are the emotion adapters of the agent's shipped stack (Part II.3). Source: [TR-0829] §13.5.

Seventeen of them were measured against the best general checkpoint of the time — SFT-3 + CFG-DPO — on identical prompts, identical sampling and identical seeds. The seventeen are Affection, Anger, Astonishment/Surprise, Awe, Bitterness, Concentration, Contempt, Contentment, Disgust, Elation, Emotional Numbness, Jealousy and Envy, Longing, Malevolence/Malice, Pleasure/Ecstasy, Pride and Sourness. That set is not a chosen subset: the generator shards adapters across ranks with names[RANK::WORLD], and the adapter directory was still growing while the workers launched, so ranks duplicated each other and skipped others. Seventeen is where the job landed.

The design of that evaluation is the interesting part. A specialist adapter should do two things: make its own emotion stronger when it is asked for, and change nothing when it is not. So each adapter gets ten prompts naming its emotion at the intense band and ten prompts naming no emotion at all — the corpus's own neutral direction. A rise in the first is the adapter working. A rise in the second is the adapter leaking into plain speech.

So the adapters do push their emotion — and they push nearly as hard when nobody asked. They also cost about half again as much transcription error and about a fifth of the burst realisation. As they stand they are not a drop-in improvement. They are tinting, not control.

Word error rates here are for the emotion-naming (matched) prompts only. Different adapters draw different slices of the prompt pool, so cross-adapter comparison of the WER columns is not prompt-matched. The prompts are drawn from three source corpora — the run log's “three voices” label is a fallback string, not three verified speakers — so the design controls for corpus, not for speaker identity, which is weaker than it looks. Jealousy and Envy reads 0.000 everywhere because of the key mismatch of I.6.4, not — as [TR-0826] §12.5 said — because of the percentile correction. The agent-era combination study (Part II.3.5) re-measured the emotion adapters' contribution on 399 clips and found a larger, resolved main effect (+0.077 target-z, t = 2.81) — with a different prompt set, different ears and the full shipped stack around it.

I.12.6 Is there enough data for a LoRA per bucket?

Feasible — a median of 6,700 rows is comfortable for a rank-16 adapter — but badly skewed, and skewed by exactly the effect that motivated the gate. The four smallest emotion buckets are Sourness (2,663), Pleasure/Ecstasy (2,976), Contempt (3,105) and Awe (3,247), and their genuineness median in the top 1 % is 0.23–0.30: the most intense recordings of those four are almost entirely inauthentic, and the gate removes three quarters of them. Intoxication (15,725 rows, genuineness median 0.804) and Amusement (13,272, 0.774) have ample genuine material. A 40-adapter programme therefore needs either unequal training sets or a relaxation to the top 2 % for the four problem emotions.

I.12.7 The VoiceNet axis adapters

The same bucket machinery was pointed at the VoiceNet axes on 26 August at 16:47. The policy is the one I.12.2 established: both tails for AROU, VALN, ARSH and VALS; the top only for the rest — the survey found that the largest “bottom” buckets of the style axes are simply ordinary unremarkable speech, so there is no such thing as an extreme absence of ranting. Twenty candidate buckets went in; three were dropped by a row-count floor of 1,500 applied after the top-1 % tail selection and the blend-and-genuineness quality gate, and 17 came out: 146,175 rows in total, median 8,216, minimum 1,922, maximum 22,849.

The three dropped axes are exactly the ones where the quality gate bites hardest: high pitch range keeps 1,147 of 31,453 candidates, whisper-talk 939 of 31,452, and cartoonish 680 of 31,450 — a 2.2 % survival rate for cartoonish, whose genuineness median in its own top 1 % is 0.109. That is the I.12.1 finding again, in its sharpest form. Training of those 17 adapters started at 16:51 on 26 August (rank 16, α 32, 5 epochs, AdamW lr 1e-4, cosine with 10 % warmup, batch 4, non-DDP, one bucket per GPU stream). They were not published as of 29 August; the delivery-axis adapters the agent ships (SFT3_VN_LEVELS, Part II.3) are this family's descendants, and their measured WER cost (+0.0546, t = 6.91, n = 160) is in Part II.3.5.

I.12.8 Every adapter is trained on SFT-3 alone, and here is why

All 40 emotion adapters, all 500 voice adapters and the 17 VoiceNet adapters are trained on SFT-3 — never on a merged SFT-3 + DPO checkpoint. The reason is structural, not a preference: the DPO adapter targets audio_lm_heads.*, and those twelve tensors are weight-tied to audio_embeddings.*. They are literally the same tensor. Merging the head delta therefore rewrites the embedding that produced it, and the model is corrupted (III.2.2 gives the measurement). The adapters are stacked at inference instead, each with its own scaling factor. The published recommendation on 29 August was DPO at 1.0, the voice adapter at 1.0 and the emotion adapter at 1.5 — the emotion weight from sweep B; the voice weight has not been swept and 1.0 is simply its trained value. The agent's shipped stack (Part II.3) keeps DPO 1.0 and voice 1.0 but sets emotion at 1.0, with the three quality adapters and quality_dpo at 1.5 around it — a different recipe, chosen by the combination study rather than by sweep B.

I.13 Two new contrastive pair families — and the model they produced

These are aimed directly at two measured defects — the leakage of I.12.5 and an emotion-blind length signal — and they are the only intervention before the agent that raised the emotion percentile above the supervised baseline without costing word error rate. The adapter they produced is the DPO of the agent's stack. Source: [TR-0829] §15; the earlier build in [TR-0826] §16.2.

I.13.1 What they are

Family A — emotion-contrastive, for selectivity. Every contrast built before this pits an intense clip against a mild one, so a model can win by being generically expressive — which is exactly the 1.4 : 1 tinting the per-emotion adapters showed. Family A pits intense against intense: same speaker, same length, emotion A against emotion B, with the instruction naming A. Turning the intensity up is then worth nothing; only reading the instruction is. The builder's own docstring states the motivation in one line: “A ratio of 1.4 : 1 is not control, it is tinting.” The selection rule is strict on both sides. The chosen clip must sit at percentile ≥ 0.90 on the head being asked for and have that head as its own top emotion; the partner must be ≤ 0.50 on that head and ≥ 0.90 on its own top head. So both sides really are intense, on different emotions. Lengths are matched to within 10 % of frames. Each base pair is emitted twice, the mirror naming the partner's own top emotion. Because chosen and rejected are different recordings with different words and a DPO pair shares one prompt, the spoken words are removed, as in the CFG construction.

Family B — length-contrastive, conditioned on the emotion. The existing truncation and continuation families know nothing about emotion, so they teach “the right length is better” in the abstract. Family B keeps the speaker, the text and the emotion fixed and varies only the length: chosen is the clip at its true duration, rejected is the same clip cut to 50–75 % or run to 125–150 %. Here the words are in the prompt, with the full timed script, because the timing tags are exactly what the rejected side violates. Both families take the emotion from the measured value, never from the prompt that generated the clip. Constants: HI_PCT 0.90, LO_PCT 0.50, FRAME_TOL 0.10, 8 emotion-contrastive and 6 length-contrastive pairs per cell, cut range 0.50–0.75, extension range 1.25–1.50, RNG seed 4711.

I.13.2 What was built

The base is the CFG corpus rather than the older one, and the mixer says why: because the CFG corpus produced the best checkpoint measured so far. Schema equality between the new rows and the base was checked field by field before the merge — 30 fields each, none missing, none extra — and all 342,284 referenced clips were found. An earlier build with a smaller per-cell budget produced 138,088 rows (118,304 emotion-contrastive over 19,884 cells, plus 9,953 truncation and 9,831 extension pairs over 500 voices — [TR-0826] §16.2) and was superseded.

I.13.3 What the run did

The run is dpo_sft3_p2: 32 ranks over 8 nodes, LoRA rank 64 (137.4 M trainable of 4.267 B, 3.220 %), β 30, chosen-NLL anchor 0.013, learning rate 1e-6, 10,331 steps per epoch with 1,033 warmup, 256 preference pairs per step, arm-balanced sampling. It ran at about 62 pairs per second and hit its six-hour wall clock at step 5,126 of 10,331 — half an epoch. Checkpoints were kept at 3906, 4132, 4355, 4579, 4800 and 5022.

The model learned the harder contrast. Emotion-contrastive accuracy rises from 0.864 to 0.938 across the two halves and sits at 0.951 post-warmup, above the older CFG families' 0.82–0.84 at the corresponding cumulative point. The length-contrastive family saturated at 1.000 over the second half — the task is solved and it is no longer informative, consistent with timing having been solved since round 2; it could be dropped from a future mix. And the run failed the health metric, exactly as every previous DPO run has: the mid-run evaluation at step 2582 reported val loss 0.7277, preference accuracy 0.9805, rew_chosen −1.0015 and rew_rejected −38.851, and the trainer wrote "healthy": false.

I.13.4 The result: a new best model

Because both this run and the CFG run use a global batch of 256 preference pairs, equal step numbers mean equal numbers of pairs seen — so step 5022 here and step 4912 there are a nearly matched comparison, and the difference is the corpus rather than the budget. The emotion percentile is 0.3541 — the highest of any preference-tuned model in this line, and the first one above the supervised baseline of 0.3494. The gain is +0.005, against a requested band of 0.90–0.98. Both statements are true and the second one is the important one. I.10.9's first standing hypothesis predicts precisely this: a preference objective can teach a model to recognise which of two recordings matches an instruction without teaching it to produce one, and high emotion-contrastive accuracy is evidence of the former only. The adapter is published as laion/moss-va-sft3-dpo-lora-p2. The remaining half epoch was never run; whether it helps is unknown.

I.14 Three inference-time levers, measured end to end (28–29 August)

Sections I.8–I.13 are about making the model learn to follow an intensity instruction. That programme reached a wall: five GRPO runs and seven preference runs did not move emotional intensity. This section is about the other approach — leaving the trained weights alone and pushing the model at the moment it speaks. Three such levers exist, all three were measured across every adapter and dimension the project has, and the picture they produce is consistent enough to be a property of the model rather than of any one method. This is the study that the agent of Part II was built on; Part II.3 re-measures the same levers inside the shipped stack. Source: [TR-0829] §24 (added 28 August); the underlying study pages are cited in [P§40–§53].

Scale: 7,419 CFG cells / 29,676 clips, 5,740 adapter dose cells over 79 adapters, 10,459 combination cells / 31,377 clips over 60 attributes. Every figure is paired — the same prompts with and without the intervention, a prompt's samples averaged before averaging across prompts, so n counts prompts. Word error is Parakeet-TDT-0.6b-v3 (I.14.8).

I.14.2 The central result: manner is writable, feeling is not

Three mechanisms with nothing in common — a weight patch, a hidden-state nudge, a logit extrapolation — all move the delivery axes several times further than the emotion heads. The ordering matches what a probe can read from the same activations: R² 0.85 for the 57 VoiceNet dimensions against 0.47 for the 40 emotion heads (Part II.4). The model has a rich, writable representation of how a line is delivered and a thin one of what it feels, and no amount of inference-time pressure changes that ratio. Decodability nevertheless fails to predict steerability per attribute: Anger (R² 0.523) gains +0.088 from a steering push while Elation (0.503) gains +0.311, and Jealousy_and_Envy — the least decodable head at 0.209 — gains +0.070. Reading a layer and writing to it are different operations.

I.14.3 The dose-response of the adapters

79 adapters swept at six merge weights. 33 have a usable weight; none is harmful at any weight. Every failure is a failure to move the target, not a broken guardrail.

Two cautions the table cannot carry alone. “Below resolution” is not “no effect”: the 24 quiet emotion adapters still move pooled (+0.024 at w = 1.0, t 2.70); ten prompts each is not enough to prove it one at a time. And 20 of the 71 burst adapters cannot be evaluated at all — trained almost entirely on isolated snippets under two seconds, with no speech row in the corpus carrying the class. The most actionable number: raising the delivery clamp from 0.75 to 1.5 buys +0.375 on the target axis (t 5.18, better on 15 of 17) for a word-error change of +0.003 (t 0.55, i.e. none). The clamp was chosen when several delivery adapters ran at once; it is not justified for one at a time. Note the contradiction with the agent-era measurement: with the full shipped stack around it, one delivery axis costs +0.0546 WER (t 6.91, n = 160, Part II.3.5). Both stand; the difference is the stack.

I.14.4 Guidance, and its true cost

Best guidance at a pre-registered threshold of Parakeet WER ≤ 0.20: g = 3.0 for emotion (+0.0254, t 2.18), g = 2.5 for delivery (+0.0569, t 3.12). Below g = 1 guidance actively hurts (−0.0370 at g = 0.5, t −2.56), the directional control the arm needed. Side effects are mild and partly favourable: genuineness rises with guidance on the emotion family (3.85 → 3.94), blend is flat, duration error unmoved; burst realisation falls slightly (0.414 → 0.392) and the delivery family pays intelligibility above g = 2 (share at WER ≤ 0.2 falls 0.828 → 0.744 by g = 3). g = 3.0 is the BON_GUIDANCE the agent ships (Part II.3.2).

The LoRA difference vector as a guidance signal. With SFT3+DPO-v2 as the unconditional branch and SFT3+DPO-v2+LoRA as the conditioned one, the same arithmetic applies at the logit level. At equal nominal factor, weight-scaling reaches further — but at 3.0 it hits WER 0.310 and collapses burst realisation to 0.287, where logit guidance stays at 0.111 and 0.400. At matched intelligibility the ranking reverses for the delivery axes (0.5827 against 0.5634) and ties for emotion. Two independent estimates of the adapter's own effect, through different code paths, agree to four decimals (+0.0878 vs +0.0879).

I.14.5 Combining the three

The levers are cumulative on the target and more than cumulative on the cost. Excess over the sum of the parts, over 60 attributes:

The super-additivity is specifically an emotion phenomenon (1.52); delivery axes are almost exactly additive (0.97) and the quality attributes saturate (0.42). The dominant interaction is steering × CFG (+0.208, t 6.30), and it is a wiring choice rather than a fact: steering both CFG branches instead of only the conditioned one costs 18 % of the effect and returns 0.209 of word error and 0.75 of genuineness.

The most useful single trick: subtract flatness rather than add feeling. Emotion and perceived authenticity are anti-correlated inside the representation at a cosine of −0.62 to −0.95, and causally a +0.10 emotion push costs −0.86 of genuineness out of 6. Pushing away from Emotional_Numbness at −0.10 returns +0.60 of genuineness (t 9.64, 67 of 80 prompts) at no cost in emotion when an adapter carries the emotion; on the steering-only version it removed essentially the whole cost (−0.86 → +0.04). It stacks with CFG at −0.05 (+0.68, t 10.94) but not at −0.10. Adding the genuineness vector back instead is the weaker but still favourable trade: 16 % of the emotion gain buys 56 % of the loss.

Two further results close open questions. Re-extracting the steering vector with the adapter loaded does not help: the adapter shifts the bucket mean by 3.8 % of its norm but shifts the shared reference nearly as much, so the direction survives at cosine 0.9958 — below the split-half noise floor at 60 of 60 attributes. It contributes a small correctly-signed extra only when the adapter is also loaded at generation (+0.095 z, t 2.16, against −0.041 without). Conversely steering is not made redundant by the adapter: it adds +0.864 SD (t 12.12) on top of the adapter-only cell.

Where the recipe lands. A balanced operating point exists for 53 of 60 attributes, a high-effect one for 56. Commonest balanced recipe: adapter 1.0 + CFG 2.0 (19 attributes); commonest high-effect: adapter 1.0 + steering +0.10 at one layer (12). Median balanced cell: WER 0.086, genuineness 3.92/6, blend 4.95/10, burst realisation 0.436, |duration error| 0.07 s. Two composition rules are measurements rather than preferences: ESTH and S_RANT cancel (ranting alone +0.464, t 7.01 on 12 of 12 prompts; with aesthetics added, −0.012), and the matched random direction is null alone (−0.033) but not null at the combined operating point (+0.106, t 2.78), so every recommendation must beat a random push of the same magnitude. The shipped agent chose adapter + CFG, not steering; the reason is the listening test of Part II.3.3.

I.14.6 A defect in merging adapters, found in the demo server

The demo server at LAION-AI/Humaneness-Voice-Demo-Server — the production home of the agent of Part II — loaded adapters by parsing the tensors itself and adding the product straight into each module's stored weight (lora_bank.py:274), undone by subtracting the recomputed product at end of turn (:305). PeftModel was never constructed. In this architecture that collides with tie_weights() (modeling_moss_tts.py:114): text_lm_head.weight = transformer.embed_tokens.weight, and for every codebook head.weight = embedding.weight — the same tensor object. Every adapter targets audio_lm_heads.0…11, and those twelve tensors are audio_embeddings.0…11. Three consequences, none externally visible:

• Head and embedding table are modified together. In training the adapter sat on a frozen weight and the table never moved; at merge time it does, so the merged model is not the model that was trained or evaluated. Measured: both tensors moved by exactly 6.103515625e−05.
• It is inside the generation loop. modeling_moss_tts.py:578 reads audio_embeddingschannel to feed each generated channel back as the input for the next, twelve times per frame.
• Merge/unmerge does not restore the weights. float16 deltas added into bf16 weights (≈8 mantissa bits): w + d − d ≠ w, because the addition already discarded bits the subtraction cannot recover. Every turn drifts further, and via the tie, so does the embedding table. This is the “rounding drift” the demo server's own investigation listed as never confirmed; the mechanism is real, its magnitude over a long session unmeasured.

Fix: load through PEFT and leave unmerged, scaling by multiplying the stored scaling — identical in effect to a merge at that weight, with the base weights never written. Keeping a custom loader requires at minimum excluding audio_lm_heads.* from the merge and applying those modules as forward hooks, and restoring from a snapshot rather than subtracting. The state of the server's loader as of 9 September is recorded in Part II.2.

I.14.7 The stop-talking bug, and what it says about duration

The same server chased a different symptom: the model said its line and then kept going, a few often-invented words on every second or third reply. Their conclusion constrains the rest of this report.

• Not the stop token. Swept across its range, three settings of the stop bias produced bit-identical output through two code paths — only possible if the stop-or-continue decision is never close.
• The clock. Across every configuration the audio matched the requested duration to within 0.02 s. The model was not overrunning its budget; it was honouring it exactly and filling unneeded time with invented speech.
• Two inflators, neither sufficient alone: the render rate (4.5 audio frames per word, the rate in the format's own worked example) and a closing pause tag after the last word. At the old rate, dropping the closing pause still left invented words in 20–40 % of takes; at the tighter rate, keeping it left 10–20 %. Together: 0 % in the final measurement across three seeds.
• Removing the timing tags is the worst configuration measured — 0.9 invented words per take and nearly four times the trailing audio. The timed format is not the problem; an over-generous budget inside it is.
• Adapters contribute: base plus quality adapter gives WER 0.013 and zero invented words in ten takes; the full shipped stack gives 0.258 and roughly two per take. Turning the same adapters down recovered most of it (0.273 → 0.018) at a cost in the naturalness score, which was highest on the heavy stack.

The constraint this places on everything else: the model treats the duration budget as a contract, not a ceiling. The project's own duration errors are small because its budgets are tight, not because the model stops well — and any future work that loosens a budget should expect filler rather than silence. The server's extra_w metric, counting words after the last one that still matches the reference, is a more direct measure of over-generation than anything in the harness and should be adopted.

I.14.8 The word-error instrument

Every WER in I.8–I.13 comes from openai/whisper-large-v3-turbo. From 2026-08-29 the primary instrument is nvidia/parakeet-tdt-0.6b-v3, a transducer with no language decoder, chosen on the theory that Whisper would hallucinate fluent text on over-driven audio and therefore flatter exactly the failures these levers produce. The theory was wrong, and measuring it is the useful part. Pooled offset −0.0177 over 29,676 clips (independently −0.0237 and −0.0231 in the other two studies), and the offset does not grow with guidance: slope −0.00081 per unit g, t −0.19, n = 1,863 cells, with the flattered share flat between 3.4 % and 4.5 % at every guidance value. Whisper is the harsher instrument here, worst where the audio is worst — on the most broken clips it returns mean WER above 1.0, hallucinated insertions rather than the plausible substitutions that were feared. The Whisper thresholds of I.8–I.13 were not too permissive and those tables stand as written. Parakeet remains primary for a different reason: on identical audio it is 2.4× less noisy, a real gain in resolution for any harness running few samples per cell. It is the WER ear of the agent.

I.14.9 Two things nobody had measured on 29 August

Blending two emotions. Nothing in any of the studies tested what happens when you ask for two feelings at once — bitter but amused, frightened but trying to sound calm — or for a point somewhere between them. What was known constrains how it should be attempted, and rules out the obvious version: adding two scaled directions is not interpolation (anger at 0.10 plus genuineness at 0.10 realises 0.1604 at the layer they share, because the two directions are not parallel; the correct construction is a spherical interpolation between the two normalised vectors, then a single α); some pairs are destructive (ESTH × S_RANT: +0.464 alone, −0.012 together); the response is not uniform across emotions (Sexual_Lust +0.487 and Affection +0.476 at the free operating point, but Fatigue_Exhaustion −0.113 and Doubt −0.070); and the two tails of one attribute are orthogonal, not opposite (median cosine −0.0004). The cheapest version is not a vector at all: the corpus captions already contain blends — a real GENERAL: line reads “affect is mildly positive … reads as amusement, teasing, interest” — so blending at the prompt level is in-distribution and blending at the vector level is not. As of 9 September this is still unmeasured (IV.2).

Spectral content. Every score in this project comes from a model, and not one of them would notice a gentle low-pass. The MOSS tokenizer is a residual quantiser: codebook 0 carries the coarse structure and each of the eleven that follow carries the residual of the ones before it — sibilance, breath, the band above roughly 6 kHz. Those deep channels are the most fragile: predicted last, conditioned on every earlier channel of the same frame, weakest signal, least conspicuous errors. Every adapter targets all twelve audio heads and stacked adapters sum their deltas linearly, so the load on those channels grows with the number of adapters — the same quantity the demo server measured driving word error from 0.013 to 0.258. A band-energy ratio — 6–16 kHz against 0–6 kHz on a fixed line — would be a few lines of code and would close the hole. Still unmeasured (IV.2).

I.14.10 What the lever study does not settle

• The emotion problem. A balanced operating point for 53 of 60 attributes is a usable system, not a solution. Every lever moves delivery several times further than feeling; the next real step is data rather than inference.
• Nobody has listened. Every number here is one model's judgement of another model's output, and several of the scorers produced the training labels in the first place.
• Stacking is untested in this study — all adapter measurements are one adapter at a time. The combination study of Part II.3.5 is the first stacking measurement.
• Total injected magnitude was never held constant in the layer-count ladder, so “more layers” is confounded with “more push”. The realised per-tap magnitudes are logged, so a matched-budget re-analysis needs no new generation.

I.15 The trajectory corpus

Every measurement in Part I is of one utterance in isolation. That is the right unit for “can this model perform contempt at the 95th percentile” and the wrong unit for almost everything a director actually asks for, because direction is about change: start polite and let it curdle, come down off it slowly, angrier with every sentence. None of those is a single target percentile, none can be scored by running an instrument on one clip, and nothing in this project has ever trained on or evaluated one. A trajectory is the smallest object that makes that measurable: an ordered list of clips from one speaker in which one measured dimension moves monotonically. Because every clip in the corpus already carries the full 40-emotion and 57-VoiceNet annotation of I.6, such sequences can be mined rather than recorded. Source: [TR-0829] §17 and its companion page.

A row is a specification and contains no audio. Twenty-four columns: the ordered clip identifiers, the speaker they share, the dimension being walked, the distance covered and the per-step size, the monotonicity and consistency thresholds and the step count, quality bounds, total duration, whether the clips are adjacent in the source recording, and selection bookkeeping. Roughly 55 bytes per trajectory, which is why 7.6 million of them fit in 419 MB.

The two-sided family is the smallest and the most interesting. A single-dimension walk asks the model to do more of something; a two-sided walk asks it to trade — and every acting note about a turn, a realisation or a change of heart is a trade rather than an intensification. Trajectories are short: two to five clips, and the step histogram sums exactly to the row total in both versions, so there are none longer. The public release is a clean removal of three source datasets — podcast, snippets and evasnippets — whose transcripts are not part of the open real-speech release. A trajectory row names exactly which segments were used, which is a weaker disclosure than publishing the transcripts but is not nothing. The exclusion of snippets in particular was a cautious default rather than a settled decision.

I.16 Prehistory kept for the record: the DramaBox/LaionBox audio DiT, the search agent, and the first evaluation stack

The 16 August reference report ([REF-V1]) covered two model families under one name, and later reports dropped the second. It is kept here because the brief for this report forbids losing anything, because two of its findings (the bf16 ULP floor and the rank-buys-identity rule) are the same lessons the MOSS line learned independently, and because “DramaBox” names both the data source of voice-acting v1 (I.3.2) and this unrelated model — a name collision worth defusing once. Nothing in this section is a MOSS number, and the MOS/quality tables must not be read as such.

I.16.1 The DramaBox / LaionBox audio DiT (LTX-2.3 22B)

A different model: the LTX-2.3 22B audio-only diffusion transformer (base ltx-2.3-22b-dev-audio-only-v13-merged, 48 blocks, inner dim 2,048, 3.286 B trainable parameters) — a flow-matching speech enhancement / re-synthesis model, physically located in Voice-Acting-Pipeline/. HF lineage: laionbox-v0.1 → v0.3 → v0.5 → v0.6-wip. Source: [REF-V1] Part II, itself compiled from that project's protocol.md §7, §10.2, §11, RUN12_ADALN_STATUS.md, run9/run10_training_log.txt and eval_output_combined/eval_full_report.html.

Authoritative final evaluation — 72 samples (5 speakers × 6 prompts × 2 seeds, 3 EN + 3 DE), the same sidon_normalized references for every model. Metrics: MOS / UTMOS / NISQA / ScoReQ (quality), CE / PQ (AudioBox aesthetics), spk (WavLM-base-plus-sv cosine). Higher is better. n = 72 per row; no confidence intervals were recorded.

SIDON post-processing lifts every row by about +0.15–0.2 MOS (Run 5 → 4.747). ChatterboxVC → SIDON pushes speaker similarity highest: Run 15 r256 reaches spk 0.947 (against 0.920 for the full FT). Campaign verdict: (1) the full fine-tune (Run 5 / v0.3) is the quality champion at every stage (raw MOS 4.60 vs ≈4.48 for LoRA); (2) higher LoRA rank (r256) buys speaker similarity, not quality (raw spk 0.92 vs 0.88; VC→SIDON 0.947 vs 0.92); (3) AdaLN-Zero + 4 aux losses did not beat plain LoRA, and aux over-training degrades quality (Run 12 at step 1700: MOS 3.62); (4) SIDON is a reliable +0.15–0.2 MOS post-processor.

• Training-loss notes. Run 9 (LoRA128+AdaLN, 510 steps): best flow loss 0.4944 at step 428. Run 10 (LoRA-only, 340 steps): best flow 0.5028 at step 172.
• Run 12 AdaLN divergence: at lr_adaln 5e-4 the flow loss blew up 0.53 → 2.70 and adaln_norm 0.09 → 3.79 by step 128; at 5e-5 it is stable (flow ≈ 0.53–0.6, adaln_norm ≈ 0.18).
• The bf16 ULP floor (Run 7): for weights ≈ 0.01 the bf16 unit in the last place (≈ 8.58e-5) exceeds a typical Adam update (≈ 1.72e-5), so updates round to zero — a silent freeze past epoch 1. Fix = fp32 master weights (CPU-offloaded, ≈ 2.5 GB), used by all later LoRA runs. This is the mirror image of the MOSS line's AdamW eps = 1e-4 fix (I.3.5).
• Flow loss ≠ MOS: diverse_mix_data is bimodal (easy subset fits to ≈ 0.14, hard ≈ 0.55) so the flow mean tracks which subset is streaming, not quality — do not rank checkpoints by flow loss alone. The DiT project's version of Part III.1.
• Evaluation references must be fixed and identical across models — a mixed-reference methodology once flipped the DiT ranking.
• No per-model WER is computed in this pipeline; WER appears only as a data-generation reward.

I.16.2 The first evaluation stack, the 99-vector scorer and the search agent

[REF-V1] §16 records the evaluation stack of the pre-JUPITER era, which is the ancestor of the agent's Ears: WER from Parakeet-tdt-0.6b-v3 (transformers ParakeetForTDT, fp32) — so Parakeet was the first ASR instrument, was replaced by Whisper-large-v3-turbo for the timed-script round, and came back as primary on 29 August (I.14.8); va_rescore.NinetyNineScorer — 40 EmoNet emotion experts + 57 VoiceNet regression dimensions + genuineness (slot 97) + blend (slot 98); torchmetrics DNSMOS for quality; ECAPA-TDNN and Orange/Speaker-wavLM-tbr for speaker similarity; laion/vocalburst-locator for bursts; and a Gemini LLM judge (gemini-3.6-flash via a gateway), audio-in, 0–10 verdicts.

The reward-design default of that era, user-specified: fitness = (Σ wᵢ·norm(targetᵢ) + w_g·norm(GENU) + w_b·norm(BLEND)) × (1 − WER). The (1 − WER) factor is mandatory for speech missions (it curbs overacting); GENU + BLEND together carry about the same total weight as the targets. All fitness is measured over a cohort (n = 8 default), never a single sample. Compare the agent's R_old of Part II.6 — the same three ingredients, plus CLAP, with the same (1 − WER) multiplier.

Is the VoiceNet representation good? — the EXT-MLP benchmark. On the ≥ 2-rater VoiceNet-EXT cut (13,917 items / 395 prompts) the ordinal-regression MLP predictors top balanced accuracy per prompt, ahead of both VoiceCLAP contrastive readouts: VoiceCLAP-Small (110 M) contrastive cosine 0.6367 / per-prompt ρ 0.1051; VoiceCLAP-Large (7 B) 0.6510 / 0.1475; VoiceNet-Dim predictors (MLP) 0.6693 / 0.1317. Source: [REF-V1] §6, from the paper-rebuttal numbers in WIP docs/08.

The autonomous LoRA + prompt search agent (LAION-AI/voice-acting-search-agent) searched LoRA merges and prompts against these rewards on a single A100-80GB. Overnight programme (1–2 August 2026): 163 agent generations, ≈ 10 k scored takes. The Gemma-4-31B QAT brain won decisively (10/10/9 on single-dimension missions, only 22 tool calls; the 12B arm “wins” raw scores only by paying WER). Supervisor: Gemini was an effective active director (tracks the objective at r = 0.72, gives sonic-diagnostic feedback); a local MOSS-Audio-8B judge was only usable as a lenient score-only gate (active steering was harmful). The Part II agent's choice of a Gemma-4-12B local Mind is a cost decision made with this result in view; the 31B/12B gap was never re-measured on the new instrument (IV.2).

I.16.3 The gotchas list of 16 August, kept whole

• bf16 needs AdamW eps = 1e-4 (MOSS FT) and fp32 master weights (DiT LoRA) — both defeat silent underflow/ULP freezes. Gradient checkpointing must use the module-wrapper path, not the legacy enable().
• Rank tradeoff is consistent across both model families: higher rank → speaker identity, not quality. Full fine-tune beats LoRA on raw quality; LoRA + fp32-master is the cost-effective quality/identity compromise.
• Emotion and VoiceNet LoRAs raise the target axis but cost genuineness/blend/quality/WER — merge at ≈ 0.5 for the best reward. λ = 0.5 already delivers ≈ 36 % of a VoiceNet dimension's full shift; tempered doses (0.4–0.5) preserve speaker similarity.
• Got-Talent high-emotion synthetic data is the strongest emotion-LoRA source. Prompt-only sometimes beats any LoRA (e.g. Sadness); always A/B against prompt.
• Score cohorts (n = 8), never single samples; report mean ± sd + side-metrics. MOSS codec decode is stereo → downmix with mean(0), and run a self-repetition check before publishing. Flow/val loss is not quality — use MOS/DNSMOS + speaker-sim + WER-gated reward.
• /tmp and /run/user/1001 on the A100 box are RAM-backed tmpfs (504 GB / 101 GB) — fast but wiped on reboot; disk / was near full. Publish deliverables to HF/GitHub. (This is why the v2 training scripts no longer exist — I.3.3.)
• Background servers (http.server, cloudflared) get SIGTERM'd if launched inside another call — run each as its own task; cloudflared is at /home/deployer/cloudflared.
• speechbrain 1.x passes use_auth_token to a hf_hub API that dropped it → monkeypatch for ECAPA. Pushing to LAION-AI needs a classic repo-scope PAT with SSO (fine-grained PATs fail).
• Gemini judging via /v1/chat/completions silently drops the audio (answers as if nothing was attached; everything scores 0) — caught only because real recordings also scored 0. Use native generateContent + inline_data. gemini-3-flash needs thinkingBudget=0; gemini-3.6-flash rejects it with HTTP 400.
• MP3 bitrate_mode="CONSTANT" pins to 225 kbps (570 → 220 GB VBR); pyarrow.from_pylist infers schema from row 0 only (drops later-appearing columns); upload_large_folder deadlocks against the Hub 320-commits/hr limit; the audio tokenizer OOMs above batch ≈ 12; a 0.16 s “double-decode” if audio_codes_list already holds waveforms.
• Corpus bias: in the German-Mediathek > 2.5-emotion subset, music broadcasts are 8× over-represented and pile into dysphoric heads (Distress 51 %, Sadness 51 %, Longing 41 %, Intoxication 40 %). “Balanced” is unachievable → dominance-cap the top class.

The canonical documentation package of that era was the private LAION-AI/Voice-Acting-Pipeline-WIP (docs/01–23 + docs/LEARNINGS.md + code); the DiT project lived in Voice-Acting-Pipeline/. Both links were dead (404) by 26 August ([TR-0826] §18.2).

Part II — The actor: Mind, Instrument, Tools, Ears, Selection

II.0 The metaphor and its technical counterpart

Part II walks through the actor one part at a time. The project lead gave this undertaking a framing narrative, and it is not decorative: it is the architecture. An actor has a mind that reads a scene and decides how to play it; an instrument that responds without wanting anything of its own; tools with which the actor works on the instrument and produces several takes; and ears with which the actor hears those takes and selects one. The table below maps each part of the metaphor to its technical counterpart and to the section of this part that measures it.

II.1 Actor's Mind — the directing model

The actor's mind is given the situation and writes the prompt from it. In the running server this is a language model that delivers four things in a single constrained pass: the reply (reply), the delivery prose (delivery → GENERAL), the script with the bracketed cues (script → SCRIPT), and two real tool calls — the choice of reference voice and the choice of generation mode. This is enforced through a JSON-schema grammar, not through a second call: a separate tool turn would have put a full prefill and decode onto the time to first sound.

II.1.1 The rule that holds the whole system together

From this follows the most dangerous unwritten rule of the system: a parenthesis with a number is a sound, one without a number is an instruction. (chuckle) is a noise, (clearly amused) a stage direction — and (clearly amused, with a small chuckle) produces no chuckle at all, because the whole parenthesis is read as an instruction. The duration is mandated by the format and forbidden in the prompt. Whoever confuses the two silently turns every stage direction into a sound.

A second precondition that long went unspoken: a cue is written in English, even when the spoken line is German. The corpus is labelled in English; its German lines read literally Das zerreisst einen einfach, weisst du? (relief sigh). A German cue lies outside the distribution and drags the words down with it.

II.1.2 The chain a turn passes through

From input to sound there are eleven stages. At several points the order is itself a result and not a convention. The count of “eleven stages” comes from the German evaluation page ~/voice_acting_agent.html; in the source itself six section markers are set and the README describes four steps — so it is a reading of the code, not the code's self-description.

II.1.3 What is actually measured about the mind

And here stands the most uncomfortable single result of the whole programme. The written stage direction moves nothing. In a fully balanced grid of 1,440 clips — 6 emotions × 5 utterances × 2 languages × 2 punctuation variants × 2 direction variants × 6 voice conditions, plus a 240-clip control — the direction changes the emotion score by -0,002 [-0,023, +0,020] (paired, 720 pairs). Not the genuineness, not the blend, not the burst realisation, not the word error rate. This is the factor around which the entire prompt surface of the third round was built.

The obvious objection was checked and does not hold: had the emotion adapter already saturated the score, the direction could add nothing more. In the control with the adapter at weight 0 the effect of the direction remains non-significant (+0.022 [−0.031, +0.077]), and switching the adapter off lowers the emotion percentile only from 0.864 to 0.821. Neither the adapter nor the direction produces the high scores.

What produces them is the text: 35,1 % of the variance of the emotion score is explained by *which sentence* it is. The utterances were written to match their emotions, and the classifier hears content, not only delivery. From this follows a consequence that reaches beyond this study: as long as the evaluation uses emotionally coloured text, no training method can show an effect — not because it does not work, but because the instrument is saturated before the model speaks.

Two things the mind very much does move come from the work on the demo server and so far appear in our protocol only as a reference: an example moves a language model, a rule does not. The rule “put pauses in the middle of the sentence” took effect roughly zero times per reply; rule plus one example roughly once; rule plus one example per feeling plus an explanation of the tempo value one to three times. And 19 prompt additions from an evolutionary search over 2,949 scored clips changed nothing — the first run flattered itself; corrected, the control wins. The take-to-take spread of the same cell there is SD 1.57 out of 15, so that a best-of-3 fitness buys +2.78 points from luck alone.

II.2 Actor's Instrument — SFT3

The instrument has no intent. It is given a prompt and answers. The basis is laion/moss-tts-local-transformer-4.55b-voice-acting-v2 — a semantic transformer with around 4 billion parameters over 36 layers, a “talker” transformer with around 550 million and twelve audio heads, together 4.13 billion trainable parameters. The audio tokenizer is MOSS-Audio-Tokenizer-v2, 12 codebooks of 1024 at 12.5 frames per second: one frame is exactly 80 milliseconds.

II.2.1 The prompt surface, in three rounds

Round 1 carried the text in plain form, preceded by a delivery instruction in parentheses. Round 2 introduced the timing script — every speech duration, every pause from 0.20 s and every burst length are computed and written into the prompt — and dropped the delivery instruction without replacement in the process. That went unnoticed at the time and is the central negative result of that round: giving an instruction afterwards to a model that has never seen one destroys it. With a forced instruction the word error rate of all round-2 models lies at 0.48–0.51 against 0.11 without, including bare SFT-2 without any adapter.

Round 3 brought the instruction back, this time *inside* the timing script. That fixes the collapse: 0,447 -> 0,099 word error rate, with better timing fidelity and better burst realisation at the same time. Emotion control does not improve. That is the dividing line along which this programme has run ever since.

II.2.2 What the instrument can do

II.2.3 What it cannot do

Tempo. A tighter budget forces faster speech (slope ≈ 0.66), but there is a floor at about 18 cps: below it the model does not slow down, it fills with silence. In the arm in which the budget was held fixed and only the tempo label was varied, all four slopes are zero — telling the model 14 or 26 characters per second changes nothing, it speaks 18. That holds for the adapter that was trained on exactly this, too.

Why this could not work is, in hindsight, arithmetic: the character rate of the corpus was computed from the printed duration and the printed text. In every training line the rate agrees with the duration, and the duration is in the prompt — the model never had a reason to read the rate. The label is an invertible paraphrase, not a second control channel. 121,940 lines and 1 h 47 min of training bought no new capability. The adapter is harmless and useless; stage 2 sits safely on top of it, but it must not be described as tempo control.

Emotional intensity on command. Three objectives were set on it — GRPO with group-relative reward, DPO with contrastive pairs, DPO with symmetric, instruction-conditioned pairs — and none moved it. The only method that has moved it is the emotion LoRA on the filtered top one percent (0.3572 against 0.3494 baseline): more extreme data, not a cleverer loss function.

II.2.4 The training path, in one table

II.3 Actor's Tools — adapters, guidance, prompt forms, steering

The actor works on the instrument with craft. For us that means three kinds of tool: LoRA adapters (small additional weights), guidance (the difference between “with instruction” and “without” is amplified) and the prompt forms (how a request is phrased). A fourth, the steering vectors, was built, measured — and switched off again after a listening test; it therefore gets a section of its own (II.3.2).

II.3.1 The shipped stack

What is actually inside the model when a reply is generated is not a question of documentation but of configuration. The following table is read at render time from $SC/hv_mirror/config.py.

That is seven permanent adapters, i.e. 5.75 to 7.25 of merged weight, before the first burst adapter is added. This number is not a side issue; it is the actual experimental setup, and II.7 shows why.

II.3.2 Guidance (CFG) — the tool that needs no training

For every frame the model is run twice: once on the full prompt, once on the same prompt with the condition removed. The audio logits are extrapolated away from the unconditional branch: logits = logits_uncond + g · (logits_cond − logits_uncond). At g = 1 the unconditional term cancels exactly — that is a genuine control which runs through the same code path, including both KV caches.

II.3.2.a On the emotion condition

The first lever in the whole programme that moved emotional intensity at all. 96 clips, 24 utterances × 4 strengths, on the three emotions furthest from saturation: emotion 0.809 (control g = 1.0) → 0.887 (g = 1.5) → 0.917 (g = 2.0) → 0.912 (g = 3.0). Paired against the same prompt-and-seed combination: +0.078 · +0.108 · +0.103, with 95 % intervals that all contain zero.

The large follow-up study over 7,419 cells / 29,676 clips puts the magnitude firmly into perspective and confirms the direction: for emotion the best value under the word-error bound is g = 3 with +0.0254 (t +2.18, 63 of 120 improved), for the VoiceNet axes g = 2.5 with +0.0569 (t +3.12, 57 of 87). Below g = 1 guidance actively hurts: −0.0370 at g = 0.5 (t −2.56). These two measurements of the same object differ by an order of magnitude because they were measured from different baselines (0.809 versus 0.263) and one takes three hand-picked emotions, the other all forty. To this day they have not been merged into a single number, and that stands in the paper outline as an open item.

II.3.2.b The price, measured five times independently

The expectation was well below double, since only the semantic transformer runs twice. The expectation was wrong. The local “talker” transformer runs twice as well — twelve times per frame per branch, because the unconditional branch needs its own local state for the next channel's logits — and so do the twelve audio head projections. Only the codec decoding is shared, and that accounts for 1.6 % of the work. Sharing it saves nothing.

The price is the second branch, not the strength. So everything that argues for guidance at all argues for running it at the best point of the band rather than the smallest. In the server it therefore does not run streaming: the realtime budget there is 1.0 and the merged baseline 0.764; 1.93× lands at ~1.47 and the player starves. A guided take is rendered in full and then played. The automatic lever selection therefore never emits guidance.

II.3.2.c On the burst — the clearest guidance finding of the programme

Here it is not the emotion condition that is neutralised but two axes at once: the unconditional branch receives a neutral prompt and has the burst adapter switched off. The amplified difference vector thus carries the prompt cue and the adapter delta together. For this to be implementable at all, the adapter state had to be switchable inside the twelve-channel loop — the twelve audio heads are LoRA targets themselves. The index is therefore split: 252 modules in the semantic transformer (once per branch and frame) and 16 in the channel loop, so that 16 rather than 268 numbers are rewritten there. Cost of the switch: below the measurement noise.

Paired against C10: D4 +0.0972 strict (t +2.72) and +0.2083 family-relaxed (t +4.41) — but at +0.2235 word error rate (t +5.63). E2 +0.0625 / +0.1319 (t +3.25) at ΔWER −0.0141 (t −0.77) — i.e. for free. And the decisive decomposition: D beats E on the strict metric at no g (t 0.72 / 0.94 / 1.41) and pays +0.09 to +0.20 word error rate for it at every g. The prompt carries the larger part, not the adapter delta.

The one exception is exactly where the adapter can do something: on scream / inline — 496 verified training lines — D4 − E4 is +0.333 strict (t +3.74), and the class goes 0,208 -> 0,583. That is the largest movement any lever of this programme has produced on a burst class. shriek (48 lines) and frustrated_groan do not exceed 0.083 strict under any arm. Guidance amplifies a direction that exists; it does not create one.

The wiring control is not optional and was measured rather than assumed: at g = 1 the guided path must reproduce the single-branch path exactly. Largest deviation |guided − conditional| over every channel of every frame: max 9.5·10⁻⁷, mean 1.1·10⁻⁷; the results are cell-identical, D1 − C10 = E1 − C10 = 0.0000 on every metric. Everything above g = 1 is therefore guidance and not wiring.

II.3.2.d The preference pairs from guidance contrasts

The third application of CFG is a training corpus: contrast pairs that differ only in how strongly one dimension is expressed, with the spoken text removed so that the preference cannot go through the words. Every pair is emitted twice with swapped roles: the instruction “high” makes the intense clip the chosen one, the instruction “low” the mild one. Both clips appear as chosen and as rejected — so the model cannot win through a global preference for intense audio, but must condition on the instruction. The two halves are balanced to 0.1 percentage points; in the table above they are exactly the same size.

And exactly this symmetry is the reason it did not work. The trained model became the best general model of the programme — on reward, word error rate, quality and burst realisation — and did not move emotional intensity. The gradient towards intensity under “high” is mirrored by an equally large gradient away from it under “low”; the net movement of the unconditional distribution is zero by construction. Correct for a conditioning signal, useless for raising a ceiling. That the construction took hold is attested independently: preference accuracy on the CFG families rose from 0.562 (chance) over the first 40 logged batches to 0.975 over the last 40.

II.3.3 Steering vectors — the lever that won in the measurement and failed in the ear

A steering vector adds a direction directly into the hidden state: h ← h + α · unit(v) · ‖h‖. The direction is a difference of means — the mean of the activations over the clips that show an attribute strongly, minus a shared reference set. Where these vectors come from and in which layer they sit is described in II.4.

II.3.3.a What was measured

Three rules follow from this, and the server implements exactly them. Emotion: adapter and steering are cleanly additive — the interaction is indistinguishable from zero — and steering there is roughly five times the adapter. Delivery axes: significantly sub-additive — the two levers do the same thing, so exactly one runs. Quality: steering moves nothing, the adapter is the whole lever.

The ladder itself is the most interesting part, because it was once read completely wrong. At α = 0.10 on layer h20 the emotion percentile moves from 0.4354 to 0.5840, paired +0,1486 (t 3,17) over 24 prompts, while a matched random direction of the same length moves only +0,0135 (t 0,48). The word error rate falls in the process from 0.166 to 0.115, and the share of clips under WER 0.2 rises from 0.84 to 0.91. The price is paid monotonically in genuineness: 3.77 → 3.64 → 3.15 → 2.92 → 2.41 → 1.90 up the α ladder, while the random direction holds 3.64–4.18.

II.3.3.b And then a human listened

This finding is documented in five places ([P§40]; ~/reports/ensemble_protokoll.md; ~/reports/crossfade_v2.md; ~/wikiskills/REVISION_2026-08-31.md; ~/wikiskills/REVISION_2026-09-02.md). It is probably the single most important lesson about the distance between the agent's “ears” and real ears. And it carries not a single number.

Two qualifications that sharpen the finding rather than explain it away. First: the server ran with α = 0.10 under a ceiling of 0.15 — i.e. inside the range the measurements call good. The report is therefore not an artefact of the far tails and does not disappear when α is lowered. Second: the listener's session concerned the emotion axes, and there the measured effect is +0.10 percentile, i.e. almost nothing — almost everything that is audible there is therefore artefact. The open question is narrow: does steering on the delivery axes sound right at α ≈ 0.1? Nobody has listened. Until someone does, GEN_MODE = adapter stays.

The consequence reached far beyond the default. When the study on emotional transitions was repeated, five of its nine arms were dropped without replacement — they *were* the steering schedule. What remained in the second version crossfaded adapter merge weights instead of steering strengths, and the result was unambiguous: not a single crossfade beats a random reordering of its own weights. The best value of a crossfade against its own shuffle control is +0.044 with 14 of 28 texts improved, p = 1.000 — as exact a null as the design can produce. The two-part cut that is already shipped, by contrast, beats them at all four guidance levels (p = 0.013–0.036).

II.3.3.c Two hard refusals that hold regardless of the setting

• No steering on the quality axes. The measurement is a null (see table), and it agrees with two independent observations.
• No steering onto the low ends of the delivery axes. The vector table contains only the difference *high minus low*. The two ends of an attribute are orthogonal, not opposite — median cosine −0.0004. “Halfway between anger and calm” is not “anger at half strength”, and targeting a _low end via a sign flip hits an unrelated direction.

A third, smaller result of the same family is in the configuration: two adapters can cancel each other out. esthetics_high and S_RANT_high work individually (+0.20 to +0.32 and +0.464, t 7.01, 12 of 12) and together not at all (−0.012). The aesthetics adapter is therefore set to 0 server-side when ranting and halved on the neighbouring axis.

II.3.4 Prompt forms — the tool that costs nothing

The third kind of tool is the wording. It is cheap, it needs no training and no second forward pass — and across the whole burst programme it is the most reliable lever we have found. Fifteen classes, in the server's stack at its own doses, 5,360 cells / 32,160 clips over two jobs, paired over ten prompts per class; the pooled rows run over n = 15 classes.

The first two forms add up cleanly: +0.026 and +0.022 give +0.044; +0.002 and +0.023 give +0.030. Together they are the only prompt result of this study that is significant on the strict metric. The costs add up with them.

Two forms hurt, and both in an instructive way. The action form ((sie stöhnt frustriert, 0.5 Sekunden) instead of (frustrated groan, 0.5 seconds)) drives the failure rate up by +0.120 (t +3.4, 11 of 15): the model degrades to silence, not to a wrong burst. The model was trained on labels from the detector's vocabulary; an actor's direction in that field is not a label it knows. And at the same time blend (+0.183) and DNSMOS (+0.032) rise while the burst stops happening — the clearest case in the programme of a metric moving in the right direction for the wrong reason.

Mid-sentence placement drives the failure rate up by +0.306 to +0.366 (t +8.4 to +10.4), and all fifteen classes get worse. Roughly a third of all mid-sentence bursts simply do not happen. Read the other way round, that is a *positive* result about the shipped prompt: the sentence boundary is worth about 0.3 in failure rate. One class reverses it — clears_throat goes from 0.250 to 0.467 mid-sentence, its best cell of all. Placement is a decision per class.

The longer-duration effect also has an uncomfortable origin, which was cleanly separated. Over 15,409 cues where anything at all was detected, the probability of a family hit rises with the detected duration — 0.282 (under 0.3 s) → 0.388 → 0.476 → 0.510 (0.8–1.2 s), then it saturates. Against an announcement 1.0 s longer, however, the delivered burst grows by only +0.054 s (t +2.3). The detector's slope times 0.054 s predicts +0.020; observed is +0.012. No residual remains that one could attribute to better realisation — the gain belongs to the measuring instrument, not to the model.

In total 31 of 117 burst pages carry a recipe; 19 classes realise not a single hit at any weight and under any prompt form, and a further 17 stay below the threshold. The pattern is families, not classes: every mouth class and every whistle class of the bank is on one of the two lists. They are not weak adapters that would need a larger weight — the sound is missing from what the model can produce, and the way out is data, not a dial.

II.3.5 What an adapter actually moves — the combination study

The message of this table is the refusal: where a column says “none”, the measurement says that there is no usable setting. A consumer can no longer tell an interpolation from a measurement once it has been written down — so it is not written down.

The large dose–response study over 79 adapters × 7 merge weights (5,740 cells, 17,220 clips) found 33 of 79 adapters with a usable weight. Its thresholds are derived, not chosen: first the noise floor was measured (every w = 0 cell holds several independent clips per prompt without adapter, and any difference within it is pure sampling noise), two independent estimates agreed to within 4 %, and the threshold is then max(3σ, a substantive lower bound). The price of this is stated openly: a threshold below the noise lets honest adapters fail by chance — at 2σ that would be about 40 false failures over ~1,800 tests, at 3σ about two. Conversely, a real degradation *below* 3σ goes undetected, which is why the table says “not detectable at this resolution” and never “zero”.

Two readings sit in it. The emotion adapters are cheap, weak and stop acting at w = 1.0 — 1.5 minus 1.0 is −0.0026 (t −0.40, with only 20 of 40 adapters higher), while 1.0 minus 0.5 brings +0.0206 (t 3.65). The delivery axes are what works, monotone up to 1.5, 17 of 17 adapters improve from w = 0.5 on. And the only reliably monotone cost quantity is not naturalness but speaker similarity: 0.548 → 0.484 (emotion, t −8.3, 36 of 40 worse) and 0.559 → 0.475 (VoiceNet, t −8.5, 17 of 17), while genuineness actually *rises*. Turning an adapter up does not make the voice less genuine, it makes it less like the speaker that was asked for.

II.4 Layer forensics — where in the network each attribute lives

The steering vectors of the previous section do not come from nowhere. They are the by-product of a piece of work that answers a question of its own and stands on its own: at which point in the network is which property of a recording readable at all?

II.4.1 How it was measured

The regime is the decision everything hangs on: teacher forcing, no free generation. One forward pass per corpus line, with that line's own MOSS codes on the assistant side, packed byte for byte as in training; the hidden states are captured by hook. Activation and label therefore describe the same waveform, and every target — the 40 emotion heads, the 57 VoiceNet axes, genuineness, blend — has already been measured on exactly this waveform.

Capture happens at 38 taps: h00 is the sum of the embeddings before layer 0, h01 to h35 are the outputs of decoder layers 0 to 34, h36 is the normalised output of the last layer — that is the 80 ms vector that goes into the acoustic decoder, because the global → local mapping in this checkpoint is the identity (checked at runtime) — and loc is the state of the single-layer talker transformer at the frame slot, i.e. the vector the twelve audio heads actually read. Without loc, h36 would have been a duplicate.

II.4.2 What is readable where

All 99 attributes are readable to some degree — the weakest, Jealousy_and_Envy, still reaches 0.209, against a metadata baseline of −0.003. Best readable are GEND 0.939, REGS 0.924, S_MONO 0.920; worst Jealousy_and_Envy 0.209, Awe 0.239, Thankfulness_Gratitude 0.281.

Two findings alongside, both of which count for the outlook in II.4.6. Speaker memorisation is practically absent: the difference between the random and the speaker-disjoint split is −0.0035 (emotion) and −0.0026 (VoiceNet), i.e. within the seed noise and slightly negative. The exception is genuineness at +0.042 — genuineness is partly a speaker property, and a genuineness predictor on activations is in part a speaker lookup table. And the MLP beats the ridge on all 99 attributes, median +0.043, the largest gap of all at genuineness with +0.180. That is the measurement which pulls the ground from under the premise of a *linear* steering direction: where a probe finds 0.18 R² beyond a linear map, a single direction from that layer will fall short of the probe.

II.4.3 Reading several layers, and the controls for it

Over all 38 taps, ensembles of size 2 to 6 were searched — greedy, beam and random, six attribute groups, two extractions: 4,888 fitted combinations in 13.1 GPU hours. The hidden layer stays fixed at 512, so that only the input projection grows with k.

The seed bands are the real statement here. Blend and genuineness have the largest raw gains and the largest error bars, and their gains lie at or inside those bands; emotion and VoiceNet have small gains far outside theirs. A table cannot be read at two confidence levels without saying so. Both capacity controls — the best tap duplicated k times, and the best tap plus k−1 blocks of N(0,1) noise, each at identical input width — are beaten by all 60 winners.

Two side findings worth having. Redundancy does not predict complementarity: the best emotion pair h18 + h21 has a representational similarity of 0.979 and still gains +0.020. And the value of the greedy search lies in reaching across all 38 taps: h06 alone ranks 13th of 38 and would not have made any pre-selection — it enters at k = 4 and stays through k = 6.

II.4.4 What follows for vectors — and the standing puzzle

From the same activations the vector library was built: archetypes.npy, 156 × 38 × 2560 float32, from 309,631 clip means. Deliberately float32, because a steering direction is a difference of means — one to two orders of magnitude below the means themselves. The verification is strict: all 97 pre-existing dimensions × 38 taps × 3 buckets were recomputed from the raw data and agree bit-exactly (max |d| = 0).

Three measurements on this library decide how it may be used. First: the two ends of an attribute are orthogonal, not opposite — cos(high − mid, −(low − mid)) has a median of −0.0004 over 2,166 pairs. “Halfway between anger and calm” is not “anger at half strength”. Every attribute steerable in both directions needs two stored vectors, and pushing down must select the low row rather than negate the high row.

Emotion and perceived quality are nearly antipodal directions in the model, and they sharpen towards the output. Pushing an emotion up pushes perceived quality down in the representation itself — and the α ladder of the previous section turns that into a measured causal price: genuineness 3.77 → 1.90 up the ladder. The rightmost column is the control that makes this readable at all: it shows that these are 97 different directions and not one direction with 97 labels.

Third, and it limits the library: 34 of 156 rows carry a warning flag. 32 of the 40 emotion heads have a low end that is a tie at zero — more than one percent of the corpus sits at raw value exactly 0, 32 % in the median, 82.2 % for Awe — and for 25 of them the low end and the middle select the same clips. “Negative anger” is not defined by this library.

II.4.4.a The puzzle that stays open

Two readings fit equally well, and telling them apart is the most valuable open experiment of the programme. Either the model computes the emotion in the middle and then discards most of it — that would be a defect one could fix. Or the late layers compress a rich description down to the few dimensions the codec needs — that would be normal and would mean: steering must happen early or not at all.

II.4.5 Vocal bursts by layer — and the limit that determines everything

In p3, 79,402 of 217,583 clips (36.5 %) carry at least one burst; 118,660 bursts over 33 labels, median 0.28 s. The presence of a burst is best readable at h19 — ridge AUC 0.9777, balanced accuracy 0.9232 against a chance rate of 0.500 and a metadata baseline of 0.587; the identity at h18, mean AUC 0.871 over the eight classes with enough material. Both plateaus lie in taps 13 to 23 — in the emotion band, not in the VoiceNet band, while the blend scalar sits six taps deeper at h25.

The controls are unusually complete and deserve mention, because they show how to rule out a clip-level confound. The falsification — clips whose bursts all lie *outside* the window — reaches 0.778, i.e. 0.200 below the headline: that is the floor produced by the property “this clip contains a burst somewhere” alone, and exactly why the window restriction is not optional. The sharpest test — burst-carrying clips on both sides, matched on burst count — still separates at 0.968. And strictly same-speaker negatives give 0.969 against 0.978: speaker identity is not the signal.

II.4.6 Outlook (PROPOSAL): a perception module inside the instrument instead of beside it

The observation that carries the proposal: the probes of II.4.2 are already exactly the object in question. They are predictors on the mean of a hidden layer — not on audio — and their target is the output of the audio-side scorers themselves. The measured R² is therefore directly the share of an audio scorer's judgement that can be reproduced from activations: speaking styles 0.859, VoiceNet overall 0.846, genuineness 0.719, emotions 0.473, burst blend 0.435, and burst presence as a classifier AUC 0.978. With ensembles over several layers these values rise to 0.870 / 0.739 / 0.507 / 0.455.

A perception module that sits inside the instrument instead of beside it would build on exactly that: predictors for genuineness, vocal-burst blending, the 40 emotions, the speaking styles and speaker embeddings, read from the hidden layers while generation is running. The appeal is not the accuracy — the audio-side scorers remain the reference — but that such a module need not decode any audio: it could judge within generation rather than after it, and thereby make abort, resampling or a selection among candidates cheaper than the current chain allows.

What is already there, and could be cited in a proposal as preliminary work: the speaker-disjoint split is verified (0 of 13,757 test voices leak); the caption confound is ruled out (≤ 0.015 R²); the resolution limit is known (seed noise p95 0.011); the non-linearity is quantified (MLP over ridge on all 99, median +0.043); the capacity controls for multi-layer inputs are in place; and the costs are measured — the four extractions together needed under one GPU node hour for 388,000 lines at 75.5 GB of output, the probing 26 and 13.1 GPU hours. A single 99-output probe is thus in the range of minutes.

A ninth point belongs here, because otherwise it makes the idea look too good: such a predictor would be a third model judge in a chain that already consists of models scoring one another. It also inherits both things measured in II.4.4 — the coupling to the quality axis and the finding that a random direction of equal length raises the emotion score more than the real vector does, once one turns the dial too far.

II.5 Actor's Ears — the perception module

What does the actor listen to his own takes with? Not with one model but with close to a dozen, and they have different jobs. The project leader explicitly asked for this module to be described: which models it consists of, what each of them measures, how they are combined — and what the selection actually achieves with them. The last point gets a section of its own (II.6).

II.5.1 The instruments, one by one

II.5.2 The burst detector and its recall floor

Two things follow, and the second is the more useful one. The 7-billion encoder buys about 0.07 exact accuracy for roughly 64 times the parameters — worth it for a scorer that runs offline, probably not for something that runs per request. And the non-commercially licensed small encoder buys nothing over the commercial one — the differences lie within the noise of a 425-sample cell. That is a licensing result, not an accuracy result: there was no accuracy price for the commercially usable choice, so the licence question never had to become a trade-off.

The head that counts today is a 17-way classification — 16 burst classes that occur at least a hundred times in both halves of the corpus, plus a no_burst rejection class. Chance is 0.0588. What ships is an ensemble of five initialisations within *one* grouped split — never an ensemble across split seeds, because its members would have trained on each other's test data, and the reported held-out numbers would be partly memorisation.

5 classes lie below 25 % recall on real speech — and what they have in common is the sigh-and-breath family, i.e. exactly the classes whose acoustic difference is duration and effort rather than timbre. For them a measured hit rate below about 20 % carries no information about the generator. Two of them arithmetically cannot break a threshold of 0.15.

That is why the family-relaxed column always stands beside it: the failure is granularity, not deafness. Deep Breath goes from 0.186 strict to 0.712 at family level, Breathy Giggle from 0.469 to 0.969; the confusions lie almost entirely within the family — Breathy Giggle → Chuckle, Exhausted Groan → Frustrated Groan, Deep Breath → Sharp Inhale, Humming → Soft Hum. A recipe that reliably produces *a* groan and gets the wrong name is a working recipe.

II.5.3 A vocabulary problem that is brand new

The new head cannot simply re-score the old inventory. Measured on a sample of 61,579 lines with 12,894 burst events from three parquet shards: the old burst columns carry 21 distinct labels, of which only 10 lie in the new 16-label vocabulary, and 83.8 % of the existing events (10,811 of 12,894) carry a label that the new head cannot emit at all. The four most frequent old labels — Low Mumble 26.3 %, Ahem 26.2 %, Contented Sigh 20.5 %, Surprised Gasp 7.2 %, together 80.2 % — do not exist in the new vocabulary. In the other direction, six new classes occur not once in the old inventory of the sample.

II.5.4 The change of instrument, and what it really bought

Since 29 August the word error rate has been measured with nvidia/parakeet-tdt-0.6b-v3, no longer with Whisper large-v3-turbo. The stated reason was hallucination: Whisper has a language decoder and writes fluent text over audio that has stopped being speech — exactly what an over-driven adapter produces.

But the strongest argument was not hallucination, it was variance. On identical audio Whisper's per-clip word error rate has a standard deviation of 0.3115 against Parakeet's 0.1340 — 2.33x as noisy — and the paired noise floor is 0.0833 against 0.0346, i.e. 2.41 times. Measuring with Whisper costs about two and a half times the error bar for the same number of takes.

II.5.5 What the ears do not hear

Two measurements determine how far this module can be trusted, and both are uncomfortable.

• The quality scorers do not see a catastrophic failure. When seven adapters are stacked, the high-frequency energy rises (0.0036 → 0.0131, paired +0.0095, t +2.2) — the assumption that a stack would make the sound dull is wrong. What collapses is intelligibility: 0.063 -> 0.554. Meanwhile Genuineness barely moves (4.16 → 4.04) and Blend rises (6.59 → 7.17). The stack makes the model say the wrong words, and to the quality scorers it keeps sounding good while doing so.
• The emotion instrument is saturated before the model speaks. 35.1 % of the variance of the emotion score is explained by the choice of text, the median is 0.972, and 58.7 % of the clips lie at ≥ 0.95. An evaluation with emotionally coloured text can show no effect — not because there is none.

A third case belongs here because it shows the same mechanism from the other side: DNSMOS must not decide between burst recipes. Its overall score moves by less than 0.04 for every lever, but it moves by +0.38 (t +16.7, 9 of 9) between two script kinds that differ only in how much speech the clip contains. On this material it measures the composition of the clip.

And the outlook from II.4.6 belongs here too: a perception module that sits inside the instrument instead of beside it — predictors for Genuineness, Blend, emotions, speaking styles and speaker embeddings, read from the hidden layers during generation. The groundwork exists (0.846 for VoiceNet, 0.719 for Genuineness, 0.473 for emotion, AUC 0.978 for burst presence, all speaker-disjoint and with controls); none of it is measured in the generation path, and the proposal is to be read as a proposal.

II.6 Selection — Best-of-N

From the ears a single decision emerges: the server generates a reply 8 times in one batched pass and keeps one of them. The selection is made by:

The normalisation is a min-max within the candidate set, not against an absolute scale — the three scorers have incomparable ranges (0–6, 0–10, a cosine) and are not calibrated against human judgement. If all are equal, the term collapses to a constant 0.5. The CLAP term counts double, because it alone measures *whether this is the requested performance*; the other two measure whether the take is good at all. The intelligibility factor is the raw inverse word error: the earlier flattening above 0.85 covered six of eight candidates of one run, so the factor stopped separating exactly where the candidates lay closest together.

II.6.1 No term asks for the burst — and two pull against it

None of the four terms asks whether the requested vocal burst occurs at all. Two even pull in the opposite direction: Genuineness drops when a burst succeeds, and Blend measures how well *a* non-speech sound fits, not whether it is the right one. The final report on the server put the loss at up to +0.74 delivered hit rate on scream — and said explicitly that this number is computed, not measured. This section measures it.

So positive, not negative and not zero: best-of-8 under the shipped reward buys about 29 % more delivered sounds than a random pick. The premise holds — it is just smaller than assumed. Caution with the AUC: its n is 725, not 1,940, because it is only defined on the sets that contain at least one hit and one miss.

Two corrections to our own assumption. First: Genuineness is not the culprit — its interval encloses zero. The measurement that had raised the suspicion (Genuineness 3.10 solo at w = 1 against 3.23 without adapter) is an effect between conditions, and Best-of-N never chooses between conditions — it chooses within a set in which all candidates share the same configuration. Second: it is Blend — of all things the term whose name promises the blending-in of vocal bursts actively works against it.

II.6.2 The cross-validation that closes the circularity

A detector that makes the selection, and the same detector that measures the result — that would be circular reasoning. Both encoders had already scored all 15,517 candidates anyway, so the counter-check costs zero GPU minutes: rank with one, measure with the other.

The gain survives the crossing. Ranked with large-v2 there is practically no shrinkage; ranked with commercial about half of the apparent advantage is device-specific. Both crossings stay significant, so the effect is about the burst and not about the detector — but the diagonal number of the commercial head must not be cited, only its crossed one.

The oracle is the measured version of the wiki column *P(≥ 1 hit in N)*. The entire headroom is 0.2606 — and so the gap of +0.74 on scream estimated in the final report is reachable at no λ by re-ranking. That is the value of this measurement: it replaces a calculation with an upper bound.

II.6.3 The adapter arms, and a sign that flips

The ceiling of 1.25 is roughly right. Lowering to 1.0 costs in the set mean; raising to 1.5 gains in the set mean and is practically zero on the delivered recording — the extra burst is generated and then not selected, so it reaches no listener. That is the thesis of this section in one example.

A delivery axis at 1.5 costs, paired over 160 sets, +0.0546 (t 6.91) Parakeet WER and brings +0.121 Genuineness (t 4.61) — more word errors than the burst adapter itself, and more than half of the headroom the +0.104 gate leaves.

Intelligibility is nowhere the problem. Every arm passes both bounds, and every one has a *negative* paired word error against its own w = 0 cell; absolute on inline 0.063 to 0.087 against a bound of 0.25.

II.6.4 The largest open risk: guidance

Guidance does not produce more bursts — it makes the shipped reward better at finding them. The set mean barely moves, the delivered rate clearly does. A selection effect, not a generation effect, and it fits the term table: guidance amplifies the delivery condition, the burst bracket is part of it, so CLAP separates the candidates better.

II.6.5 The recommendation

A fifth term in the reward. The gain is measured crosswise, so the circularity is closed; the relative number refers to the rate that the same independent head measures for today's reward.

This is a trade, not a free gain, and the traded quantity is CLAP: −0.0055 is 3.6 % of the within-set span of CLAP (0.1507), i.e. of the share that a re-ranking can move at all. λ = 2 rather than λ = 3, because the marginal rate breaks there: λ1→λ2 buys 3.82 hits per CLAP unit, λ2→λ3 only 2.09 and λ3→λ5 only 1.16.

Not recommended: the stepped variant (costs ΔWER +0.046, t 3.55, against +0.006 for the soft one, and brings 0.002 more hits), the agreement add-on term (moves 0.002 over the whole sweep — a clean negative) and bestmem. No recipe changes: the change concerns the selection, not the weights.

II.7 The babble bug — the fault that named the shipped configuration

A benchmark sentence — a parent shouting a child away from the window — came back as babble, word error rate 1.82. The obvious suspects were all wrong: not the brackets in the script, not the duplicated scream instruction, not the wrongly chosen emotion adapter. Four lines settle it:

One adapter at 1.5 destroys nothing, two at 1.5 each destroy the line — regardless of whether a burst was requested at all. Replicated over five seeds: 5 von 5 Seeds, Median-WER 0,82. The dose table over six seeds per cell, in the full shipped stack, shows the cliff between 2.5 and 3.0: sum 1.0 → median word error 0.000 · 1.5 → 0.000 · 2.0 → 0.010 · 2.5 → 0.110 · 3.0 → 0.820.

It was the sum of the adapters, not the bracket tags. And the reason the previous justification in the code claimed the opposite and cited a real measurement for it is the stack: that measurement had measured burst adapters on a bare model. Here 5.75 to 7.25 of weight already sit in front of them.

Three further faults were found in the same pass, and the second is the most instructive. The double scream: the guard against duplicate insertion recognised only brackets *without* a comma and *without* a digit. (scream) matched it, (scream, 0.7 seconds) did not — and it is exactly this second form that the prompt has demanded since 5 September. From that day on, every director who wrote the burst correctly had a second bare (scream) inserted. End trimming showed the wrong copy: the candidates were encoded before trimming, and only the streamed take was trimmed — one reply showed a player at 18.4 s next to a stream of 16.5 s; since then every candidate is trimmed before judging, so that the reward scores what is heard. And the budget scaling was undone again right afterwards, because the code looked the weight up a second time in the recipe after scaling: sum 2.42 against a budget of 2.0. Fixed by a final overall clamping step.

The second commit deals with the same sentence in German, which kept babbling for a different reason. Only the brackets translated, every German word of the line untouched: German cues median word error 0.267 (2 of 4 takes unusable), English cues on the same German words 0.000 (0 of 4). The corpus is labelled in English. Since then the server rewrites the cues server-side — before the retrieval stage, because with German cues the retriever falls back to the emotion named by the directing model, and that is exactly how a horror scene came to be conditioned on Jealousy_and_Envy.

II.8 The tools programme after 29 August, part 1 — conditioning datasets, transitions, tempo, crossfade, the quality adapters (protocol §32–§42)

This chapter transcribes eleven protocol sections written between 29 and 31 August 2026. They fall into four strands: three conditioning datasets built from the voice-profile corpus (§32) and the rate adapter that was trained on one of them and then measured to do nothing (§37–§39); a spectral check of adapter stacking (§33) and the protocol audit that followed (§34); the shipping of steering and guidance into the demo server with its WikiSkill layer (§35); and two rounds of the emotional-transition study (§36, §40) with the quality adapters' evaluation (§41) and a baseline error in the DPO trainer (§42). The sections are presented in numerical order, which is not the order they occupy in the file. [P§32–§42]

II.8.1 2026-08-29 — Three conditioning datasets: quality, speaking rate and speed — [P§32]

Agent A4, 2026-08-29. Code ~/quality_speed_datasets/code/, report ~/quality_speed_datasets.html, mirror research-log-2026-08/quality-speed-datasets/. The brief asked for three training sets built from the voice-profile corpus in the project's own prompt format (prompt_lib2, PROMPT_FORMAT_HASH 073aeb09dc923376): a quality preference set of about 50,000 pairs at about 100 per voice profile across all 500, an SFT set conditioned on speaking rate, and a speed preference set. It described the ~500 profiles as having been repaired by Chatterbox voice conversion and by Sidon restoration, each carrying DNSMOS before and after. [P§32.1]

The corpus is real and is exactly as described in size: $NB/train2/corpus/shard-*.parquet filtered to src == 0 is 1,200,530 takes over exactly 500 voice profiles, a mean of 2,401 takes per profile, 607,399 German and 593,131 English, mean duration 10.52 s. The repair, however, is not there at that scale, and this governs everything that follows. Three repair artefacts exist and none of them is "the 500 profiles' generated takes, scored before and after". [P§32.1]

pvcodes.py states the reason in its own docstring: the pilot and 490-voice generation runs have "exactly one rendering: the take as generated". vprof/vc500/ does cover 500 voices at 15 TB, but it is cross-voice attribute transfer — it carries target_voice and target_kind — not repair: its quality falls by 0.282 on average and rises on 1.0 % of rows. [P§32.1]

The repair mostly makes things worse, and the confound is fatal. Measured, not assumed: on arm A (the 500 selected profiles) Sidon changes DNSMOS by −0.011 on average and improves 48.4 % of clips; Chatterbox by +0.004, improving 51.4 %. The best of the three options is "do not repair" for 173 of the 500 voices (Chatterbox 171, Sidon 156). On arm C: DNSMOS −0.038 mean, 28.0 % improve, and 51 of 832 takes clear +0.10. On arm B, best of the three repairs against the raw take: +0.0014 mean, and only 3.67 % of 26,594 candidates gain more than 0.10 (no uncertainty recorded for any of these means). [P§32.2]

The confound the brief warned about is not marginal but disqualifying for voice conversion. Against the raw take, VC moves ECAPA speaker similarity by +0.259 on average and stays inside ±0.02 on 0.08 % of candidates; it is built to move the speaker toward a reference. Sidon moves it by +0.003 and stays inside ±0.02 on 36.5 %, ±0.05 on 75.8 %. On the 40-dim emotion vector, cosine to the original is 0.924 for VC (51.3 % at or above 0.95) against 0.965 for Sidon (79.6 %). "So the guards do not trim a voice-conversion dataset, they delete it, and what survives is the one repair that preserves identity and, for that same reason, barely improves quality." An independent confirmation already sat in the tree: vcbon/code/bench_pipeline.py records "SIDON on converted output measured NEGATIVE on quality" and ships sidon_out=False. [P§32.2]

What was done about it. Two things rather than reporting a small dataset and stopping. First, DNSMOS was measured where it was missing: dnsmos_pass.py scored all 106,424 arm-B clips with speechmos on the 16 kHz decode, the same scorer vpvc.py used. An operational note is recorded for whoever runs it next: speechmos builds its onnxruntime sessions with default options, and onnxruntime's default intra-op thread count is the core count, ignoring OMP_NUM_THREADS; sixteen workers asked for about 768 threads on 48 cores, load average hit 224 and nothing finished. Pinning intra_op_num_threads = 1 and chunking each tar four ways is what made it run. Second, the repair was run rather than inherited: Sidon costs about 40 ms per clip on a GPU, so mksidon.py runs it over a condition-spread sample of every one of the 500 profiles and measures DNSMOS on both sides plus an ECAPA speaker cosine per pair. Only Sidon is run, for the reason above. [P§32.3]

Dataset 2 — SFT conditioned on speaking rate. This is the one dataset the corpus supports at full scale unchanged, because it needs no repair: every take already has a duration, a transcript and word-level timings. $SC/a4_datasets/ds2_rate_sft/, 128 shards, 1,200,530 rows, all 500 voices, the corpus schema plus seven descriptive columns (rate_mode, n_rate_seg, n_rate_fallback, cps_median, cps_min, cps_max, rate_hash). No existing column is altered or dropped. The rate cannot ride in the data: both trainers render the prompt at training time (sft_train2.py:153 and dpo_train2.py:211 both call prompt_lib.user_message(row, rng)) and prompt_lib2.render_prompt builds the Text: field from text, words_json, the burst spans and dur_s. So the deliverable is the data plus prompt_lib2_rate.py, a drop-in module exporting user_message, make_rng and PROMPT_FORMAT_HASH; the change at the call site is one line. This is a new prompt format: PROMPT_FORMAT_HASH = 090c4ca315519a57, against the project's 073aeb09dc923376. No existing adapter has seen a characters-per-second tag. [P§32.4]

Three surfaces are mixed in equal thirds per row: dur — [4.7 seconds duration], byte-identical to today; rate — [14.2 characters per second]; both — [4.7 seconds duration; 14.2 characters per second]. Decisions were each settled by measurement. Characters are len(chunk) of the emitted text, spaces and punctuation included: over 50,112 corpus segments the three candidate character rules are within 1 % of each other on stability (coefficient of variation 0.3150 counting everything, 0.3132 stripping spaces, 0.3159 alphanumerics only; counting words is worse at 0.3572), and len(chunk) is the only definition recomputable from the prompt itself. The rate is per segment and is denominated by the printed duration, rounded to the one decimal the tag shows, because computing it from the underlying float differs by up to 0.6 cps and would put a number in the prompt that the prompt cannot reproduce. [P§32.4]

A segment gets no rate below 1.0 s, below 10 characters, or outside 4–50 cps, and falls back to [D seconds duration] in every mode. The rate of a short segment is boundary noise: coefficient of variation by segment length is 3.178 under 0.5 s (standard deviation 113 cps; the 99th percentile is 128 cps, which is not speech) against 0.239 at 2–3 s and 0.207 at 3–5 s. The rule costs 6.88 % of segments — 4.53 % under a second, 0.62 % with too little text, 1.64 % physically implausible — and leaves 0.79 % of rows with no rate at all. Language is not corrected for: English runs 22.28 cps and German 20.73 (medians 21.45 and 20.15); the Language: field already states which. [P§32.4]

A bug this design avoids: directions._SEG_RX is \[(\d+(?:\.\d+)?) seconds duration\] and requires the bracket immediately after "duration", so it matches neither extended tag; rendering in rate mode and then annotating would take the untimed branch and put one direction at the head of the line instead of one per segment, silently making direction placement a function of rate mode on two thirds of rows. The prompt is therefore rendered by the production path in dur mode, with prompt_lib2, anno and directions untouched, and the finished string's tags are rewritten afterwards. Controls (selftest.py, 3,000 corpus rows, all pass): dur-mode prompts byte-identical to prompt_lib2; two independent implementations of the rate transform agreeing on 6,000 renders; tags-stripped content identical before and after; every both tag's rate equal to len(chunk) over the printed duration; and 3,091 directions inserted in each of the three modes. Shipped mix: 33.34 % dur, 33.38 % rate, 33.29 % both, and within every voice the three stay between 0.301 and 0.371. [P§32.4]

Dataset 3 — speed preference. chosen is the take as generated, rejected the same take with one speech segment time-stretched by 20–30 % up or down, prompt byte-identical on both sides. The stretch is librosa.effects.time_stretch, a phase vocoder; audiostretchy is not installed and nothing was installed to get it. A deviation from the brief is stated plainly: the brief said to build this only on the good-quality material from dataset 1, which would have yielded a few hundred single-voice pairs. Quality is instead held constant directly: every corpus take has an audiobox-aesthetics score in $NB/sftdpo/out/sft/part-*.parquet, uid-keyed with 100 % coverage of the 1,200,530 rows, and this dataset takes only qual_overall >= 3.2 — 303,483 takes, the top 25 %, across all 500 voices. [P§32.5]

Three faults were found by measurement, all of which would have shipped silently. (1) The corpus's stored codes are not the codes the tokeniser produces from the corresponding audio — 74.7 % of frames match exactly, 92.3 % per codebook — so taking chosen from the corpus would have made the negative differ by an mp3 decode, a splice and a stretch; both sides now go through the identical decode / cut / crossfade / re-encode path with the stretch factor 1.0 on the chosen side. (2) Total clip length gave the answer away: stretching one segment moved the whole clip's length by 9.35 % on average (max 23.2 %) while the shared prompt's Tokens: field states the chosen frame count. The time is now taken back out of the silence, proportionally across the pauses; residual length difference is 0.003 s, 0.03 % of the clip, 100 % of pairs compensated, at the cost of a direction imbalance of 30 % slower / 70 % faster at that stage. (3) The compensation clicked, twice: np.tile room tone put the rejected clip's worst-discontinuity ratio at 7.14 on faster pairs against 1.76 for the chosen clip; phase-vocoder re-timing without boundary crossfades put both directions at 4.85; re-timing with crossfaded boundaries lands at 1.93 against 1.76. The stretch itself measured 1.53 throughout, i.e. clean. [P§32.5]

A hard per-pair gate drops any pair whose rejected clip exceeds twice the chosen clip's discontinuity ratio (or 4.0 absolute), or whose residual length exceeds 2 %. The half-speed canary holds: samples_per_frame is exactly 3840.0 = 48000/12.5 on every original clip and 3860 on stretched clips from frame padding — not the 7680 that would mean a stereo frame read as mono. Seam clicks with the 5 ms crossfade sit at 0.178× the local 99.9th-percentile jump (max 0.98); without it, mean 1.123, 95th percentile 5.587, max 35.4. [P§32.5]

Setting the speaker guard against a null, not against a number. The first Sidon run over the 500 profiles measured ECAPA cosine between each original take and its repaired version at 0.844 on average, with nothing reaching the 0.95 that had been assumed as a guard. Read as an absolute threshold that says Sidon destroys speaker identity, and it would have deleted the dataset. It says no such thing: Sidon resynthesises through a DAC decoder, so the embedding moves even where the speaker plainly does not. The baseline adopted is the ECAPA cosine between different original takes of the same voice, and the guard is a margin against it; every row carries speaker_cosine_orig_vs_repaired, speaker_cosine_same_voice_baseline and their difference. The protocol calls this "the same shape of error as the 'steering does not work' false negative in section 26: a threshold read off an absolute scale, with no control to say what the scale means." [P§32.6]

The sampling band. The first pass took the lowest-quality takes of each condition and got takes averaging qual_overall 2.36 against a corpus mean of 3.09 and a 25th percentile of 3.00 — broken audio as much as dull audio. The band was moved to the 5th to 50th percentile within each voice, and sampling is spread over conditions by taking one take per condition before a second from any. [P§32.7]

What makes each of the three a worse idea than it sounds. Dataset 1 is built on a selected subset — the takes where DNSMOS clearly improved — so its pairs are systematically the ones where the original was worst, and the emotion guard is inherited rather than measured on the 500-voice arm (null emotion_cosine_to_original). Dataset 2 is a new prompt format that nothing has been trained on; the rate surface removes the per-segment duration, and characters per second cannot tell a fast talker from a dense sentence read at normal pace. Dataset 3's negative is synthetic: the identity control shows that running the whole path with stretch factor 1.0 still moves 62 % of frames, the direction imbalance is a real skew, and a compensated pair's pause tags no longer exactly match its audio. "Of the three, dataset 2 is the one whose material genuinely supports what was asked; dataset 3 is sound but synthetic; dataset 1 is the one whose premise did not survive contact with the tree." [P§32.8]

A caveat on the speaker null. The baseline takes differ in words and performance as well as recording, so it measures "same speaker, different content". Over 82 voices it sits at 0.391, while original-vs-repaired sits at 0.892 — a margin of +0.50 with 100 % of pairs clearing it. That is a permissive reference and a lower bound on identity preservation; a stricter null (the same take repaired two different ways) was not run. [P§32.10]

Dataset 1 covers 130 distinct conditions, mean DNSMOS gain among the selected pairs +0.253, English 13,872 / German 10,327. Dataset 3 covers 130 conditions, 58.8 % faster / 41.2 % slower, German 24,845 / English 17,456. The speaker guard excluded exactly one pair of 24,200; it functioned as a check that passed, not as a filter (margin +0.49, nothing near the null). Every string column is written as arrow string and every shard at the corpus's own row group size (200 for the SFT container, 150 for the DPO ones), because corpus_ds2.list_units enumerates (file, row_group) pairs: a single-row-group write would have offered 128 units instead of about 15,900 and pulled 9,339 rows per read instead of about 200. [P§32.11]

The guard, quantified on the arm where both halves can be measured (this subsection sits physically inside §35 in the protocol file). dnsmos_pass.py was run to completion over the one-voice dsbuild arm — all 106,424 clips, 100 % coverage, 443 of them (0.4 %) failing to score and excluded rather than imputed — and the guards re-applied on DNSMOS. 26,594 candidates carry all four variants. [P§32.12]

The converted renderings enter with 3,440 pairs that genuinely improved DNSMOS and leave with 19; Sidon alone enters with 1,685 and keeps 918. What only the four-variant view shows: vc_sidon — Sidon applied on top of converted audio — is the most frequent winner on raw DNSMOS, 2,482 of the 5,125, more than Sidon alone (1,685); it also inherits voice conversion's speaker shift (+0.232 mean, inside ±0.05 on 1.66 % of takes), so those 2,482 become 18. "The best-looking repair by DNSMOS is the one that must not be used." [P§32.12]

II.8.2 Does stacking adapters dull the sound? No — it destroys the words instead — [P§33]

Prompted by a listening impression that the demo audio had lost some high end. 128 conditions × 2 samples = 256 clips; 16 fixed prompts across the emotion heads, English, seed 1234, generated through the same generate_steered path with an empty injector that the steering study used. Code code/spectral_stack.py, raw cells stats/spectral_stack.jsonl, report status/status_2026-08-29_spectral.html. A new instrument was needed because every scorer in the project — genuineness, blend, aesthetics, WER — is blind to a gentle low-pass. Added: long-term average power spectrum per clip reduced to band-energy ratios, spectral centroid, 85 % and 95 % rolloff, and the log-spectrum slope from 2 to 16 kHz; ratios and frequencies only, never absolute level, because adapters change loudness. [P§33]

The hypothesis was that the codec's deep codebooks carry the fine detail and are the most fragile, every adapter targets all twelve audio heads, stacked deltas sum, and a degraded fine channel would read as dullness. It does not. [P§33]

High-frequency energy rises with the stack, +0.0095 paired at seven adapters (t +2.2); the centroid and rolloff rise with it. What collapses is intelligibility: 0.063 → 0.554. Genuineness barely moves (4.16 → 4.04) and blend rises (6.59 → 7.17) — "the stack makes the model say the wrong words while still sounding fine to the quality scorers." This reproduces, on a different prompt set and a different harness, the demo server's 0.013 → 0.258 over its own stack. [P§33]

The model's output sits between real speech and its own synthetic training material, closer to the synthetic side. If the audio sounds veiled next to a real recording, the candidate mechanism is the corpus rather than the adapters — a data property every checkpoint trained on it inherits — and the lever would be corpus weighting or a high-frequency filter on the synthetic arm, not an inference setting. Limitations: nobody listened; a spectrum that is objectively brighter can still sound worse if the added energy is noise, which at WER 0.55 some of it very likely is. One stack order, one emotion, one voice profile, 16 prompts × 2 samples per condition. DNSMOS was not used (no local implementation in this tree). The white-noise arm of the control failed on a codec API detail; the corpus arms completed. [P§33]

II.8.3 2026-08-29 — The protocol audit, and the state of the archive — [P§34]

Agent A5. Three jobs: check every number in the protocol against the file it came from, make the material findable for a paper, and get everything that exists in only one place into the repository. Deliverables: ~/protocol_audit.md, ~/PAPER_INDEX.md, ~/paper/ICLR_OUTLINE.md. Sections 22–33 were audited claim by claim (every cited path resolved or reported dead, every table recomputed from raw rows where raw rows exist); sections 1–21 were spot-checked. The asymmetry is deliberate and is itself a finding. [P§34]

Everything recomputed from raw rows reproduced exactly: all of 29.2 (seven metrics × nine columns plus all three family rows), 29.6's cross-check, 28.2's pooled ladders and shape classification, both tables of 33, 30's block arithmetic, 26.1's eight headline findings and three baselines, and 22.1's four dataset tags. "The tables are sound. The errors are all in prose that quotes another section." [P§34.1]

The lesson the audit keeps: the project's correction discipline is good — every documented correction landed on every page, old values shown struck rather than silently replaced — but nothing re-derives a number that another document quotes. "Every quoted figure should be computed at render time or carry its source path." [P§34.2]

Fixes applied. Thirteen in the protocol, no prose rewritten: 22 (Phase 2 labelled main); 22.1 (58,800 → 58,714 rows; 58,800 is worklist_main.parquet, a real number on the wrong object); 23.1 and 23.2 (probe and selection rule labelled); 23.7 (264 generations → 264 cells = 1,056 generations); 26 (the 3,034 cells credited to jobs 1527888/89/90, which were cancelled at 0 cells; the real runs are 1527978 / 1527979 / 1527908); 26.1 (3,034 configurations → 170 configurations over 3,034 cells); 28 (cell arithmetic; Whisper offset at w=0 0.0285 → 0.0265); 28.5 (two delivery axes 0.041 → 0.060); 29.5 (recommendations.json → comb_recommendations.json); 30 (the saturation figure); 31.3 (the delivery clamp: +0.375 / t 5.18 / 15 of 17 / +0.003 / t 0.55 → +0.381 / t 5.42 / 16 of 17 / +0.0041 / t 0.72 — the section that makes a recommendation to another team was wrong on all five figures). Twenty-four replacements across nine report and generator files, home and repo copies alike; the generators were fixed as well as their output. [P§34.3]

Four claims with no file behind them, none changed: 22.1's "285 files" (release.json names 167 artefacts; 276 predate its mtime); 18's raw variance components 0.04719 and 0.02553 (the derived 35.1 % is sourced to status_2026-08-26_evening.html, its denominator is not); 8's "all 1,041,100 dpo_emotion rows"; 21's first burst gate "1,124,863". The larger point is the citation chain: for sections 1–21 it very often terminates at a narrative HTML (technical_report.html, a status page) rather than at data. 3,147,802, 243,986, 247,049, 1,362,516, 1,061,613, 1,853,486 and 2,327,904 are all corroborated, but by a second prose document written from the same run. If the paper draws on corpus construction or training, that chain needs rebuilding first. [P§34.4]

One claim that looked false and is true. 26 says p3_vectors_ext.npz adds q:genuineness and q:blend; a top-level key check returns "absent" (eleven keys, neither among them), and that reading is wrong: the 99 dimensions are addressed as kind:name from two parallel arrays (actf_steer.py:226), and dim_kind == "q" with dim_name genuineness and blend are entries 97 and 98. Recorded because the false reading is the natural one and would have retracted a correct claim. [P§34.5]

Archive state. Six additions to research-log-2026-08: actforensics/ (protocol.md, report.html, README — this study was in no repository and is the upstream of all four layer-forensics workstreams and every steering result); combination-study/vectors/per_attribute_manifests/ (60 JSON, 643 KB); combination-study/grids/attrs_all.json and the 13 raw/*_r000.json seed records; layer-forensics/PLAN.md (the pre-registration that 23 opens by correcting); stats/voice_profile_names.md; the three audit deliverables. Confirmed already present: layer-forensics/w3/archetypes.npy (156 × 38 × 2560, byte-identical to the home copy), both 2026-08-29 status pages, all five studies' code/ directories, lora_dose_study at 113/113 files, and 80 of 126 top-level HTML pages byte for byte. Deliberately left out: the 741 MB full vector set (the committed topk/ form is 8.6 MB), p3_vectors_ext.npz at 112 MB, and 37 audio-grid HTML pages totalling 978 MB. ~/quality_speed_datasets.html was left alone as A4's in-flight deliverable. [P§34.6]

For the paper. ~/paper/ICLR_OUTLINE.md carries the section plan, eight figures (seven drawable today from committed files) and the gap list. Three things come first: (1) there is no listening test, and this is submission-blocking — §33 showed the quality scorers cannot see a failure that takes word error from 0.063 to 0.554, §18 showed the emotion instrument is saturated by text content; a usable blind A/B harness exists (~/listening_test.html, built 2026-08-09) and comb_recommendations.json names the operating points. (2) Three re-analyses needing no new generation: the matched-total-magnitude re-analysis of the k ladder; the reconciliation of 20's CFG effect with 30.1's, which differ by an order of magnitude off different baselines; and the blend tap of 34.2. (3) Five single-seed results are written as findings — 22's tasks A/D/E on one probe seed, 15's "rank 16 wins", the GRPO v4/v5 comparison, every steering configuration, and 29's factorial — while the tie standard of 23.1 ("treat a difference below ~0.01 as a tie") is applied rigorously in the layer work and not at all elsewhere. [P§34.8]

II.8.4 2026-08-29 — Shipping the three levers into the demo server, and the first WikiSkill draft — [P§35]

LAION-AI/Humaneness-Voice-Demo-Server had exactly one way to shape a performance: load LoRA adapters and write a good prompt. The combination study had measured two more levers and how all three combine, so the job was to give the server the other two as switchable modes, let its voice-acting agent choose between them by tool call, tell it in its own prompt when to use which, and produce the first draft of the WikiSkill knowledge layer. What shipped: pull request #3 against that repository, branched from main at 44ea226. New files levers.py (the policy), steer_engine.py (the vectors and injection points), setup/build_steering_pack.py, setup/check_levers.py, docs/LEVERS.md; changed config.py, tts_engine.py, llm_agent.py, app.py, timed_script.py and three docs. And ~/wikiskills/, mirrored to research-log-2026-08/wikiskills/: 60 pattern pages, an index, an interactions page and coefficients.json, all generated from the measured JSON by code/build_wikiskills.py. [P§35]

The default mode is auto and resolves per family, because the best single lever flips by family. Emotion → adapter+steer: adapter × steering is +0.038 at t 1.36, additive. Delivery → adapter: adapter × steering is −0.164 (t −3.7) and adapter × guidance −0.125 (t −3.5), both significantly sub-additive, so exactly one lever runs. Quality → adapter: steering does not move the quality axes. auto never spends guidance — 1.93× is not a default. Subtracting Emotional_Numbness at −0.10 is attached automatically to any steered emotion; when steering and guidance are both on, both branches are steered. [P§35.1]

Three design decisions. Guidance ships non-streaming: that repository's docs/ADAPTERS.md §1 records realtime factor 1.0 as the streaming budget and 0.764 as the live merged baseline, and forward hooks for the whole adapter stack were rejected at 1.75×; guidance at 1.93× lands at about 1.47 and the player starves, so a guided take is rendered whole, the payload says "streaming": false, and batching the two branches into one forward pass is left as the documented next step. The vectors are an asset, not a commit: the research library is 112 MB, the server needs 99 × 5 × 2560 float32 = 5.3 MB, which setup/build_steering_pack.py distils and MOSS_STEER_PACK points at; the distilled pack is committed to the research log at research-log-2026-08/wikiskills/vectors/. Refusal beats interpolation: seven of sixty attributes have no configuration that clears the balanced guardrails, four delivery _low tails have no measured steering route at all (the two tails are orthogonal, not opposite, median cos −0.0004), and in both cases the server runs the adapter, records why, and the wiki page says No usable setting. [P§35.2]

What is too thin to build a default on. The quality-axis null (+0.006, t 0.01) pools 30 cells — three attributes at ten prompts; the direction is safe, the precision of that zero is not. Guidance has no significant main effect on delivery or quality (+0.026 t 0.44, +0.062 t 0.94); the shipped g = 2.5 for delivery comes from the CFG study's own arm, not from the factorial. strength: gentle halves the recipe's alpha to 0.05 and is an interpolation — only four cells in the recommendation table use alpha = 0.05 — labelled as such in config.STRENGTH_ALPHA_SCALE. DELIVERY_LEVER = "adapter" is conservatism, not arithmetic: on a delivery axis steering measures stronger than the adapter (+0.614 against +0.377), but the adapter family was swept at scale on the demo box's own hardware and the vectors have not been anywhere. Per-attribute interaction rows are ten prompts. The random-direction floor is per-attribute for only 20 of 60; the other forty fall back to the pooled −0.033. 1.93× is the study's hardware. And 29 of 40 emotion adapters have no usable merge weight in the dose sweep, yet the server loads them at 1.0 — defensible because no adapter in the 5,740-cell sweep was harmful at any weight, but the adapter mode on most emotions is doing less than its name suggests. [P§35.4]

What could not be tested. Nothing in the pull request has run on the demo server's hardware. The offline check passes in both the assets-present and assets-absent configurations. The PR asks for six smoke-test items: that the check script runs there, that adapter mode is bit-identical to main, what the steering hooks cost in realtime factor, that a guided take completes and is intelligible, that the neutralised branch looks right in end.prompt_unc, and that the director uses the new tool call without reaching for guidance on a cheerful one-liner. The steering hooks' cost is arithmetic, not a measurement: one to five hooks, each a norm and a fused multiply-add on a [1, 1, 2560] slice, against the 536 kernel launches per token that took that server's realtime factor from 0.737 to 1.29. Nobody has measured it. [P§35.5]

"The instruction you hand someone else is an artefact and needs a test like any other." One thing the demo team flagged about themselves: they had *not* byte-compared the adapter path against the pre-modes tree. setup/ab_codes.py now does that comparison — fixed seed, prompt and adapter set, comparing generated RVQ code tensors rather than audio, importing nothing that generation modes added so it runs unchanged on the older checkout. That test still has not been run; it is now possible to run it. [P§35.6]

II.8.5 Organic emotional transitions — making a performance turn inside one utterance — [P§36]

The question: a person who is grieving and then finds something bitterly funny does not switch; the grief is still under the amusement at the end. Everything the model could do before was a step — prompt sentence one for A, sentence two for B, generate, concatenate. This round builds the fade nine ways and measures which construction reads as one performance. Design: ten emotion pairs × three purpose-written texts × nine methods, plus two controls and two anchors: 399 cells at 4 samples each, zero failures. Texts byte-identical across arms, the voice pinned by reference codes, the seed a function of the text slot only, so arms are paired sample for sample. [P§36]

Machinery: the decode loop is actf_steer.generate_steered with one line added, inj.set_frame(sched.weights(t)) at the top of each frame; the injection arithmetic h <- h + α · unit(v) · ||h|| is unchanged. The three-branch guidance arm keeps cfg_run.generate_cfg's channel-level interleave verbatim and extends it to l_unc + g_A(t)·(l_A - l_unc) + g_B(t)·(l_B - l_unc) with g_A + g_B = 3, the CFG study's WER-constrained optimum for emotion (§30, Q1). Both emotion directions are injected at one shared layer per pair, chosen as the layer minimising the sum of the two attributes' held-out R² ranks. [P§36]

Result. Ranked on the median turn per text, not the mean, because the paired differences are heavy-tailed (M3's mean Δr is 5.41 with a median of 2.32; M9's are 14.46 and 2.71). Judgement is the paired sign test against the random-schedule control. M6 (multi-rate) has the largest typical turn, Δr 2.53 per text (mean 3.93), against the matched control median +0.74 (mean +2.53, 21/30 up, sign p = 0.043); runner-up M4 at 2.38. Four arms beat the random-schedule control consistently (M6, M4, M3, M2). No arm significantly beats the step baseline — the best is M6: median +0.22 (mean +2.22, 18/30 up, sign p = 0.362); M4: median +0.52 (mean +1.91, 20/30 up, sign p = 0.099). The gains are of order 0.80 where the scorer's own scatter on a non-changing clip is 1.82: "a real effect on a blunt instrument, not a solved problem." [P§36]

Two of eleven methods sit below the random-schedule control: M1 and M7. For M1 that is correct behaviour — it holds both adapters on at a constant weight and has nothing time-varying in it. M7 is the informative one: its only lever on the emotion is the wording of the brief, and the anchor result is precisely that the brief is not a lever. The smoothest arm that beat the control was M2 (1.73); the arm retaining most A at the end was M6 (0.13), the only positive value in the column. The random-schedule control — same vectors, same total magnitude, same multiset of weights, A-to-B order destroyed — travelled Δr 1.40, inside the floor, i.e. null: the movement in the working arms comes from the ordered structure of the schedule. [P§36]

Adapter re-weighting mid-generation was tested (M10 against M1, the same two adapters held constant): Δr median −0.58 (mean +0.32, 12/30 up, sign p = 0.362) against the step, median WER 0.018 against 0.017. It did not damage the audio, but neither arm cleared the floor — "a weak acquittal rather than a licence." [P§36]

Five predictions recorded before the run; 2 of 5 were wrong. (1) PARTLY — M0 (the step) would not be as bad as it sounds, with the seam showing in smoothness more than the endpoints: M0's typical turn is Δr 1.92 against 1.59 for the random-schedule control and 1.36 for a clip held at one emotion, so it is not separable from the controls, and its smoothness 1.69 is *better* than the working fades (1.73). (2) PARTLY — M2 would show a mid-clip bump and be beaten by M3: the bump is a 20 % dip, and M3 does beat M2 pairwise (median −0.14, mean −3.27, 14/30 up, p = 0.856). (3) WRONG — M7 (prompt-arc) would do better than its cost suggests as the only in-distribution arm: M7 is the worst arm, Δr 0.67, below the control (1.59); against the control median −0.24 (mean −0.38, 14/30 up, p = 0.856). (4) RIGHT — M6 is the one to build a product on: largest typical turn (2.53, ranked 1 of 10), beats the control (p = 0.043), the only arm whose A-ness at the end is positive (0.13); but it does not separate from M3 pairwise (median −0.48, mean −1.49, 12/30 up, p = 0.362), costs the most word error of the vector arms (0.040 median), and does not significantly beat the step. (5) WRONG — the floor in M4 would matter more than the fade shape: A-ness in the final window M3 −0.14, M4 −0.16, M5 −0.02, M6 0.13; adding the floor to M3 moved retained A-ness by median +0.00 (mean −0.02, 14/30 up, p = 0.856) and the distance by median +0.01 (mean −1.80, 15/30 up, p = 1.000). [P§36]

Honest limits. Nobody has listened blind; the grid at ~/emotional_transitions.html exists so that they can. Every number is a scoring model used at a duration (3 s) it was never validated at. Ten pairs are dramatic archetypes, one language, the fade schedule pinned to script frames (safe because mean duration error was −0.08 s). Next, in order: run M9 (three-branch CFG) on the full material — the only arm that moved decisively and it was starved of cells; fix the instrument before trusting further trajectory work; choose reference clips by content, or drop reference conditioning from emotion studies entirely. [P§36]

II.8.6 2026-08-30 — Duration control is already solved; what is not controlled is tempo — [P§37]

Measured while designing the evaluation for the rate adapter, and it revises the adapter's stated purpose. Duration obedience from the 1,596 clips of the transition study (SFT3 + DPO-v2 + LoRA, no rate adapter, requested totals 9.6–12.0 s): [P§37]

That M0 alone deviates is what rules out a measurement artefact: if the harness were padding or truncating to the target, M0 would read 0.000 too. How the obedience is achieved was measured by energy-based voicing analysis of 133 of the study's 399 published clips (20 ms hops, threshold 6 % of the 95th-percentile hop RMS, the same heuristic family as et_gen.trim_edges). [P§37]

The model speaks at 21.0 characters per voiced second and pads the rest with silence; it does not slow down to fill the budget, it stops talking. 38 % of every clip in the transition study is silence. Two independent pipelines agree on the speaking rate: 21.0 chars/voiced-s from decoded audio versus cps_median = 20.70 computed by anno_rate over the 121,940 rows of ds2_rate_sft_120k. This also explains et_prompts.CPS = 14.9, a chars-per-total-second constant (13.0 measured) that was never the speaking rate. [P§37]

Consequences as written at the time: (1) the dur arm of the rate corpus teaches nothing the base model cannot already do — it is a control, and the rate and both arms are the only ones that address an uncontrolled quantity; (2) the evaluation must measure chars per voiced second, with the prediction to falsify that the base model's tempo stays flat near 21 across a requested 14–26 cps sweep while the adapter moves it; (3) every clip in the transition study asked for about 38 % silence, a specific candidate for the "two performances rather than one" impression. [P§37]

II.8.7 2026-08-30 — The rate tag is a re-encoding of the duration, not a second control — [P§38]

Written 2026-08-30 14:25, before the paired half of the tempo evaluation (job 1540103) finished, so the prediction is on record ahead of its test. §37's prediction is already falsified for the base condition on partial data, n ≈ 20 cells per mode, base only: [P§38.1]

Why the prediction was unsound: §37's 21.0 cps came from the transition study, which built every prompt with the same constant (et_prompts.CPS = 14.9), so there was no variation in the requested rate at all; one operating point cannot establish a slope over a range the data never covered. What the base actually does is compress the range: it will not go slower than about 17.6 cps however much time it is given, but nearly reaches the request at the fast end. [P§38.1]

All three modes carry the same signal because the harness varies the budget. The base model has never seen a rate tag — it only knows 073aeb09dc923376 — so its response in rate mode is to the Tokens: budget, which varies with the requested rate in every mode. The discriminating quantity is therefore the interaction (adapter rate − adapter dur) − (base rate − base dur), not the slope. [P§38.2]

The prediction: that interaction is about zero, by construction of the corpus. anno_rate.seg_rate computes cps = n / dur_s from the printed chunk text and the printed duration, both already in the prompt. In both mode the rate tag is strictly redundant with the duration tag beside it; in rate mode it is an invertible re-encoding — D = n / R with n visible — at the cost of making the model count characters, so rate is harder than dur, not more expressive. No row in ds2_rate_sft_120k can have a rate that disagrees with its duration. The corpus teaches an alternative surface form, not a second control axis. What the adapter is then worth is an interface gain: a voice-acting agent can ask for "26 characters per second" without computing a duration. Falsification: if the interaction term is clearly positive, this analysis is wrong. [P§38.3]

II.8.8 Result: the rate adapter changes nothing measurable — [P§39]

Job 1540103 (budget follows the request, 240 cells / 120 paired) and job 1540486 (budget pinned, 160 cells / 80 paired); base = SFT3 + DPO-v2, adapter = ratelora_sft_r16/step3788 at w = 1.0; 8 texts × 5 requested rates × 3 samples, Parakeet WER, paired within cell on a shared seed. [P§39]

Paired adapter − base: dur +0.45 (t +1.63), rate +0.36 (t +1.10), both +0.21 (t +0.66). Interaction (rate − dur) = −0.09 cps. The tag form makes no difference. [P§39.1]

All four slopes are zero. Told 14 or told 26, the model speaks at 18. Paired adapter − base: +0.20 (t +0.64) and −0.49 (t −1.47). With the budget removed as a channel, the rate tag is inert — for the base as expected, and for the adapter that was trained on it. [P§39.2]

What is and is not controllable: (1) total duration is fully controlled by the Tokens: budget, dur_err 0.00 s in every cell of both jobs, needing no adapter; (2) tempo only indirectly — a tighter budget forces faster speech (slope about 0.66) but there is a floor at about 18 cps below which the model pads (silence fraction 0.35 in the pinned arm); (3) the rate tag is inert, confirming §38.3 in the one configuration where it could have carried information. [P§39.3]

Why it could not have worked: anno_rate.seg_rate sets R = n / D from the printed chunk and the printed duration, so in every training row the rate agrees with the duration and the model was never given a reason to read R. A corpus that teaches tempo must vary R independently of the natural duration — the same text, a generous fixed budget, differently time-stretched audio, and the tag naming the stretched rate. That data already exists: the 42,301 compensated=True pairs in ds3_speed_dpo_final, each with its stretch_factor (the protocol calls them "audiostretchy pairs" here, although §32.5 records that the stretch was librosa.effects.time_stretch and audiostretchy was not installed). A second attempt should be built from those. [P§39.4]

II.8.9 Crossfade v2: the same performance turn, without the steering vectors — [P§40]

Why it was repeated: §36 ranked nine constructions and then reported that its own denominator was noise (anchors 0.12 SD apart, 0 of 30 texts clearing a unit). Two things then changed the design space. Steering vectors were dropped from the demo server (GEN_MODE back to adapter) after a listener called their output "more emotional, with strange artefacts, and the timbre off", which removes five of §36's nine arms (M2–M6) outright. And classifier-free guidance came in, with a field ceiling of about g = 4.0 against the CFG study's WER-constrained optimum of g = 3.0 (§30). Design: 13 arms × 4 guidance levels (0, 2, 3, 4; CFG3 has no g = 0) × 30 texts × 4 samples = 1,530 cells / 6,120 clips, on v1's material byte for byte ($SC/etrans/out/prompts.json), the turn falling on an explicit 0.5 s pause. v2 fades adapter merge weights — XF, XF_FLOOR, XF_ASYM, ARC_XF, BOTH, ARC, STEP — plus CFG3. Adapters are faded, never merged; the audio heads are weight-tied to the embeddings and merging destroys them. [P§40]

Repairing the instrument. Four anchors instead of two: ANC_A/ANC_B hold one emotion end to end with that emotion's own adapter, ANC_Ap/ANC_Bp with the brief alone. Separation is read on four windows (first 3 s, first 6 s, first sentence, full clip) and three head sets (own, av, loo), and the anchors are generated first across all texts. Re-analysis of v1's stored records had already shown the 0.12 was mostly the 3 s window — brief-only with reference on reaches 0.890 SD on full clips, 12/30 texts ≥ 1 SD on loo, and brief + own adapter 0.762 on own — so v1's headline understated the model by roughly 2.5×. A prediction was recorded before v2 ran: with adapters own should beat loo; brief-only should reverse. [P§40]

What guidance costs. Measured as t_gen_s / frames_tot, batch of 4 on one GH200, paired on arm × text over 127 cells: pooled medians 51.5 → 98.6 ms/frame, a 1.92× multiplier, independently replicating the CFG study's 1.93× (§30, Q4). The multiplier is identical at g = 2, 3 and 4 — the price is the second branch, not the strength. Report ms/frame as medians: ARC_XF's g = 0 mean is 70.1 against a median of 55.5 because one warm-up cell hit 106.8. [P§40]

Engineering, dearly bought. (a) sbatch --export splits its value on commas and an --arms list is commas — v1 lost 13 of its 14 arms to exactly this; v2 passes --export=ALL plus an args file. (b) import numpy from the study venv on /e/data1 was measured at 244 s and again at 448 s, jobs spending 82–105 minutes in imports; the analysis half (xf_agg.py, xf_report.py, xf_page.py) was moved to the EasyBuild stack on /e/software, where import numpy takes 0.087 s and full aggregation takes 0.24 s on the login node. (c) Copying the venv to scratch failed with Unexpected EOF in archive. (d) Two of the first three 4-node jobs produced zero cells, one stalling 94 minutes in d_alloc_parallel/cxiWaitEventWait; the mitigation is a longer wall clock, not more nodes. (e) Resume is by cell key (--tag <new> --resume-tags x1,x3). [P§40]

Result — the anchor separation first. On the pre-declared primary readout (loo on sent), the brief-only anchors stand 0.797 / 1.013 / 1.934 / 2.239 SD apart at g = 0 / 2 / 3 / 4, with 23 of 30 texts clearing a full SD at g = 4 (on the full clip, 3.428 SD and 25/30). Against v1's 0.12 SD, 0/30 that is about a 19× improvement, decomposable into three causes: the window (v1's own settings reproduce here at 0.188 — its 3 s slice of a 30 s zero-padded scorer input was 10 % signal); guidance, which raises separation monotonically on every contrast × segmentation × anchor set, making g = 4 the measured optimum; and the adapters, which turned out to hurt. [P§40]

The grid, read against a floor. An anchor does not turn, so its |dr| is pure noise: 0.71–0.88 on loo/sent. STEP — the two-take splice already shipping — is the only arm that clears it, at every g (median dr 1.421 / 1.306 / 1.945 / 2.623; rho6 rising 0.36 → 0.73). The best non-STEP cell is CFG3@g4 at 0.818, below the 0.838 a non-turning anchor scores at the same g. Every XF* arm, BOTH, ARC and ARC_XF land in −0.41..+0.70, the range the anchors occupy — and so does the shuffled control, at 0.599 and 0.668, above nine of the twelve genuine XF* cells. Paired against STEP, every arm loses at every g, and the deficit grows with guidance (g = 0 mostly n.s.; g = 3–4 almost all p < 0.01). [P§40]

Against the shuffled control the question closes. C_SHUF runs XF's own weights with the A→B order destroyed. Across 48 arm × g comparisons the only arm that beats it is STEP, at all four guidance levels (+0.536 to +1.736, 20–21 of 28 texts up, p = 0.013–0.036); the fading arms' best showing is XF@g0 at +0.165 (16/28, p = 0.57), and XF@g4 sits at +0.044, 14 of 28 texts up, p = 1.000. This strengthened at every refresh: at 82 % of the grid STEP's advantage was p ≈ 0.06–0.08 on 16–19 texts, at 92 % p ≈ 0.036 on 26–28, at 98 % p = 0.013 at g = 4, with no fading arm moving off zero. "The fade shape does not matter, because the fading carrier does not move the emotion." [P§40]

Costs and safety. Guidance multiplier confirmed per arm at 1.90–1.94×, identical at g = 2, 3, 4, 30/30 texts up, p < 0.001 in every arm. Guidance costs zero WER (median Δ +0.000 in 33 of 36 arm × g cells). Adapters cost more than guidance does unguided: 30 ms/frame with none, 39 with one, 48–49 with a live fading pair; STEP is simultaneously the cheapest arm (30 unguided / 57 guided) and the best-scoring; CFG3's third branch is nearly free (83 ms). v1's M10 acquittal holds at 4× the scale: changing LoRA scaling mid-generation against a KV cache built at other weights does no measurable damage (XF* WER 0.015–0.037). STEP's one real cost is duration, −0.33 to −0.54 s, the eaten pause at the splice. Half-speed canary: 0 cells. [P§40]

Recommendation: ship STEP at g = 3 (median dr 1.945) — the only method that measurably turns, the only one that beats its own shuffled control, the cheapest in the grid, and what already ships. g = 4 scores higher at identical cost but sits at the listener-reported edge. Do not ship an adapter crossfade in any shape; do not ship ARC/ARC_XF. What the metric cannot decide: STEP wins by construction — two independent takes are maximally different — and whether the splice sounds like one performance or two clips is a listening question. The listening page (508 clips, one binary per clip) asks exactly that. Status at writing: 1,476 of 1,500 cells (98.4 %), all 480 anchor cells complete, 24 of the 51 arm × g cells at the full n = 30; every number was computed three times — at 82 %, 92 % and 98 % of the grid — and no conclusion moved. [P§40]

II.8.10 2026-08-31 — The quality adapters: a null measurement, then a listener who disagreed — [P§41]

(The protocol numbers this section's subsections 40.1–40.3; they are cited here as §41.1–§41.3.) Two evaluations of the same two adapters, five hours apart, disagree, and the difference is what they were stacked on. Evaluation A ("§39 harness", 16 prompts, adapters on bare SFT3): no adapter beat the base model on DNSMOS, best +0.017 (t +0.45), worst −0.130 (t −2.05); all ten evaluated states raised word error slightly, +0.001 to +0.042, none significant, ten of ten in the same direction. Evaluation B: 60 clips, adapters on the full demo-server stack (SFT3 + DPO-v2 + speaker + emotion + genuineness/blend/esthetics, doses from the server's own config.py, prompts built by its own timed_script). [P§41.1]

Paired within (emotion, take) on a shared seed, n = 20 per condition: B lowers word error by 0.027 at unchanged genuineness; C raises genuineness by 0.42 (up in 13 of 20 pairs) and pays +0.016 word error (no t or CI recorded for the paired differences). The two adapters pull in opposite directions — B toward cleanliness, C toward expression. Three explanations that cannot be separated: different text material; n = 20 is small; or an adapter behaves differently inside a five-adapter stack than alone on the base model, and evaluation A never tested the deployed configuration. "Measure the stack you ship." [P§41.1]

II.8.11 2026-08-31 — rew_chosen is measured against bare SFT3, not against the run's own starting point — [P§42]

Found by the burst+stop agent while deciding how to select a checkpoint; it changes how every DPO run in the log should be read, including §41's. In LoRA mode dpo_train2.py never materialises a separate reference model: _REF stays None and the reference log-probabilities come from peft.disable_adapter(). That disables all adapters, so the reference policy is bare SFT3 — not the adapter the run was initialised from. [P§42]

What to use instead: per-family preference accuracy on a held-out slice, scored offline through each kept checkpoint with the same pairs. The in-training eval cannot do this — it is capped at 288 pairs and reports pooled numbers only, and fam_acc in loss.jsonl is one rank's 8 pairs per heartbeat. famscore.py (+ bsfam.sbatch) does it forward-only on one GPU and writes one record per pair per checkpoint, so a later listening test or a different metric can join against it without regenerating anything. [P§42]

II.9 The tools programme after 29 August, part 2 — the vocal-burst adapters: dose, labels, manufactured data, guidance on the burst (protocol §43–§56)

This chapter follows the vocal-burst adapters from a field report ('four classes do not come out') through fourteen studies. The arc is: the merge weight is swept and is not the lever (§43); a preference run teaches stopping and burst class on different schedules (§44); the quality adapter's weight sweep cannot separate doses (§45); a listening page ships dead (§46); the training labels are audited against the detector and sixty adapters turn out to have been trained on a label the detector contradicts (§47); re-filtering leaves 9.8 % of the rows (§48); manufactured bursts are built and score 0.000 on the strict metric while working at family level (§49, §50); the family-relaxed hit rate is adopted as the metric (§50a–§50c); the other levers are stacked (§51, §52); classifier-free guidance triples the hit rate (§53); an external generator and an external annotator are tried against the detector's blind spots (§54, §55); and adapter rank is shown not to be the bottleneck (§56). Sections are presented in numeric order; the protocol's file order differs and is noted in the module docstring. [P§43–§56]

II.9.1 The vocal-burst merge weight: a dose sweep under the real stack, and a metric that could not see the answer [P§43]

Full write-up ~/reports/burst_dose_study.md; data $SC/out/burst_dose/bd_r000..015.jsonl (2,549 rows), aggregate bd_agg.json, code $SC/code/burst_dose/, with SC=/e/scratch/reformo/schuhmann1_moss. A field report said four burst classes — scream, shriek, frustrated_groan, exhausted_groan — do not come out, and the suspected lever was the burst adapter's merge weight. Reading lora_bank.py and config.py before touching a GPU turned up three discrepancies: (1) the documentation and the code disagree by a factor of two — the docstring and the plan_blend() comment quote the manual's 0.5 inline / 0.75 solo, while config.py ships BURST_LAM = 0.25 / BURST_LAM_INTENSE = 0.5, and neither halving is recorded in docs/; (2) two different rules decide 'solo' — lora_bank.plan() uses < 14 plain words, plan_blend() uses < 10, and docs/ADAPTERS.md documents only the 14-word rule; (3) 'the burst stands as its own beat' is implemented purely as 'the script is short', so a burst at the end of a long line is always dosed as inline. §28 had swept this family against the bare SFT3 export, and its own §28.5 says 'nothing here predicts a stack.' [P§43.1]

Design. 264 conditions × 8–10 prompts = 2,420 cells, four arms crossed with six classes (the four named failures plus chuckle and contented_sigh as working controls) × two script kinds (inline = a 23–57-word corpus line; solo = a 5-word lead plus the burst) × 8–10 held-in-common prompts. Every d is a within-prompt paired difference against the same condition at w = 0 on the same seed; n = 55 pooled prompt-level pairs. The decision rule — largest pooled gain in r_burst_cls among weights that are significant (t ≥ 2.00) and hold all six per-class guardrails — was fixed in bd_agg.py before the grid ran, with §28's 3σ guardrails and an absolute WER_pk ≤ 0.25 cap on inline. [P§43.2]

Controls, all passing. The stack_nolora arm gives the zero-sharing check: a five-adapter stack with the burst adapter loaded and scaled to zero is bit-identical (0.000e+00 on every metric, all 12 prompt sets) to the same stack with no burst adapter in the model, which licenses w = 0 as a baseline. The half-speed canary (samples_per_frame = 3840, §21) was asserted per clip: 0 violations in 7,260 takes. The mismatch control at w = 1.0 gives d r_burst_cls −0.044 (t −2.19) inline — a wrong-class rank-16 perturbation does not raise detections. [P§43.2]

The single qualifying weight in 36 tests is in the arm the demo server never runs, and the same weight in the corresponding stack panel is −0.0022 (t −0.14): §28.5's warning, confirmed with a control. The study is powered: the SD of the paired differences gives a minimum detectable pooled d of +0.031 to +0.050 depending on panel, and §28's bare-stack burst curve rose about +0.12 across the same range — 'That effect size is comfortably inside this study's resolution and it did not reproduce in the stack.' [P§43.3]

The methodological result. On r_burst_cls, stack/solo shows nothing (|d| ≤ 0.024, |t| ≤ 1.52). On hit-rate — the fraction of prompted target-class cues that came back as a detection of the right class — the same 7,260 takes give +0.061 at w = 0.5 (t 2.32), +0.061 at 1.0 (t 2.46), +0.091 at 1.5 (t 2.98), rising with dose. Reading reward.burst_realisation explains it: it is an F1, so every extra detection enters the precision denominator; a matched detection of the wrong class still earns 0.35 credit, so converting a wrong burst into a right one moves the score by 0.65 of a duration term; and that term exp(-|dur - prompted| / 0.5) drives a right-class burst of the wrong length toward zero. The dose does exactly the two things the metric punishes. [P§43.4]

frustrated_groan produced zero correct bursts in 1,188 prompted cues, every arm, both kinds, all ten weights; its d hit is exactly +0.000 at all nine weights because the numerator is zero at both ends. 'The dose parameter cannot lift a floor of zero.' The failures are not silence but substitution down the arousal axis: inline frustrated_groan → Contented Sigh ×127, Wistful Sigh ×14; inline exhausted_groan → Low Mumble ×135, Ahem ×25; inline shriek → Surprised Gasp ×86, Contented Sigh ×29; inline scream → Surprised Gasp ×75, Contented Sigh ×29; solo exhausted_groan → Low Mumble ×148; solo shriek → Contented Sigh ×47. no_burst appears only in single digits. Genuineness at w = 0 tracks the split: chuckle 4.68/6 and exhausted_groan 4.88/6 against frustrated_groan 1.49/6 and shriek 1.79/6. The protocol calls this an adapter-quality or label-quality problem, not a merge-weight problem, noting that frustrated_groan's labels came from a sidecar rescan that never entered the corpus burst_labels column. [P§43.5]

Recommendation: keep BURST_LAM = 0.25 and BURST_LAM_INTENSE = 0.5. Paired directly against what ships: in stack/solo no weight is significantly better than 0.5 on hit-rate while turning the adapter off is significantly worse (−0.061, t −2.32); in stack/inline nothing separates from 0.25, including w = 0. Raising the dose has one-directional costs (stack/solo d blend falls monotonically to −0.427 and d |dur err| rises to +0.017 s by w = 1.5). It does not differ by class: the per-class best w column scatters over the whole range with 8 stars in 108 cells against about 5 expected by chance, two of them negative. inline is flat on everything (max |t| 0.79, high confidence); solo is the only place anything moves (moderate confidence). The solo hit-rate maximum sits at w = 1.5, the top of the swept range, and 'is not an optimum' — the design cannot distinguish a peak from the last point before one, the same failure mode as §45. Priority follow-ups: a forced-choice listening test on stack/solo at 0.5 / 1.0 / 1.5; re-scoring with hit-rate primary, labelled post hoc; auditing frustrated_groan's and shriek's adapters against their training data; adding esthetics_high. [P§43.6]

Limits (four of eleven). No human heard any of the 7,260 takes. The stack measured is five adapters while production runs six (esthetics_high @ 0.5 absent; §33 places the intelligibility collapse at the addition of a further adapter). 'Standalone' cannot be separated from 'short script'. And 'the adapter is weak' cannot be distinguished from 'the detector cannot hear it' — every outcome passes through one burst detector never validated against human labels. The one positive claim: in a five-adapter production-shaped stack, the burst merge weight is not the lever that fixes the field report's complaint. [P§43.7]

Operational. Both jobs ran the same Python 3.11 env_transcribe venv, so the grid is single-version; a fallback launcher bdrun4.sbatch (Python 3.13.5 / torch 2.9.1 / transformers 5.14.1) was prepared and never submitted. import transformers off /e/data1 cost job 1546767 3 h 53 m and timed it out; the same import cost 38 min the next morning and about 13 s under source /e/scratch/reformo/schuhmann1_moss/code/fastgen.sh; full aggregation of 2,549 rows takes 0.7 s that way against 244–448 s for import numpy from the venv. MOSS_IMPORT_STAGGER does nothing against this: job 1550920's four ranks staggered 0/20/40/60 s all finished import numpy at exactly 604.4 s. The HF caches must stay split (HF_HOME on scratch, HF_HUB_CACHE on /e/data1). Cancel-and-resubmit churn (bdose3/4/5) produced nothing; 2,291 cells took 2 h 57 m wall at about 74 cells/min on 16 ranks. [P§43.9]

II.9.2 31 Aug 2026 — Burst realisation and stop/no-improvisation: one DPO corpus, and a per-pair scoring harness that answers 'which checkpoint' [P§44]

Stage 3 of the rank-16 chain: ratelora_sft_r16/step3788 → ratelora_dpo_mix_r16/step245 → this run. Full report ~/reports/burst_stop_dpo.md; all numbers measured on JUPITER, 2026-08-31. Corpus $SC/a4_datasets/ds_mix_dpo_bs: 14,932 pairs in 16 shard-*.parquet — quality 4,478 (30.00 %, reused ds1_quality_dpo_final), speed 2,985 (19.99 %, reused ds3_speed_dpo_final), burst 3,734 (25.01 %) and stop 3,732 (24.99 %), both new. Inside the burst arm: burst_swap 1,476 / burst_silence 1,310 / burst_excise 948 = 39.5 / 35.1 / 25.4 % (planned .40/.35/.25); stop kinds equalised at 1,244 each. Two design invariants: the swap negative is a burst of a different class by the same speaker, re-timed and RMS-matched (silence-only negatives would train a presence detector rather than class control; donors are 100.0 % cross-semantic-group over 170 target→donor class pairs and all 500 voice-profile speakers); and stop_twosent states the full original duration while the script holds only sentence one, verified on 410/410 timed renders (one speech segment each, segment sum equals full length, residual max 0.22 s). The burst arm is voice-profile-only by necessity — on the real-speech half voice_key is per-utterance (277,280 distinct 'voices' for 819,313 low-mumble detections), so no same-speaker donor exists. Gates applied before training: burst_no_free_slot 1,345, twosent_bad_split 673, truncate_bad_span 431, caption_leak 188, not_two_segments 144, first_seg_not_sentence 132; plus 18 of 44 detected burst labels dropped for under 40 candidate clips. caption_leak matters: on 188 stop rows caption_general contained a word from the cut part of the script. [P§44.1]

The run. Job 1557904, 4 GPU, [done] steps=902, EXIT=0, 11:43 → 12:26 (42 min of a 6 h wall). From ratelora_dpo_mix_r16/step245, base $SC/out/sft3/export, rank 16 / alpha 32, β = 30.0, length-normalised, lr 1e-6, warmup 90, 2 epochs, 32 global pairs/step, --val-frac 0.02, prompt_lib2_rate (hash 090c4ca315519a57); 17 checkpoints kept, step56 → step902. Run on fastgen ($SC/code/fastgen.sh: py 3.13.5, torch 2.9.1+cu128, transformers 5.14.1, peft 0.20.0): import gate 24.4 s and in srun 26 s after job start, against 604 s for import numpy alone on the venv. rew_chosen is flat for the entire run (3.29 to 3.49) while rew_rejected falls monotonically from +0.295 to −3.87: every bit of margin comes from pushing the rejected side down. Read with §42 (rc is measured against bare SFT3 via peft.disable_adapter()), the flatness says no checkpoint moved the chosen-side likelihood away from where stage 2 left it. [P§44.3]

famscore.py. Job 1562661, 1 GPU, 14 min, 40 s per checkpoint: the same 317 held-out pairs through the init adapter and all 17 checkpoints — 18 conditions, 5,706 records in $SC/work_bs/famscore/per_pair.jsonl, plus summary.json and paired.json. Identical pairs across conditions make the test paired, which is what makes n = 22–32 per family usable. Three implementation points: reuse the trainer's own code (load_dpo_units, build_pair_examples, DpoStep imported from dpo_train2); assert the adapter name-mapping hit every tensor (PEFT names …lora_A.default.weight, a saved adapter …lora_A.weight; a silent miss scores the base model 18 times and reads as 'training changed nothing'); pack one row at a time, because build_pair_examples drops undecodable rows and its meta is not index-aligned with its input. Cross-check: famscore puts step56 at 0.7476 on 317 pairs; the in-training eval put it at 0.7743 on its 288-pair subset. [P§44.4]

Result. Held-out n: quality_repair_sidon 95, speed_stretch 55, burst_swap 32, stop_truncate 31, stop_twosent 31, burst_silence 27, burst_excise 24, stop_append 22. (1) stop_twosent was inverted by the stage-2 adapter: the init adapter scores 0.258 (8 right, 23 wrong) — 'the model does not fail to stop, it actively prefers the take that keeps talking' — fixed by step 112 (0.871) and ≥ 0.968 thereafter; the pooled pref_acc could never have shown this (31 of 288 pairs moving 0.258 → 1.000 is worth 0.08 on the pooled number). (2) The stop and burst arms learn on different schedules: all three stop families are finished by step 280 (stop_truncate 1.000 from step 168, stop_append 1.000 from 280); burst_silence is at 0.370 at step 224, below its init of 0.407, first significantly positive at step 336 (d_logit +0.479, t +2.90) and peaks at step 784 (0.889, +1.338, t +7.62); burst_swap sits at init through step 168 and settles at 0.844 only from step 560. 'A checkpoint chosen on the stop arm alone would be chosen 500 steps before the burst arm was done.' (3) Nothing was damaged: quality_repair_sidon dips insignificantly at step56 (−0.164, t −1.51) and is significantly positive from step280 (+1.398, t +9.43 at step896); speed_stretch reaches 1.000 by step504. burst_excise is unsolved: 0.458 → 0.625, best 0.708, paired d_logit significant late (+0.865, t +3.84 at step896); the suspected cause is that it is the only burst family whose two sides differ in length, overlapping the speed_stretch objective. [P§44.5]

Recommendation: the range step 336 → step 896, and nobody has listened. Lower end step 336: first checkpoint where all eight families are at or above init on paired logit and the burst arm has moved (burst_silence +0.479, t +2.90; quality recovered +0.564, t +4.04), stop arm at 1.000 / 1.000 / 0.968. Upper end step 896: largest paired gain on burst_excise (+0.865, t +3.84) and quality_repair_sidon (+1.398, t +9.43), tied for best pooled accuracy (0.9432), no family below init; step902 is the six-step tail and the only late checkpoint with a reversal (burst_swap 0.844 → 0.688, 5 pairs). Explicitly not selected: best-by-val_loss (step 840 — for DPO a lower val_loss is a larger margin and picks the most over-optimised checkpoint) and best-by-pref_acc (step 616 — swings 7 pairs between evals and ranks seventh on the 317-pair ladder). [P§44.6]

II.9.3 The qual1504 weight sweep cannot separate the doses [P§45]

Job 1557968, prompted by the listening result in §41.2. Seven weights on the same ten emotions, same voice, same seeds, in the full demo-server stack; w = 0 is the control through the identical code path. n = 20 per weight (10 emotions × 2 takes), paired within (emotion, take). The adapter's own scaling is alpha/r = 32/16 = 2.0, so w multiplies that. Duration error is 0.00 s at every weight. [P§45]

No weight is distinguishable from the control. The largest |t| anywhere is 1.63. More informative than under-powering is that the response is not monotone: genuineness runs +0.30, +0.07, +0.41, +0.42, +0.01, +0.48, with the collapse at w = 1.25 sitting between two of the three largest effects — 'A real dose-response cannot do that.' Either the effect is smaller than the noise at n = 20 or it is genuinely not smooth. What survives: the three largest genuineness gains cluster at w ∈ {0.75, 1.00, 1.50} (+0.41 / +0.42 / +0.48), and among those w = 1.50 dominates on both axes — largest genuineness gain (+0.48) and smallest word-error cost (+0.006, against +0.016 at w = 1.00 and +0.065 at w = 0.75). [P§45]

Extended to w = 3.0 (job 1563064). w = 1.50 was re-run as the join and reproduces exactly (WER 0.064 / 0.064, genuineness 1.96 / 1.96 across two independent jobs), so the two tables compose. [P§45.1]

Practical reading: w = 1.5. It takes almost all of the available genuineness gain (+0.48 against a best of +0.58) for essentially no intelligibility cost (+0.006); w = 1.0 remains the conservative alternative and the difference is not demonstrable from these data. 'This is a null on a metric, not on the phenomenon' — §41.2's precedent is that a listener overturned a null DNSMOS result on these same adapters the same day. All clips are on the listening page (spaces/laion/moss-quality-adapter-listening, Part 2), seven weights side by side per emotion. (The protocol gives both '220 clips' and '140 clips' for the page; 140 is the first seven-weight sweep, 220 the eleven-weight total.) [P§45.1]

II.9.4 31 Aug 2026 — The listening page shipped with its audio players dead, and nobody would have known [P§46]

Cause. Successive edits left a stale copy of the Part 2 renderer in front of the current one inside the same <script> block. The stale copy read a data shape that no longer existed, threw Cannot read properties of undefined, and because an uncaught exception aborts the entire script block, the correct copy below it never ran. Part 1's 60 players were unaffected; Part 2's 140 silently rendered nothing. Verification had checked status codes, byte counts, JSON validity and the presence of id="app2" — none of which touch whether the JavaScript runs. [P§46]

Fix and tool. nodejs is in the EasyBuild stack. A 40-line simulator — a stub document/localStorage/CSS, then eval of the real data.js, scale.js and each inline block — reports per block whether it threw and how many <audio> elements each container ended up with: before, n1.js THREW and app2 <audio> count: 0; after, n1.js ran OK and app2 <audio> count: 140. It also caught a second latent fault: Part 2 read the saved-ratings object from Part 1's top-level const; Part 2 now reads localStorage itself inside a try/catch that prints the failure into the page. Rule: 'A published page is verified when its scripts have been executed against its real data, not when its files return 200.' Kept at /tmp/sim.js. Every listening page in the project so far — including the 399-clip transition study — had been published on HTTP checks alone. [P§46]

II.9.5 The burst adapters do not fail because of a weight: sixty of them were trained on a label the detector contradicts [P§47]

§43 measured six of the seventy-one burst adapters and left one question open (three classes work and three do not, but the working three have both more rows and a strict embedded_quality_gate, so six cases cannot separate the factors) and one unasked: is the burst even in the training audio? Full write-up ~/reports/burst_data_audit.md; dose results ~/reports/burst_dose_extended.md. [P§47]

The harness on trial, 13/13. frustrated_groan at 0.000 at every one of ten weights is as consistent with a harness that never applies the adapter as with an adapter that does nothing. $SC/code/vb_audit/va_verify.py (job 1574096, one GPU) ran on the identical code path bd_run.py uses: peft is the real 0.20.0, not the two-file stub that shadowed it earlier the same day; the adapter attaches to 268 LoRA modules at base scaling 2.0; set_lora_scale is exact and idempotent (0.5 applied twice gives 1.0, not 0.25 — a cumulative implementation would have made the ladder decay monotonically); per-class fingerprints differ; a deterministic forward moves with the weight (max |Δlogit| = 56.75 from w = 0 to 1.5); w = 0 with the adapter loaded is bitwise identical to the adapter not loaded (Δ = 0.0 on logits, same sha on the code stream), which is why round 2 dropped stack_nolora; replaying bd_run's per-cell arm sequence leaves the intended adapters active; w = 0, w = 1.5 and mismatch give three different code streams. On CPU: the prompt surface form is not the problem — the exact timed cue (<class>, D seconds) appears on 60–65 % of renders for every class (0.637 chuckle, 0.631 frustrated_groan) and prompt_format_hash 073aeb09dc923376 is identical in the manifest and both prompt-set files. 'So section 43's numbers are real.' [P§47.1]

embedded_quality_gate was never a burst-presence gate. $SC/code/vb_buckets.py walks for gate in (0.80, 0.50, 0.00) over has & (blend_pct >= gate) & (gen_pct >= gate) and stops at the first gate yielding 800 rows; the gate varies blend and genuineness percentile, not presence. Over the 3,144,739-row corpus scan only 44 of the detector's 83 labels ever fire — low mumble 846,299 times and frustrated groan 19 times. The eight classes at gate = 0.80 are exactly the eight most frequent detector labels; the gate is a proxy for corpus frequency. Twenty-seven classes were built from scripted cues: vb_buckets.py --cues admits a row when the script names the burst and the detector found some burst at the aligned index — explicitly ~has, meaning the detector called that span something else, and the scripted name is then written over the detector's label. Splitting into det / cue / std: chuckle 1600 = 800 det + 0 cue + 800 std; scream 891 = 800 + 0 + 91; shriek 386 = 174 + 0 + 212; frustrated_groan 568 = 19 + 339 + 210; fast_breathing 1600 = 0 + 1600 + 0. Twelve classes have no embedded rows at all. [P§47.2, P§47.3]

The measurement. $SC/code/vb_audit/va_audit.py, job 1574095, 4 nodes / 16 GPUs, 9 minutes, 43,139 rows — every embedded row of all 70 buckets plus up to 400 standalone per class. Each row's own MOSS codes are decoded with train_grpo.decode_waves and moss_pack.codes_from_blob (imported; the source comments record three GRPO runs lost to a hand-rolled decode), resampled to 16 kHz, and handed to the same reward.RewardModel.bursts instance that scored §43's generations. p_same = fraction of rows where the detector found a burst of the labelled class within 1.5 s of the labelled position. [P§47.4]

Result 2 — cue rows realise the label at 0.00, and the substitution is the finding: fast_breathing 1,600 cue rows → Contented Sigh 1,330; normal_breathing 1,163 → Contented Sigh 962; heavy_breathing 1,162 → Contented Sigh 874; sobs 661 → Contented Sigh 427; frustrated_groan 339 → Contented Sigh 159, Wistful Sigh 103. 'The instrument that scores this entire pipeline hears a sigh where the script wrote a groan, a pant or a sob.' Result 3 — the standalone column is an instrument reading: contented_sigh (0.86 on embedded rows) scores 0.05 on its own curated isolated sighs; the locator is a Whisper-small segmenter and the classifier's embedder a 30-second encoder, so a 1.1 s clip padded to 30 s is out of distribution for both. Standalone p_same is flat across clip length (0.046 at 0.7 s, 0.093 at 9 s); 33.4 % of standalone rows produce no located event against 2.7 % of embedded rows, and when the locator fires the classifier names the wrong class nine times out of ten (3,423 other to 397 same). The standalone rows were admitted on a different classifier's verdict (clf_is_burst in vblora_perclass), leaving about 40,000 rows — 59 % of the burst training set — unadjudicated. [P§47.4]

'An adapter trained 2,000 steps on 568 rows of which eight contain the burst it is named after has been trained to reproduce the other 560, and the other 560 are sighs.' The exception is scream: 556 verified rows and still weak, so bucket size (891 rows / 1,114 steps against chuckle's 1,600 / 2,000) is the remaining candidate. Re-filter: keeping a row only if every labelled span of the target class has a same-class detection within 1.5 s, 15 classes have ≥ 300 verified rows (breathy_giggle 1026, chuckle 832, low_mumble 798, wistful_sigh 757, ahem 727, contented_sigh 727, exhausted_groan 690, yawn 667, resonant_hum 636, surprised_gasp 608, deep_breath 593, childlike_giggle 576, scream 556, sharp_inhale 550, soft_hum 376); 54 classes < 100; 32 classes = 0. A probability floor does not help (at θ = 0.3 every class halves, at 0.5 everything collapses including the working classes); the ranking is θ-invariant. The defensible statement is 'the training label and the evaluation instrument disagree on twenty-seven classes, and the model learned the instrument's answer.' [P§47.5, P§47.6]

The instrument on trial (§47.7). Re-reading the 44,344 scored rows: the detector fires on isolated bursts and is more confident there (median 0.489 on isolated chuckles vs 0.258 embedded; only 3.3 % no detection) but renames them one step inside the family — isolated contented_sigh → Wistful Sigh 55 / Contented Sigh 5; isolated chuckle → Breathy Giggle 48 / Chuckle 30. reward._same_class is strip-non-letters then equality or substring, so 'contented sigh' vs 'wistful sigh' scores 0, the same 0 as silence. At family level over eight families: det 0.666 → 0.766, std 0.068 → 0.221, cue 0.000 → 0.028; ten classes reach ≥ 0.60 (guffaw 1.000, cackle 0.955, chuckle 0.901, humming 0.842 from an exact rate of 0.000) — 4,644 rows certifiable with no new work. Crop experiment (job 1575339, vb_deaf/vd_probe.py): ten contented_sigh rows embedded in speech that the full pipeline named correctly, cut to the burst alone (0.24–0.55 s). [P§47.7]

Isolation costs the pipeline six of ten correct answers and every loss is the locator's; the cheap fix is to bypass the locator for any clip already of burst length. The fix does not rescue the standalone rows: on the curated isolated sighs both stages agree (Wistful Sigh 8/10 with the locator, 8/10 without), so the residue is the source, not the length. For frustrated_groan, on twenty cue rows with the locator bypassed the classifier says Contented Sigh 9, Ahem 3, Wistful Sigh 3, Low Mumble 3, Surprised Gasp 2 — Frustrated Groan zero, at confidences 0.11–0.67; 46 clips are staged at $SC/pages/deaf/. The standalone caveat narrows from a blind instrument to a near-sighted one whose error is directional and localised to a named stage. Rule: 'find the smallest edit to the input that turns one reading into the other', and 'when a metric reports 0, check whether its comparison function can express nearly right.' [P§47.7]

Round 2, the full-coverage dose sweep (§47.8). All 70 adapters on one common carrier — the ten neutral held-out prompts from §28, re-timed per class with that class's cue in the same slot, requested duration set to the class's median labelled burst duration — configuration stack@carrier-v3 (§43's is stack@class-rows-v2; absolute levels are not quotable between them). Arms stack / alone / mismatch, kinds inline / solo, weights 0 / 0.25 / 0.5 / 0.8 / 1.0 / 1.25 / 1.5, 10 prompts × group 3: 15,400 cells / 46,200 takes, 8 nodes × 4 GPUs, job 1574185. The job stopped on TIMEOUT at 4 h: 14,972 / 15,400 cells (97.2 %), 1,501 conditions, 0 canary violations, 0 duplicates; the loss lands in the diagnostic arms (alone inline 1,752/2,100, mismatch inline 636/700). The stack arm is 9,784 / 9,800 — 69 of 70 classes complete; whispered_mumble's inline w = 0.8 cell is not measured and w = 0.5 is 4/10, marked as such rather than omitted. [P§47.8]

Pooled: no global weight is recommendable, and that is not a null. The inline effect is significant at four of six weights (peak Δ r_burst_cls +0.0115, t = 3.59 at w = 0.5), but the count of classes inside the §28 guardrails decays monotonically — 69/70 at w = 0.25, 66 at 0.5, 55 at 0.8, 47 at 1.0, 40 at 1.25, 36 at 1.5 (solo: 65 → 25). BURST_LAM 0.25 is the only inline weight where 69 of 70 classes stay inside the 3σ bands; round 2 gives no reason to move either default. The two target metrics disagree about the dose in opposite directions: r_burst_cls peaks at w = 0.5 inline and is gone by 1.5; hit_rate climbs monotonically to w = 1.25 inline (t = 4.68) and 1.5 solo (t = 4.02), exactly where the guardrails have given out (Δ genuineness −0.19, Δ blend −0.83 at 1.5). The mismatch control moves r_burst_cls by −0.0033 (t = −0.99) inline, +0.0007 (t = +0.20) solo. Per class: 17 of 70 have a weight that separates from zero — 3 in both kinds (childlike_giggle, effort_grunt, deep_breath), 9 inline only (deep_breathing, coughing, sharp_inhale, heavy_breathing, frustrated_groan, wolf_whistle, pleasure_moan, yawn, cough), 5 solo only (scream, soft_hum, hiccup, snorting_giggle, drinking_noises). [P§47.8]

II.9.6 Re-filtering the burst buckets: 9.8 % of the training rows survive, and 34 of 70 classes have nothing to re-train on [P§48]

Companion to §47, run in parallel by a second agent: it builds the filter, the decision table and the re-training, and adds one failure mode §47 did not look for. The defect no detector was needed to find. prompt_lib2.render_prompt writes every burst into the script as (<label>, <D> seconds) with D = burst_end − burst_start, while dur_s also enters the GENERAL caption and the Tokens: budget; the stated burst length is a literal instruction. Comparing every row's dur_s against its own audio (frames / 12.5) over all 67,540 rows of the 70 buckets: [P§48.1]

The filter. A row survives iff Gate A |dur_s − frames/12.5| ≤ 0.15 s, no span ending past the audio, no empty spans (0.15 s because anno._fmt rounds to one decimal and the worst embedded row is off by 0.079 s — the tolerance admits every embedded row by construction); Gate B every target-class span has a same-class detection within 1.5 s (reward._same_class, burst_realisation's own tolerance) with the weakest probability ≥ θ = 0.174; and not is_val. θ is calibrated, not chosen: the 10th percentile of the weakest same-class detection probability over the 1,891 Gate-A-clean rows of the three adapters that demonstrably work (p10 by class 0.199 / 0.162 / 0.171). $SC/code/vb_filter/vf_score.py is a pinned fork of §47's va_audit.py scoring all 67,540 rows; two jobs, about 20 minutes on 12 GPUs. [P§48.2]

6,615 of 67,540 rows survive (9.8 %). The twelve: low_mumble 624, chuckle 590, contented_sigh 582, deep_breath 540, exhausted_groan 530, yawn 504, scream 496, resonant_hum 477, breathy_giggle 456, childlike_giggle 412, wistful_sigh 405, sharp_inhale 345; survivors are 97 % embedded rows. The three field-report classes split three ways: scream is not a data problem (800 embedded rows realise at 68.6 %, indistinguishable from exhausted_groan's 68.2 %; 496 survive at 55.7 %); shriek yields 48 rows and is instrument-limited on top (only 36.8 % of corpus-labelled Shriek spans still read as a shriek after the MOSS codec round trip — they come back as Scream, Surprised Gasp, Childlike Giggle); frustrated_groan yields 4 rows — all 339 scripted-cue rows fail, 155 contain a Contented Sigh at the labelled position, its 191 standalone rows realise at 1.0 % — recommendation: withdraw the adapter, not re-train it. [P§48.3]

The standalone corpus is the single largest source of bad supervision. Realisation on isolated-burst rows: chuckle 30 %, exhausted_groan 13 %, scream 6 %, contented_sigh 5 %, shriek 0 % of 158, frustrated_groan 1 %; 60–87 % of them fail Gate A. A dead end recorded: 'the isolated recordings are known-good positives, so the detector's hit rate on them measures the detector' does not work — it reads 30 % / 5 % / 13 % on three working classes. The check that works is redet (re-detection of prov = det rows through the codec): 65–89 % on every class with a real pool, 36.8 % on shriek, undefined on frustrated_groan. §47's independent audit lands in the same place (chuckle 832 / scream 556 / shriek 65 / frustrated_groan 8 against 590 / 496 / 48 / 4; 15 rescued vs 12 solid + 2 thin). Repairing Gate-A failures by clamping dur_s is not worth it: only 335 rows across all 70 classes fail A and would pass B, none from shriek, scream or frustrated_groan. [P§48.4]

Re-training. train_bucket_loras.py --buckets $SC/buckets_strict --kind all --out $SC/out/bucket_loras_strict with the hf_burst/manifest.json recipe verbatim: rank 16, α 32, 5 epochs, lr 1e-4, batch 4, base sft3 export, prompt_hash 073aeb09dc923376. Holding the recipe fixed shrinks the step count — chuckle 2,000 → 738, scream 1,114 → 620, shriek 485 → 60. Operational: bd_run.py gained --new-loras / --weights / --classes / --arms and arms stack_new / alone_new (NEW::-prefixed adapter strings), proven behaviour-neutral (identical 264 conditions and 16-rank plan with no new flags); two agents worked the area at once, kept clean by pinned forks (vf_score.py, bd_run3.py) and separate output roots; filtered buckets must not be written under /e/data1 (inode quota). [P§48.5, P§48.6]

The instrument has a prior (§48.7). Over the 36,793 scored training rows the detector emitted 43,238 detections using only 44 of its 83 labels, and three labels are 58 % of them — Contented Sigh 28.3 %, Low Mumble 16.1 %, Ahem 13.5 %. Those three plus Surprised Gasp, Breathy Giggle, Wistful Sigh, Childlike Giggle, Exhausted Groan and Chuckle are the detector's working vocabulary and exactly the eight classes at embedded_quality_gate = 0.80, i.e. the eight whose adapters work. For each of the 34 zero-survivor classes one substitute label dominates at 19–88 %; for 16 it is Contented Sigh (fast_breathing 88 %, normal_breathing 86 %, relief_sigh 86 %, panting 86 %, slow_breathing 81 %, gurgling 66 %, sobs 64 %, wolf_whistle 62 %); gulps → Low Mumble 81 %, swallows → Low Mumble 68 %. Four are near-synonym confusions inside the detector's own label set (relief_sigh/Contented Sigh, snorting_giggle/Breathy Giggle, fearful_gasp/Surprised Gasp, whispered_mumble/Low Mumble). [P§48.7]

Before/after (§48.8). Two complete grids: 5 classes × {inline, solo} × {stack, stack_new, alone, alone_new} × w ∈ {0, 0.5, 1.0, 1.5} = 160 conditions, 46 prompts per (kind, arm, weight) — prompt sets are 9/8/10/10/9, so 1472 records is the full grid, not a shortfall of 1600. br swaps in adapters re-trained with the recipe held fixed; bs with the optimiser step count held fixed. Paired at the prompt level, each prompt's 3 samples averaged first, t over n = prompts. [P§48.8]

Two aggregation defects that print as the same character (§48.9). bd_agg.tstat returned nan for n < 2 (one paired observation; now dropped, counted and listed by name — both final runs report 0 of 80 dropped, and on partial data one n = 1 cell had produced a 't' that was pure noise) and for sd = 0 with mean exactly 0 — the run's built-in correctness check at w = 0, where stack and stack_new load the same model; it now prints 0 (ctrl), whereas before 'the run's own proof of validity was rendered as its own failure mode.' sd = 0 with mean ≠ 0 printed +inf and is now an exact sign test, p = 2^−(n−1). tcrit is now exact (scipy.stats.t.ppf(0.975, n−1), scipy 1.18.1) — the previous lookup rounded up to the next tabulated n, giving n = 11 the value 2.201 instead of df = 10's 2.228; the fallback now rounds down. Displayed old→new levels are means over the same prompts the difference is taken on. Twenty-three synthetic checks cover the guard. Rule: 'A sentinel value that your correctness check and your failure mode both render as is not a sentinel.' [P§48.9]

II.9.7 Manufacturing the missing bursts: voice conversion fails the gate, and the corpus that works scores 0.000 [P§49]

Rather than discard rows, build them: splice a real, human-recorded vocal burst from laion/vocal-bursts-clean into the middle of a real SFT-corpus utterance, annotate the seam exactly, and train on the result. frustrated_groan is the test class (§43 measured it at 0.000; neither detector stage hears a groan on any of 20 real cue clips). Three gates: A — does Chatterbox voice conversion preserve the burst's identity while moving it into the surrounding speaker's voice; B — does the splice leave an audible seam; C — does the annotation parse back to the burst that is there. B and C passed at full scale. [P§49]