# RESULTS — THE HOUSE COHERENCE BENCH (arm 1, `house-wide`) **What this measured.** An operator is training LoRA style adapters on a local text-to-music model from open-licence corpora. The previous train — `mnml0`, 159 tracks of "minimal techno" matched by an archive.org subject tag — produced an adapter the operator's blind ear **rejected**, and the write-up's most quotable causal candidates were **corpus incoherence** and **placeholder captions**. This arm trains on a **bigger, one-genre, one-host corpus with captions derived from each track's own published tags**, and this bench asks, on the pinned instrument and against pre-registered rules, whether that corpus is actually any tighter than the one that failed — and whether the adapter it produced moves its own renders toward it. **The instrument is not rebuilt.** Amendment A1 of `PREREG-COHERENCE.md` pins it, and `PREREG-HOUSE-ARMS.md` §8 reuses it: the same CLAP-laion-music checkpoint (`music_audioset_epoch_15_esc_90.14.pt`, sha `fae3e9c0…`, 2,352,471,003 B), the same CPython 3.12.14 / torch 2.8.0+cpu freeze (`freeze-A1.txt` sha `258a10ee…`), the same `CLAP_Module(enable_fusion=False, amodel='HTSAT-base')` with the checkpoint passed explicitly as `ckpt=`. **The mnml arc's crops were never re-cut** (§8.3) — re-cutting them would re-open the one thing that makes 0.4489 comparable; its archived embeddings were read at their pinned sha. **MEASURED, NEVER GATING** (D-20260831-17). Nothing here passes, fails, promotes or retires an adapter. The instrument of record for whether an adapter is any good is the operator's blind ear; these numbers ride beside that verdict and never in front of it. --- ## §1 — THE GATES THAT RAN BEFORE ANY NUMBER ### 1.1 STEP 0 — the instrument reproduces A1's receipt ``` STEP 0 freeze-A1.txt sha256 = 258a10ee9966ed60eb7839ce9ae9695c431277e4e324564c9fe2bf4e2550ebb1 STEP 0 freeze-A1.txt lines = 141 MATCHES A1 STEP 0 music_audioset_epoch_15_esc_90.14.pt bytes = 2352471003 sha256 = fae3e9c087f2909c28a09dc31c8dfcdacbc42ba44c70e972b58c1bd1caf6dedd STEP 0 checkpoint MATCHES A1 pin STEP 0 torch.cuda.is_available() = False (CPU-only by construction, §3.1) STEP 0 RECEIPT: CLAP-laion-music embedding: shape (1, 512), L2 norm 1.0000 STEP 0 A1 REQUIRES: CLAP-laion-music embedding: shape (1, 512), L2 norm 1.0000 STEP 0 PASS — A1's receipt reproduced exactly; the corpora may now be read. ``` ### 1.2 STEP 0′ — the comparator is reproduced, not remembered `PREREG-HOUSE-ARMS.md` §9.2 registers a reproduction gate: re-derive mnml's row from the archived `embeddings.npz` **in the RNG position it occupied** (fifth of five, `CORPUS_ORDER`), and abort unless it reproduces `0.448852` / `[0.365211, 0.523948]` **to six decimals**. It reproduces — and so does every other corpus in the archive, which is a stronger receipt than the gate asked for: | corpus | n | `D_pair` at N = 19 | 95 % bootstrap CI | full-set `D_pair` | |---|---:|---:|---|---:| | bach | 63 | 0.087162 | [0.067163, 0.107243] | 0.086224 | | sousa | 84 | 0.092111 | [0.077620, 0.108330] | 0.092299 | | chopin | 19 | 0.155487 | [0.131573, 0.178577] | 0.155487 | | **fm-control** (one artist) | 24 | **0.237789** | [0.190913, 0.277263] | 0.235924 | | **mnml** (the tag corpus) | 159 | **0.448852** | **[0.365211, 0.523948]** | 0.446845 | The mnml arc published **0.4489 vs 0.2378**; both reproduce to the digit. ### 1.3 The crops, and what was and was not re-cut | | | |---|---| | sealed `run_bench.py` | `3aed3e5e964f876497e67ccb01b3b157faa2c6dcb8c6017320d964d16ca8a22a` | | house crops planned | **2,241** (1,248 `norm` over all 416 tracks + 993 `raw` over the 331 training tracks) | | cut | **108.6 s, 0 bad** | | embedded | **2,241 forwards in 272.0 s**; crop-vector L2 norms `min 1.0000 max 1.0000 mean 1.0000` | | mnml-arc crops | **NOT re-cut** — archived embeddings read at their pinned sha | The crop plan was emitted by `train/plan_house_crops.py`, which **imports** the sealed bench's `WINDOW_PCTS`, `WINDOW_SECONDS`, `crop_name`, `window_start` and `sha256_file` rather than copying them, so a crop it names is byte-identical to one `cmd_plan` would have named. `crop` and `embed` then ran **from the sealed file, verbatim**. --- ## §2 — P-A: THE OPERATOR'S PRE-TRAIN PREDICTION, ANSWERED **The registered claim** (§9.2): the `house-wide` **training set**'s mean pairwise CLAP cosine distance lands **below** the minimal-techno training set's, on the same instrument at the same N. **Predicted: YES. Confidence: moderate, not high** — registered before any number existed, with the reasoning stated in advance both ways. **The registered decision rule, and it partitions:** * **HOLDS** iff the house CI lies **entirely below** [0.365211, 0.523948]; * **FAILS** iff it lies entirely **above**; * otherwise **NOT DETECTED** — *"the two corpora are not measurably different in spread at this N with this instrument"*, **never an equivalence**. ### 2.1 The three registered sets, primary stream (house draws fresh at RNG position 1) | set | n | `D_pair` at N = 19 | 95 % bootstrap CI | full-set `D_pair` | |---|---:|---:|---|---:| | **`house-wide` TRAINING set — THE HEADLINE** | **331** | **0.377387** | **[0.314739, 0.427888]** | 0.375439 | | the full frozen pool | 416 | 0.379768 | [0.324849, 0.437419] | 0.380345 | | the training set **without** `artist_487628` | 284 | 0.386780 | [0.334029, 0.434874] | 0.385221 | | the held-out side (descriptive) | 85 | 0.390658 | [0.335225, 0.435712] | 0.389588 | | `house-artist` — arm 3's training set | 37 | 0.220339 | [0.171481, 0.278180] | 0.218203 | **On §9.2's third set.** The prereg registered it as *"the training set with `artist_487628` re-included"*, written while ⧖ H-1 was open. D-20260905-95 took option **(b)**, which puts that artist in the training side already — so set (iii) as written *is* set (i). The informative complement is its **removal**, which is what is reported above, and **the substitution is named here rather than silently swapped**. It moves the number the way the prereg predicted a change would: taking out the corpus's single tightest cluster (39 tracks sharing one caption and one album) **raises** `D_pair`, 0.377387 → 0.386780. ### 2.2 THE VERDICT — and the two rules disagree, which the prereg registered as the finding **PRIMARY (the ruled rule): NOT DETECTED.** The house CI `[0.314739, 0.427888]` overlaps mnml's `[0.365211, 0.523948]`; it does not lie entirely below it. **SECONDARY (registered, not a substitute): house is LESS spread, p = 0.0001.** Two-sample permutation at equal N = 19, 10,000 permutations, `default_rng(20260905)`, one-sided for "house < mnml": observed difference **−0.066730**, p(house less spread) = **0.0001**. §9.2 registered exactly what to do here: *"Both reported; **if they disagree, the disagreement is the finding.**"* It is, and the mechanism is visible rather than mysterious: * the **point estimates** are far apart — **0.377387 vs 0.448852**, the house training set **15.9 % tighter** than the corpus whose incoherence was the leading suspect; * the **intervals** still touch because **mnml's is nearly twice as wide** (0.158737 vs the house's 0.113149). §9.2 says why in its own words: mnml's CI *"is the 5th and 196th order statistic of 200 draws — granular, not smooth"*, drawn from 159 tracks; the house CI is drawn from 331. **The sensitivity arm settles that the RNG stream is not what decided it.** Re-drawn in mnml's own fifth-position stream, the house training set reads **0.374548, CI [0.317377, 0.435384]** — **NOT DETECTED** again. The verdict is the same under both streams. **The three outcomes are exhaustive and mutually exclusive by construction. No fourth sentence exists**, and none is written here. In particular this is **not** *"the corpora are the same"*. ### 2.3 The two confounds registered with their directions, re-read against what was measured **The delivery chain.** §9.2 registered that P-A, if it held, *"may be measuring one codec vs four"* — and A-H1 replaced the "one delivery format" half of that sentence with the freeze's measured census. Measured on the frozen manifest, not read from a note: **416/416 mp3, 416/416 at 320 kbps, 416/416 stereo**, sample rate **44,100 Hz on 406 files and 48,000 Hz on 10**, bandwidth ceiling median **19.905 kHz** (min 12.54, max 20.38, mean 19.12), **413 "band edge (expected)" and 3 "natural rolloff (keeps falling)"**. So the house corpus is **one codec at one bitrate at effectively one rate**, and mnml's spanned four codecs and four sample rates. **This confound is real, it is larger than §2.7's amendment allowed for — the census came back more uniform, not less — and it points exactly the way P-A's own reasoning said it would.** A reader is entitled to the sentence *"P-A's secondary result may in part be measuring one codec against four"*, and until S-BAND runs this bench cannot separate the two. The registered sensitivity arm **S-BAND** — recomputing house `D_pair` on crops lowpassed at the mnml corpus's median ceiling — is **NOT RUN by this lane** and is named as owed (see §5). **Duration.** Registered as a bias **against** P-A, i.e. the safe direction: the crop policy samples a fixed 30 s per track, and the house pool's median duration (**290.77 s**) is longer than the mnml-arc corpora's 133–240 s, so house tracks are sampled at a *smaller fraction* of their length — more sampling noise per embedding, which **inflates** house `D_pair`. **The house number is therefore, if anything, biased upward, and it still came in 15.9 % below mnml's.** --- ## §3 — H1-HOUSE: SUPPORTED, ON BOTH RULES **The hypothesis** (§8.2): `D_pair(house-wide training set) > D_pair(house-artist training set)`, N = 19 primary. **Predicted: yes. Confidence: high.** **Decision rule, inherited exactly:** primary — non-overlapping 95 % CIs (`PREREG-COHERENCE.md` RATIFY-5's ruled rule); secondary — one-sided permutation p < 0.05; **both reported, and if they disagree the disagreement is the finding**. | | `D_pair` at N = 19 | 95 % CI | n | |---|---:|---|---:| | `house-wide` training set | **0.377387** | [0.314739, 0.427888] | 331 | | `house-artist` training set (arm 3) | **0.220339** | [0.171481, 0.278180] | 37 | * **Primary — non-overlapping CIs: SUPPORTED.** 0.314739 > 0.278180; the intervals are disjoint, and in the predicted direction. * **Secondary — permutation: SUPPORTED.** Observed difference **+0.160234**, 10,000 permutations, **p = 0.0001**. * **They agree**, so there is no disagreement to publish as a finding. **The nesting is named in advance, not discovered after.** Under D-20260905-95's ruling of ⧖ H-1 as option (b), **arm 3's 37 training tracks are a SUBSET of arm 1's 331**. So this is *"the same 37 tracks with and without 294 others"*, **not** two independent corpora — the A-C3 failure the coherence bench had to file after the fact, registered here before the number existed. Every arm-1-vs-arm-3 sentence carries that phrasing. **What it recovers from the mnml arc, and what it does not.** The mnml arc measured 0.4489 (a tag corpus by dozens of hands) against 0.2378 (one artist) — a 1.9× ratio. Here it is 0.3774 against 0.2203 — **1.71×** — at a different N, a different genre and a different acquisition route. **The wide-vs-one-artist gap is a robust property of how these corpora are assembled, not an artefact of one corpus.** It says nothing about whether either adapter is any good. ### 3.1 Where the house corpus sits among all six Ranked by `D_pair` at N = 19, tightest first — the mnml arc's five, plus this one: | rank | corpus | `D_pair` (N = 19) | what it is | |---:|---|---:|---| | 1 | bach | 0.087162 | one composer, one performer tradition | | 2 | sousa | 0.092111 | one band's catalogue | | 3 | chopin | 0.155487 | one composer, many performers | | 4 | `house-artist` | 0.220339 | one house artist, 37 tracks | | 5 | fm-control | 0.237789 | one artist (the mnml arc's control) | | **6** | **`house-wide`** | **0.377387** | **331 house tracks, 36 artists, one host, one genre fence** | | 7 | mnml | 0.448852 | 159 tracks, an archive.org subject tag, twenty years of netlabels | **The house-wide corpus is the second-most-spread set ever measured on this instrument** — and it is still the tightest *wide* corpus the estate has assembled. Two ratios, because they are different facts and conflating them is easy: * against **mnml**, house-wide is **15.9 % tighter** (0.377387 vs 0.448852); * on the axis from **mnml to one artist**, house-wide closes **33.9 %** of the distance to `fm-control` (0.448852 → 0.237789) and **31.3 %** of the distance to `house-artist` (0.448852 → 0.220339). **So a genre fence and a single host bought about a third of the gap between a folksonomy tag and one artist — not the whole of it, and not a rounding either.** The remaining two thirds is what "36 different artists" costs, and no fence can spend it. --- ## §4 — H3 / H3b / H5, BLOCK T, AND THE PQ FLOOR **The render set:** 85 clips, one per fresh process, 30 s, turbo at **8 effective steps on every clip** (recovered from the engine's own timings, never the argument — turbo silently clamps), shift 1.0, guidance 1.0, seeds `4000+i` and `5000+i`. **108 of 108 per-cell byte-distinct checks PASS** — no arm produced the same audio as another anywhere, so no dial did nothing. The determinism control is **byte-identical**, which is what licenses every byte-distinct check. ### 4.1 H3-HOUSE and H3b-HOUSE — NOT SUPPORTED, and the mnml arc's result replicates **The question:** for each base/adapter pair at the same caption and seed, is the adapter's render **closer to its own training corpus's centroid**? Exact one-sided sign test, n = 24 per arm, registered threshold **k ≥ 17** (α = 0.0320). | | arm | toward | n | p (one-sided) | clears at | verdict | |---|---|---:|---:|---:|---:|---| | **H3-HOUSE** | `house-wide` @ **1.0** | **15** | 24 | 0.1537 | ≥ 17 | **NOT SUPPORTED** | | **H3b-HOUSE** | `house-wide` @ **0.5** | **10** | 24 | 0.8463 | ≥ 17 | **NOT SUPPORTED** | **Predicted: uncertain** — the mnml arc's best-powered arm returned 5/12 and 6/12, and this is the same answer at better power. 15 of 24 is 62.5 %, above a coin but not at threshold; 10 of 24 at half dose is **below** a coin. ⚠ **The clustering caveat rides this test too.** These 24 trials are 12 prompts × 2 seeds and are therefore **prompt-clustered**, so §7.3's clustered type-I table applies and the 0.0320 label is optimistic at any ρ > 0. Neither arm reaches threshold under either reading, so the caveat does not change the verdict — but it is stated rather than omitted because it would have mattered had the count landed at 17. **The cross-centroid control is what makes the 15 interpretable** (`PREREG-COHERENCE.md` §6.4). The same 24 pairs, scored against centroids the adapter was **not** trained on: | arm | toward `house-wide` (its own) | toward `house-artist` | toward `mnml` | toward `fm-control` | |---|---:|---:|---:|---:| | @ 1.0 | **15/24** | 15/24 | 9/24 | 9/24 | | @ 0.5 | **10/24** | 13/24 | 6/24 | 8/24 | **Read it carefully.** At full dose the adapter moved renders toward its own corpus in 15 of 24 and toward the two **techno-family** corpora in only 9 — so the movement is not "toward everything", it has a **house-shaped direction**. But it moved toward `house-artist` exactly as often as toward `house-wide`, which is unsurprising and must be said: **`house-artist` is a 37-track subset of `house-wide`'s own training set**, so those two centroids are not independent. **The honest sentence is: the adapter moved renders toward house and away from techno, at a rate that does not clear the registered threshold.** ### 4.2 H5-HOUSE — NOT SUPPORTED, and it publishes in the registered words **The question:** is `d(·, C_house-wide)` **non-increasing** across `base → 0.5 → 1.0`? | axis | monotone | of | |---|---:|---:| | both seeds | **7** | 24 | | primary seed only (comparable to the mnml arc's 12) | **4** | 12 | **4 of 12 is exactly what the mnml arc measured**, and it publishes in the coherence pre-registration's own registered words: ***"the dial moves the audio but not along the corpus axis."*** The dial does move the audio — 108/108 byte-distinct proves that — it simply does not move it monotonically toward the corpus centroid. ### 4.3 BLOCK T — the trigger confound, separated Twelve renders of the twelve prompts at seed `4000+i`, **base model, no adapter loaded**, with `house-wide, ` prepended exactly as the adapter arm prepends it. §4.6 registered why this is needed: **`house-wide` tokenises to include the genre word "house"**, so the trigger adds a genre term to every prompt — new information on P01–P06 and redundant on P07–P10. | | moved toward `C_house-wide` | n | p (one-sided) | |---|---:|---:|---:| | **the TOKEN** (base → trigger-only) | **3** | 12 | 0.9807 | | **the WEIGHTS** (trigger-only → @1.0) | **7** | 12 | 0.3872 | **The separation the block exists for:** on this instrument the **token alone moved renders toward the house centroid in only 3 of 12** — if anything it moved them *away* — while the **weights moved them toward in 7 of 12**. So the modest house-ward drift in §4.1 is **not** explained by the trigger string adding the word "house" to the prompt; it is the adapter's weights. That is a real result and it is the one thing block T was registered to buy. ⚠ **Block T's other job is not done yet.** Scored **by the ear**, `base` vs `base + trigger` is the cleanest null this harness will ever have — twelve pairs in which no adapter exists, so a systematic preference is an instrument artefact by construction. **That scoring has not happened**, so §7.4's caveat still travels with every block-A result: *this harness has never been shown to return ≈ 50 % with no adapter in the loop.* **A6-HOUSE is filed before the first sitting or the caveat is permanent.** ### 4.4 The memorisation screen (A2-HOUSE) — DESCRIPTIVE, and why **A2-HOUSE was never filed.** §12 owed it *"before the first training step of arm 1"*; the only amendment on the sealed prereg is A-H1, and the first training step was 2026-09-06T09:52:02Z. §0.5 governs: *"If an owed amendment is not filed before the step it governs, the criterion it carries degrades to DESCRIPTIVE for that run — it does not silently become a pass."* **§7.6(d) is therefore descriptive and arm 1 clears on three clauses, not four.** The screen was computed anyway, to A2-HOUSE's own written spec. **The training set's pairwise CLAP similarity, published whole** (54,615 pairs): | min | p25 | median | p75 | **p95 (the threshold)** | p99 | max | mean | |---:|---:|---:|---:|---:|---:|---:|---:| | −0.0212 | 0.5390 | 0.6288 | 0.7173 | **0.8259** | 0.8812 | 0.9956 | 0.6246 | **The matched negative control** — the holdout, which the model never saw, against the training set (28,135 pairs): median **0.6214**, p95 **0.7924**, max **0.9415**. ⚠ §3.1(3) fault (a): the holdout is small-catalogue artists **by construction**, so this control is computed on small-catalogue artists only. **Result:** of block A's 24 adapter-arm clips, **5** have a nearest training track above the 95th-percentile threshold. The closest is `P12__house-1.0__seed5012` against `track_1223463` at cosine **0.8649** — below the training set's own **p99** of 0.8812 and well below its max of 0.9956, i.e. **inside the range the corpus already spans with itself**. **Still owed:** A2-HOUSE's spec has the top-20 pairs **listened to** in the same harness, question *"is this the same tune?"*. That is the operator's ear and has not happened. ### 4.5 §7.5.1 THE PQ FLOOR — BINDING, and it PASSES Audiobox-Aesthetics 0.0.4, A1-pinned, **one instrument, one session, all 85 listening copies**. The floor is **PQ ≥ (base-arm PQ − 0.3)**. **Never rank on the PC axis.** | arm | n | mean PQ | min | max | passes floor (7.6527)? | |---|---:|---:|---:|---:|---| | **base** | 24 | **7.9527** | 6.854 | 8.3865 | — (it *is* the reference) | | `house-0.5` | 24 | **8.0608** | 6.9933 | 8.4372 | **PASS** | | **`house-1.0` (the gate arm)** | 24 | **7.9462** | 7.0888 | 8.354 | **PASS** | | `trigger-only` | 12 | 7.9510 | 7.4531 | 8.2856 | PASS | | `base-repeat` | 1 | 7.7066 | — | — | PASS | **§7.6(c) — the one hard automatic fail — PASSES.** The gate arm is **0.0065 PQ below base**, against a 0.3 allowance: the adapter does not degrade perceptual quality. *(The half-dose arm scores marginally **above** base. That is descriptive, not a ranking claim.)* ### 4.6 The cross-box check (§7.4) — MEASURED, and it changes a phrase Round 4 rendered on the inference box (a 96 GB workstation-class card); these rendered on the training box (RTX 3090, sm_86), under identical pins and the same seed. ``` P01 base seed4001, the archive box 73b95c6b7a32c9d5fb5f8fb62c62ad947fccfba1c20fba0620e62fb547a75a1d P01 base seed4001, the training box 6f29ead37432… DIFFERENT ``` **Fixed-seed diffusion is not bit-identical across these two architectures.** §7.4 registered the consequence in advance and it now fires: **the phrase "byte-comparable across articles" is STRUCK from every downstream document.** Cross-article comparisons in this programme are **same-recipe, not same-bytes**. This is not a failure of anything — it is the reason the check was registered rather than assumed, and the within-box determinism control passed byte-identically in the same session, so the render path is deterministic *on a box*. --- ## §5 — WHAT THIS BENCH CANNOT SAY, AND WHAT IT DID NOT RUN Inherited verbatim from `PREREG-COHERENCE.md` §9: CLAP distance is **not musical quality**; spread is **not disorder**; **no causal claim**; **no generalisation** to genres, models or embedders; **not a memorisation check**; **not an equivalence test** — a null is *"we did not detect a difference at this N with this instrument"*, never *"the corpora are the same"*. **Registered arms this lane did NOT run, named rather than left to be noticed:** 1. **S-BAND** (§9.2) — house `D_pair` recomputed on crops after a common lowpass at the mnml corpus's own median bandwidth ceiling, so a codec-matched reading exists beside the primary. The per-file ceilings **are** measured and published (§2.3); the re-crop is not run. 2. **The duration-matched sensitivity** (§9.2) — house `D_pair` on the subset of house tracks inside the mnml corpus's own duration range. 3. **H1b-HOUSE** (§8.2) — it needs `house-core`, which is arm 2 and does not exist. 4. **The N_min(house) secondary axis** for H1 — reported at N = 19, the **primary** axis, which §8.1 makes primary *"so no hypothesis is decided on an axis chosen after its numbers exist"*. N_min(house) = 37 is arm 3's training set; the N = 19 answer is the registered one. Both (1) and (2) are cheap once the crops exist and are recommended for the arc-end pass.