LAYER_CENSUS_VALIDATION.md

music/exemplars/instruments/LAYER_CENSUS_VALIDATION.md

LAYER CENSUS — instrument (b) validation

Tier: MEASURED INSTRUMENT. Produced by harness/music_gen/instr_layer_census.py (v1).

Machine record: build/audio/exemplars/instruments/LAYER_CENSUS_VALIDATION.json.

Governing ruling: docs/spine/DECISIONS_PENDING_JOSH.md :: MUSIC ROUND 1 GRADED + THE COMPOSITION FLOOR (commit da627060).

Legal posture: Analysis for understanding only. Cards read from disk; lawfully-obtained audio decoded for measurement and never copied, redistributed, or used as model input.

Generated: 2026-08-07T23:00:09+00:00.

1. The one-paragraph answer

The tracks Josh has loved for decades run a measured mean of 5.43 concurrent spectral streams, in a range of 4.38 to 6.35 — the sparsest being Magus Castle and the densest Chrono Trigger Main Theme. That is not dozens, and the reason it is not dozens is the instrument rather than the music: the estimator was calibrated against synthetic mixes with planted answers, it tracks the truth to within about one source up to eight concurrent layers, and then it SATURATES near 8 however many are really there. Every real orchestral track in the pool sits on that ceiling. So the first honest result is a negative one: a mono mixdown cannot tell you whether a cue runs twelve layers or forty, this instrument cannot adjudicate the layer clause of Josh's floor, and the composed manifest in section 6 is the only thing that can. The second result is positive and is the one the sitting needed. Against the eighteen candidates Josh turned down, the loved tracks separate on DENSITY at AUC 0.826 — hits 5.43 mean streams against the candidates' 4.89 — while they do NOT separate on entrance spread at all (0.385; the candidates' median spread 0.821 is if anything wider than the hits' 0.792). The rejected round is not a wall and it is not badly staged: it changes texture MORE often than the music Josh loves (16 entrances against 12) and it is simply THINNER while it does so, and slower to arrive at full texture (0.317 of the way in against 0.165). 13 of 18 clear the corpus-derived floor and 1 of 18 clear the target. Within album — the controlled framing, hits against their own siblings — the strongest axis is entrance_spread at AUC 0.679.

2. Three tiers, and why the cheap one may not impersonate the expensive one

• TIER 1, the band census. The formula card's lanes are ten SPECTRAL OCCUPANCY BANDS — sub, low_bass, bass, low_mid, mid, upper_mid, presence, brilliance, air, ultra. The card says so itself: two instruments in one octave read as one band, one instrument across two octaves reads as two. Band simultaneity can never exceed ten. Every Tier-1 metric carries band in its name, and Tier 1 never emits a layer count.
• TIER 2, the spectral stream estimate. An over-complete KL-NMF dictionary is fitted to the mel magnitude spectrogram of the whole track, and each frame is credited with the number of dictionary atoms holding at least five per cent of that frame's energy. That per-frame count is the layer series. It estimates concurrent spectral streams, not instruments.
• TIER 3, the composed count. When we author a track we know the layer count exactly, because we wrote the parts. Section 6 gives the manifest schema. For our own output the manifest IS the census and Tier 2 becomes only the check that the render matches the score.

3. What the Tier-2 estimator measures, and its error

Calibrated on synthetic ladders whose answers were planted before the estimator saw them. Every figure below was re-derived by the controls in the same process that wrote this file.

The accuracy curve, planted against measured.

• The accurate range is an UPPER BOUND on accuracy: the ladder separates its sources cleanly in register, and real orchestration does not. On real music the same estimator overcounts sparse acoustic writing -- reverb tails, sustain, room -- and undercounts dense tutti, so the two errors squeeze every real track into a narrow band near the ceiling.
• Observed on real music: the 30 corpus hits peak between 7 and 9 streams, with 21 of 30 sitting at eight or above. That is the ceiling reading, not a measurement of the orchestration, and it is why the gate in section 5.4 is built on the mean rather than the peak.

4. The control census

14 of 14 controls passed. 5 of 5 mutation arms were caught.

The mutation arms. Each deliberately breaks one decision inside the module and the run records which control caught it. A mutation nothing catches means a control that is not armed, and the self-test fails on it.

5. The corpus reading

• Pool measured this run: 1180 cards, 30 corpus hits, 927 within-album controls, 223 carried but unused.
• Pool figures are what THIS run measured. PATTERN_FINDINGS_V2.md was written at 1181 cards / 28 hits / 851 controls; the pool has moved since (EX_014's quarantine cleared and further rows attached) and the numbers here are the live ones.

5.1 Tier 1 — the band census, every card

Within-album AUC over 30 hits against 927 siblings. The multiplicity price for this ten-feature bank is a permutation p95 of 0.1313 on |AUC-0.5|; the wider 103-feature bank the program uses elsewhere prices it at 0.195 (PATTERN_FINDINGS_V2.md section 1), and a band axis quoted alongside those features must clear the wider bar, not this one.

• Clears this bank's own permutation p95: n_band_entrances (0.642).

5.2 Tier 2 — how many layers does a track Josh loves actually run

Sampling rule: within each album carrying a hit, sort sibling track ids and draw up to 16 without replacement using random.Random(20260807). Measured: 30 hits, 260 siblings, 18 round-1 candidates, in 587.3 seconds.

Read the layer figures against section 3: the estimator saturates near 8, so peak is a ceiling reading for nearly every row here and carries little information. The MEAN and the staging columns are the informative ones.

• Permutation p95 for this thirteen-feature bank: 0.1719 on |AUC-0.5|.
• Clears it: entrance_spread (0.679).

5.3 The eighteen candidates Josh turned down

• Mean streams: hits median 5.43 against candidates 4.89.
• Entrance spread: hits median 0.792 against candidates 0.821.
• Entrances per cue: hits median 12.0 against candidates 16.0.
• 13 of 18 candidates clear the corpus-derived floor; 1 of 18 clear the target.

The direct contrast, loved tracks against the eighteen. This is NOT within-album and cannot be: the candidates share no composer, no album, no production chain and no era with anything in the corpus, so era and mastering are uncontrolled here. It is a descriptive contrast, and it is the only way to ask whether this instrument can see what Josh heard. AUC above 0.5 means the hits score higher.

• ENTRANCE SPREAD DOES NOT SEPARATE THEM. The axis that separates a loved track from its own album siblings at AUC 0.679 reads 0.385 against the eighteen Josh turned down, with the candidates' median spread 0.821 against the hits' 0.792. In those words: the round Josh rejected stages its texture changes as widely across the cue as the music he has loved for decades. Whatever he heard, it is not a failure to distribute entrances, and this instrument cannot see it.
• The strongest axis on this contrast is est_layers_mean at AUC 0.826, hits 5.43 against candidates 4.89. With eighteen candidates and thirty hits an AUC inside roughly 0.35 to 0.65 is not a finding, and none of these is a controlled comparison.
• The SHAPE of the difference, which is worth more than any single AUC. The rejected round is not a wall. It changes texture MORE often than the loved tracks (16 entrances against 12, 18 plateaus against 12), it spreads those changes just as widely, and it reaches its own full texture LATER (0.317 of the way in against 0.165). What it is, on this instrument's evidence, is CHURN WITHOUT BODY: constant movement at lower density. That is a different defect from the one 'strategically place the sounds and layers' most obviously names, and it is a lead the pass-2 architecture can act on.

5.4 The floor, derived rather than asserted

• Rule: floor = 10th percentile of the hits; target = the hits' median, over 30 hits measured with audio.
• Mean concurrent streams: floor 4.74, target 5.43.
• Entrance spread: floor 0.640, target 0.792.
• passes_floor is built on the MEAN and the spread, never on the peak, because the peak is saturated for nearly every real track and a saturated statistic makes a worthless gate.

6. The composed count — the manifest that makes the floor exact

The audio estimator cannot count dozens and never will from a mixdown. When we author the track we do not need to estimate: we wrote the parts. One row per layer, six required fields, no nesting.

This is not a nicety, and the sibling research lane settles it independently. docs/proposals/music/craft_research/ORCHESTRATION_TEXTURE_AND_DENSITY.md section 3.6 rules that Josh's dozens should be read against the PART count and puts a full cue at twenty-four to forty discrete musical lines, thinning to six or fewer for the sparse nature-and-humming class he also named. This instrument's ceiling is about eight. The two numbers do not overlap, and the consequence is plain: TIER 2 CANNOT ADJUDICATE THE LAYER CLAUSE OF JOSH'S FLOOR. A twenty-four-part cue and a forty-part cue are indistinguishable to it, and so are a twelve-part cue and either of them. The manifest below is therefore the load-bearing instrument for that clause, not a convenience, and the audio tier's real job is the STAGING profile — entrances, exits, plateaus, spread — which it measures well and which the manifest also carries.

• ids unique; entry_s >= 0; exit_s > entry_s; exit_s <= total_length_s
• register drawn from the ten card bands so Tier 1 and Tier 3 speak the same register language
• a layer that leaves and returns is ONE row with the outer entry/exit, plus one row per return only when the return is a genuinely different part
• census_from_manifest() in this module turns a manifest into the same metric names the audio tier emits, so a composed track and a measured track are directly comparable.
• The recommendation to the pass-2 architecture lane: emit the manifest beside the render. It costs nothing, it is the only honest answer to the layer clause of the floor, and it turns the audio estimator into a render check rather than a guess.

7. What this instrument cannot see

• It cannot count dozens. The ceiling is measured, it is about 8, and every real orchestral track in the pool sits on it. The orchestration research lane puts a full cue at twenty-four to forty discrete parts (ORCHESTRATION_TEXTURE_AND_DENSITY.md section 3.6); this instrument cannot tell twelve from forty. Any claim that a track runs thirty layers is a claim about the score, not about anything measurable in the mixdown.
• It cannot name an instrument. There is no timbre classifier here and no roster. A cue of eight synth pads and a cue of eight orchestral desks read the same.
• It cannot tell two players in unison from one. The exact-unison control reads one where the truth is two, and a slight detune reads five. Doubling — which is a large part of what makes an orchestra sound like an orchestra — is invisible in one direction and inflating in the other.
• It cannot tell a layer that is compositionally distinct from a layer that is spectrally distinct. A single instrument sweeping four octaves reads as four.
• It cannot hear whether a layer is GOOD. Density and staging are one clause of Josh's floor; hooks, melody, the three-to-five-cycle variation law and the no-filler bar are other instruments' business.
• It cannot see stereo. Everything is measured in mono at 22.05 kHz, so width, placement and the sense of an ensemble spread across a stage are entirely outside its reach.
• The recommended next rung, not attempted here: a pretrained source separator would raise the ceiling materially. It needs torch, which is not installed and which this lane did not install. It is the obvious upgrade the day the ceiling blocks a decision.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root