STEPBACK_POSTMORTEM.md

music/STEPBACK_POSTMORTEM.md

STEPBACK POST-MORTEM — four rounds of our own music stack, audited against the only ear that graded them

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

Canon served: Josh's four recorded music grades (docs/review_candidates.json and

docs/spine/DECISIONS_PENDING_JOSH.md), the melody-first / thirty-year-bar direction (memory

music-direction-melody-first-30-year-bar), the composed-never-prompted rule, the no-human-composer

ruling, the CVD §17.1 care line, and proof-before-spend.

If this document disagrees with canon, CANON WINS and this document is the defect.

TIER: ADVERSARIAL SELF-AUDIT — PROPOSAL-TIER. ENGINE, not CANON. It sets no canon, names no

region content, changes no spine or registry row, and picks no vendor. It is the post-mortem Josh

ordered when he asked whether we are on the right stack at all. **The subject of the audit is our own

work, and there is no defence of it below.**

DERIVED FROM:

• docs/review_candidates.json :: candidates[] — PASS7_FLORES_{ROAD,FALLS,ROUNDS}.grade_verbatim

and .grade_reading (graded 2026-08-09), PASS5_* (graded 2026-08-08), round_1_takedown_note,

round_2_note, round_2_takedown_note, round_3_takedown_note, round_4_note

• docs/spine/DECISIONS_PENDING_JOSH.md — MUSIC ROUND 1 GRADED + THE COMPOSITION FLOOR (L5482 ff.),

ROUND 2 GRADED (L5600) + THE ROUND-2 FLOOR ADDENDUM (L5615), THE MUSIC GENERATOR VERDICT (L3884),

GRADED 2026-08-08 — ROUND 3 DOWN, ALL THREE, AND THE DIAGNOSIS IS ARCHITECTURAL (L5744)

• build/audio/pass7/ROUND4_STAGING.md §0, §1, §2, §3, §4, §5, §6, §7, §7A, §7B, §8, §8A, §8B
• build/audio/pass5/PASS5_VERDICT.md §0-§9 and build/audio/pass4/PASS4_VERDICT.md §0-§7
• docs/proposals/music/MUSIC_PRACTICE_GAP_AUDIT.md §0, §1, §3, §4, §4A
• build/audio/exemplars/instruments/*_VALIDATION.md + HARMONY_FIT.md + MELODY_BANDS.json +

GROOVE_BANDS.json + build/audio/exemplars/CORPUS.json

• harness/music_gen/ read at HEAD — pass3_flores.py, pass5_flores.py, pass7_flores.py,

pass2_realise.py, pass7_realise.py, theme_compositions.py, author_head_cell.py,

timbre_critic.py, pass7_battery.py, floor_instruments.py

• docs/DOC_MAP.md music rows L173-L221, L934, L1038, L1053-L1062

NOT DERIVED (authored judgment, and why it had no canon home):

• The apportionment of each grade clause to composition / rendering / both (§6). Canon records the

clauses; nothing in canon apportions them. Every apportionment below cites the measurement it rests

on, and where the evidence is ambiguous it is called ambiguous rather than assigned.

• The survives / retires ruling in §8. It is an ENGINE call, delegated, and it is named as authored

judgment so the director can veto any line of it.

---

0. The question, and the answer in one paragraph

Josh, ordering this step back: *"you have to be polluting the LLM creating this. Are we using the

proper tech stack?"*

There is no music model in the loop to pollute. Rounds 2, 3 and 4 contain zero music-trained

parameters anywhere between the canon and the WAV. The only model in the generation loop is a

code-generation LLM, and what it emits is Python. The melodies Josh has now turned down three times

in a row are six-, five- and seven-note tuples typed into a source file

(harness/music_gen/pass3_flores.py:127, :135, :142) with a prose comment above each explaining

the theory it is meant to realise. Every round since has developed those same cells by rules written

by the same code model, and graded them with instruments written by the same code model, whose own

validation files state in their own words that they cannot rank real music by quality. So the honest

answer is not that the model is polluted. It is that **the model is being asked to be the composer

through the narrowest channel available to it — literal note tuples and rule systems — with no

musical prior anywhere in the stack and no ear anywhere in the loop except Josh's, four rounds

apart.** The one round that DID contain a music-trained model (round 1, ACE-Step 1.5) was rejected

for a different reason and its lesson was mis-generalised: we concluded that generation had to be

composed rather than prompted, which was right, and then built a composer with no musicality in it,

which does not follow.

---

1. Sources

1.1 Canon — the four grades, verbatim

Read in order. These are the only quality verdicts this program has ever received.

Two earlier verdicts belong in the same list because they are the same ear on the same program and

they are routinely left out of round counts: 2026-08-05, on the shipped orchestral cohort — it

was all the same track, all thirty seconds, one sound, a five-second loop

(harness/music_gen/sameness_baseline.py exists because of it); and 2026-08-06, on the three

loop-proof ACE-Step beds — *"that music is still terrible."* **Six recorded negative verdicts, zero

positive ones, across every architecture this program has tried.**

1.2 Our own tree, read at HEAD for this audit

Every claim in §3 through §7 was verified by opening the file, not by reading a report about it.

harness/music_gen/ is 88,013 lines of Python across 85 modules. The census of what those modules

actually import, taken for this audit:

• The compose path (pass3_flores → pass3_grammar → pass4_grammar → pass5_flores →

pass7_flores → harmony → groove) imports numpy, json, math, random, hashlib,

copy and each other. Nothing else. No model of any kind, no checkpoint, no inference runtime.

• torch appears in four files and in none of them is it a composer. generate.py and

sa3_generate.py are the round-1 ACE-Step drivers; instr_layer_census.py is an analysis

instrument; and **pass2_realise.py — the renderer every round from 2 to 4 runs through — names

torch only inside a set of FORBIDDEN tokens.**

• The renderer carries an AST-verified firewall against generative models, on purpose

(pass2_realise.py:1205-1224). It parses its own imports rather than grepping them — because the

first version of the check matched its own argument list — and refuses sa3_generate, generate,

acestep, stable_audio, diffusers, torch, transformers and audiocraft, with a positive

control proving the parser can say yes.

That firewall is the composed-never-prompted rule made structural, and it is honest engineering. It

also means **the absence of learned musicality in rounds 2-4 is not an oversight — it is an

architectural decision, encoded as a control, and it is the decision this post-mortem exists to

re-examine.** librosa and scipy appear widely, but only in the measurement instruments, which

read finished audio and generate nothing.

1.3 Published and external — cited for the two claims that need outside grounding

These are cited to ground what exists outside our stack (§5, §8) and for nothing else. No number

in this document comes from any of them.

• ACE-Step / ACE-Step 1.5, the music foundation model round 1 used — https://github.com/ace-step/ACE-Step and https://github.com/ace-step/ACE-Step-1.5 and https://ace-step.github.io/
• MIDI-GPT, a controllable symbolic (score-tier) transformer trained on the GigaMIDI dataset, AAAI 2025 — https://github.com/Metacreation-Lab/MIDI-GPT and https://huggingface.co/Metacreation/MIDI-GPT
• Text2midi, text-conditioned symbolic generation, AAAI 2025 — https://github.com/AMAAI-Lab/Text2midi
• SkyTNT midi-model, a MIDI-event transformer — https://github.com/SkyTNT/midi-model
• AIVA, the commercial AI composer Josh named, with MIDI/stem export and a copyright-owning paid tier — https://aivaai.org/ and https://aitoolfinder.org/tools/aiva/
• On why sampled orchestras read as fake — round-robin absence produces the machine-gun artifact, and scripted legato does not reproduce recorded transitions — https://glitchmagic.com/blogs/news/the-two-biggest-secrets-to-making-a-midi-orchestra-sound-realistic and https://www.becomeaprocomposer.com/guide/sample-libraries and https://andrewfeazelle.com/how-to-make-an-orchestral-sample-library-sound-real/

---

2. The trajectory, and the finding that should have stopped the program two rounds ago

Every internal number improved, monotonically, round over round, on identical code. His grade went

down, monotonically, over the same rounds. **They are anti-correlated, and the anti-correlation is

visible in our own published tables.**

And the axis that runs the other way, from the same tables — the one that matches his ear and that we

were scoring as a pass:

Read those two tables together and the mechanism is plain. Josh's round-2 addendum said *build

and hold with proper peaks and valleys that still never lose the listeners*. We encoded the first

half of that sentence as instr_dynamic_arc's hold-share floor and encoded none of the second half.

Four passes then optimised toward longer holds and were rewarded for it: ROUNDS went from no flat

stretch at all to a fifty-three-second one, and ROAD's segmenter boundaries fell from 22 to 8 — one

change every 28 seconds in a 223-second cue. Round 1's floor had already ruled the opposite in his

own words: *"I dont like repetitious loops lasting more than 3-5 cycles without introducing a new

instrument or catch or melody or beat drop."* His round-4 sentence is *"there is no variation or

changes or movement throughout the songs."* **That defect was on our own card, in our own numbers,

getting worse every round, being scored as a pass, because the instrument that read it had been

built from half of his sentence.**

This is the deepest process finding in the post-mortem and it is not about music at all. It is that

a hand-encoded objective function derived from a paraphrase of the customer's sentence will be

optimised against, faithfully, all the way past the thing the customer meant.

---

3. Where the melodies actually come from — the finding that answers the question literally

harness/music_gen/pass3_flores.py:114-144. Three Python literals. This is the entire melodic

source of rounds 3 and 4:

IDEA_ROAD   = G.Cell([(69, 0.0, 0.75), (71, 1.0, 0.5), (74, 2.0, 1.25),
                      (76, 4.0, 0.75), (81, 5.0, 1.0), (76, 6.5, 1.0)], 4, "ROAD")
IDEA_FALLS  = G.Cell([(76, 0.0, 2.0), (71, 2.5, 1.0), (69, 3.5, 0.5),
                      (71, 4.5, 1.5), (66, 6.5, 1.0)], 4, "FALLS")
IDEA_ROUNDS = G.Cell([(62, 0.0, 0.5), (62, 1.0, 0.5), (66, 1.75, 0.75),
                      (64, 3.0, 0.5), (69, 4.5, 0.75), (66, 5.5, 0.5),
                      (64, 6.5, 1.0)], 4, "ROUNDS")

Six notes, five notes, seven notes. pass7_flores.py:177 imports pass3_flores; round 4 is a

transform of the round-3 plan with its structural tones nudged onto the new chord grid by

harmony.snap_structural_tones (pass7_flores.py:1279). **Round 3's verdict said the melodic

material was pass 3's byte-identical and had never been heard by anyone

(PASS5_VERDICT.md §7); round 4 changed its harmonisation, not its identity.** So the tune Josh has

rejected three consecutive times is one authoring act, performed once, in a code editor, by a

code-generation model choosing MIDI integers to satisfy a prose constraint.

The same is true one layer up. harness/music_gen/theme_compositions.py holds the thirteen canonical

theme heads — the leitmotifs that are supposed to survive thirty years — as dicts of pitch tuples,

authored under Josh's own 2026-08-05 ruling (*"Humans aren't composing the music… The entire music

pipeline is yours"*). author_head_cell.py is the BATTERY that checks each composition against the

constraint it claims to realise. **Nothing in either file proposes a tune. They record one and then

verify its compliance.** The promotion note in author_head_cell.py:15-19 is precise about why the

A/B went this way: text conditioning did not reliably realise a specified melodic shape (100%

delivery of the brief's licensed leap on authored material against 12.5% and 0% from sixteen

text-only candidates), so authored composition was promoted as the melodic source. That measurement

was correct. **The conclusion drawn from it — that the melodic source should therefore be typed

literals — was one option of several and was never A/B'd against a symbolic model constrained the

same way.**

Everything downstream is rule systems written by the same model: pass3_grammar and pass4_grammar

(plant, sequence, mutate, displace, augment, reharmonise, liquidate), harmony.py (21 rule codes, a

mode-constrained chord vocabulary, voice-leading by candidate filtering, an ERB spacing floor derived

from Glasberg-Moore), groove.py (metrical hierarchy, Longuet-Higgins & Lee syncopation, a per-class

groove library). This is genuinely good engineering. None of it has ever heard a piece of music.

---

4. What the procedural grammar CAN deliver, and does

This is the honest credit side, and it is real. Nothing below should be thrown away because the

melodies failed.

• Form and structural placement, exactly as specified. Section maps, cycle grids, one manifest

row per layer with the bar it enters and the bar it leaves, novelty schedules, per-cue span

schedules where no two cues share an operation sequence, an operation split or a derivation shape

(PASS4_VERDICT.md §1b). Round 1's architectural defect — one caption in, one mix out, nothing

controlling what entered when — is genuinely solved, and Josh confirmed it with his ear:

round 2 *"sounded much better"* than round 1.

• The care line, mechanically enforced. Allow and deny registers read off the region page's own

Section 10; the antagonist's menace carried by instrument, register, articulation and rhythmic

displacement with each margin measured and declared, after the pitch-set disjunction was correctly

identified as a harness invention rather than a canon requirement (ROUND4_STAGING.md §3). A

voice palette that was reachable and deliberately not used because the region page attests no sung

register (PASS5_VERDICT.md §4). **No generative model on the market will do this, and it is not

optional for this project.**

• Measurability and reproducibility. Pre-render gates published before a sample is played with

timestamps that prove it; must-fire controls that turn a row red when the property is stripped; a

forged waiver value caught by a critic and the row rearmed to read a file rather than a claim

(PASS5_VERDICT.md §5). Loudness to ITU-R BS.1770-4 controlled against ffmpeg ebur128 to 0.02 LU

on a real file. A synthetic room, so no third-party IR licence can ever be withdrawn from under a

shipped game.

• Distinctness at scale. Intra-region and cross-track distinctness run pre-render, 0 of 3 pairs

flagged, 16 of 16 known-answer controls green. Josh's floor is 20+ complete tracks per region node

across 76 region rows — on the order of 1,500 tracks — and the failure mode at that scale is

sameness, which he has already graded once (2026-08-05).

• Honesty about its own nulls. Six instruments in this program return honest nulls on their

separation arm and say so on their own faces. That posture is worth more than any of the numbers

and it is the reason this post-mortem could be written from our own files in an afternoon.

---

5. What it structurally CANNOT deliver, and why no amount of further rules fixes it

**There is no learned musicality anywhere in the generation loop, and its absence is enforced by a

control we wrote** (§1.2, pass2_realise.py:1205-1224). Not "not enough of it" — none, deliberately.

A model trained on music has a prior over what melodic continuations sound like to people who have

heard music. Our stack has, in that slot, a code model choosing integers to satisfy predicates. The

consequences are specific and all four are measured in our own files:

1. **The grammar can transform material; it cannot judge whether the material is worth

transforming.** pass3_grammar will plant, sequence, mutate, displace, augment, reharmonise and

liquidate a bad cell exactly as diligently as a good one, and every one of those operations will

score as a pass. Round 3 is the proof: 57 of 57 score-side targets, 3 of 3 on the arc instrument,

and *"these main melodies really suck."* Round 4's dissonance_classifiability at 0.31-0.40 —

roughly two thirds of non-chord tones unclassifiable as passing, neighbour, appoggiatura,

suspension or anticipation (ROUND4_STAGING.md §8 item 3) — is the same fact in vertical form:

the notes are being placed by rules that do not know what the note after them will sound like.

2. The vertical relationship is at chance. instr_harmony_fit's random-melody null re-draws the

melody's pitches uniformly from the plan's own mode inside each part's compass, keeping the

rhythm, 200 seeded draws. **Round 2: 0 of 3 cues clear their own null. Round 3: 0 of 3. Round 4:

1 of 3.** In plain words, for five of six cues across two rounds, the melody's agreement with its

own chords is statistically indistinguishable from random pitches drawn from the same scale. That

is what a pitch sequence produced without a musical prior looks like when you finally build a test

that needs no corpus.

3. Our judges cannot tell good music from bad music, and they say so. Every one of them was

priced against a within-album permutation multiplicity bar — can the axis tell a track Josh has

loved for decades from competent filler on the same album — and **not one axis in any instrument

clears it.** instr_dynamic_arc: best deviation 0.1527 against a p95 bar of 0.1750, published

with the sentence *"this instrument cannot rank real music by quality."* melody_bands: every

axis DEFECT-DETECTOR-ONLY or NULL. instr_groove: 0 separates, 2 defect-detector-only, 5 null.

instr_melodic_intelligence: zero of five axes reach AUC 0.75. Worse, several point the wrong

way — instr_hook_presence scores the round Josh threw out HIGHER than the music he loves

(hook_strength AUC 0.306), instr_melodic_intelligence's counter-melody axis reads 0.195, and

instr_groove finds three audio axes where the loved tracks score LOWER because a pulse near

perfect regularity is a grid, not a groove. instr_noise_structure measured sensory roughness —

the one thing that could grade clashes on a render — at AUC 0.4963, chance, and withdrew its one

apparent separator when it turned out to be reading codec provenance. **The whole judging bank

separates OUR OUTPUT from real released music and cannot rank real music. Optimising against it

moves the work toward "not obviously machine-made" and never toward "good."**

4. No ear is in the loop before Josh's. Four rounds shipped to the only ear in the program

without a single internal listen — every verdict document says so in its own opening paragraph.

The image factory has critics that LOOK; music has none that LISTEN. And the one instrument in the

repo built precisely to catch a fake-sounding render, timbre_critic.py (blind to the score,

anchored on 1,182 measured cards of real released game music), **is not imported by

pass7_battery.py, pass5_battery.py, floor_instruments.py or pass7_flores.py.** The single

thing we own that could have failed round 4 for how it SOUNDS was not run on round 4.

---

6. The four grades traced to composition, rendering, or both

Apportioned against measurements, one clause at a time. Where the evidence does not decide, it says

so.

Round 1 — GENERATOR CONTROL SURFACE (neither composition nor rendering as we now use the terms)

ACE-Step 1.5 produced a finished mixdown from a prose caption. Nothing controlled what entered when.

The loop's own measurement was that roughly half the card dimensions do not respond to the

generator's conditioning — the control surface was the ceiling, not the targets

(DECISIONS_PENDING_JOSH.md, THE MUSIC GENERATOR VERDICT, 2026-08-06). The right lesson was

*generation must be composed, not prompted*. The lesson we also drew — *therefore no learned model

may participate anywhere* — does not follow from that measurement and was never tested.

Round 2 — COMPOSITION

*"They sounded much better, but the melodies are way too simple."* The architecture fix worked and he

said so. The complaint is entirely about the material: one head cell realised three ways, longest

novel span 3.5-4.0 bars, one to three distinct ideas per cue. Rendering was not named.

Round 3 — one clause RENDERING-AND-CODE, three COMPOSITION, one BOTH

Round 4 — one clause RENDERING outright, one COMPOSITION, two BOTH

**Rendering has been a silent tax since round 2 and only became visible when the composition improved

enough to expose it.** That is the honest reading of why "I hate the instruments" arrives at round 4

and not at round 2.

---

7. What the batteries actually proved, stated without flattery

• They proved the writing changed in the ways each prior verdict called for. That is a real claim and

it held every round.

• They proved our output is distinguishable from real released game music: instr_dynamic_arc at AUC

0.964 over round 2, instr_cycle_law at 0.911 over all 32 round-1 candidates, layer_census

density at 0.826. **These are generation-defect detectors and every one of them labels itself as

one.**

• They never proved anything about quality, and never could, because no axis in the bank clears its

own multiplicity bar. Two rounds were steered by a bank with that property while its own validation

files said so on the first page.

• The mismatch was formally ruled after round 3 — *the structural battery stays as floor, never again

as verdict* — and the ruling did not take. Round 4's headline achievements are still

score-tier: 57/57, 3 of 3 on the piano test, IN band on two melodic axes. Its own staging document

spends more space on hold_share 0.1333 against a floor of 0.1434 (about 2.3 seconds in a

223-second cue) than on anything Josh could hear. **That is the demotion failing in practice, and

it is a process defect, not a music one.**

---

8. WHAT IT MEANS FOR OUR STACK

8.1 Survives into any new architecture — do not rebuild these

8.2 Retires

8.3 What this post-mortem requires of ANY round-5 architecture — before a vendor or model is chosen

Stated as requirements rather than as a recommendation, because the comparator lane owns the choice.

1. A musical prior must exist somewhere between the canon and the notes. Whether that is a

symbolic model conditioned on our head cells, an audio model used inside our structural frame, or

a licensed composition service, the melodic proposal must come from something that has heard

music. Rules may then filter, develop and enforce care — that ordering is the whole inversion.

2. A score must survive. Our two discriminating tests, the entire care mechanism, the leitmotif

registry, the orchestra-playability bar and the 3-to-5-cycle variation law all read notes. **An

architecture that emits only audio retires all of them**, which is the round-1 trap. Symbolic

output, or audio plus a reliable score, is a hard requirement.

3. An ear must enter the loop before Josh's. Wire timbre_critic.py into the battery, and add a

listening step of some kind — an internal audio critic, or a graded blind A/B against corpus

excerpts — so that "sounds fake" can fail a round without costing him a sitting.

4. Decompose the two complaints before spending anything. The cheapest experiment available:

render a known-strong public-domain melody through the identical sfizz chain, and render our own

ROAD cell through a better realisation path, and put both in front of Josh **as a sound question,

not a music question.** For four rounds we have never separated "the writing is bad" from "the

players are bad." That is proof-before-spend, it costs no model licence, and it tells the

comparator lane which half of the stack to replace first.

5. Throughput has to be planned for, not discovered. Josh's floor is 20+ complete tracks per

region or cultural node across 76 region rows. Four rounds have produced eight cues. Whatever is

chosen must be arguable at three orders of magnitude more output, with distinctness holding.

6. The care line and the licence posture are non-negotiable constraints on the choice. Allow and

deny registers must be enforceable, and nothing in our corpus may become a conditioning,

fine-tune or LoRA input — a rule the fetch modules already carry in their own docstrings and which

must travel to any new generator.

AND THE ONE THING THIS DOCUMENT MAY NOT DECIDE, NAMED RATHER THAN ASSUMED. Requirement 1 asks

for a musical prior in the loop. composed-never-prompted is Josh's standing ruling and the

generator firewall in pass2_realise.py is that ruling in code. **Whether a learned model PROPOSING

melodic material — which our leitmotif registry then selects, develops, cares for and gates, with

Josh still owning every theme — is composition or is prompting, is his call and not this lane's.**

The two readings are genuinely different: one says the ruling bans prompt-and-pray full-mix

generation, which round 1 was and which failed on its own terms; the other says it bans any model

touching the notes. **The comparator lane must put that fork to him as a full brief — steelmanned

both ways, with a recommendation and the strongest objection — before any round-5 architecture is

chosen, because every option in the comparator's space depends on which reading holds.** Nothing in

this post-mortem should be read as having pre-decided it.

---

9. The strongest objections to this audit's own conclusion

Run per the standing decision protocol, calibrated against this document's own bias, which is toward

"our architecture is the problem."

• OBJECTION 1 — round 1 HAD a music-trained model and lost to round 2, which did not. True, and

it is the strongest thing anyone can say against §5. Josh's round-2 words are *"they sounded much

better"* than the ACE-Step round. The answer, and it is a real answer rather than a dodge: the

two arrangements are not comparable. Round 1 used a learned model as a one-shot full-mix generator

with a prose caption as its only control, and the measured failure was the control surface, not the

musicality. Nobody has ever tested a learned model as a **melodic proposer inside our structural,

care-enforcing architecture**, which is the arrangement §8.3 requires. The objection correctly

refutes "learned models are automatically better." It does not support "typed literals are the

right melodic source."

• OBJECTION 2 — the whole complaint may be rendering. *"I hate the instruments you are using"*

is the most concrete sentence in the round-4 grade, and a Chrono Trigger score played through our

CC0 chain with one dynamic layer per note would also sound like our stack. **This objection is

strong enough that §8.3 item 4 exists to settle it before any spend.** What holds it short of the

whole explanation: rounds 2 and 3 were graded down on melody with rendering never mentioned, and

five of six cues sit at their own random-melody null, which is a property of the written pitches

and not of any sampler.

• OBJECTION 3 — round 4 fixed a four-round pitch defect and has never been judged clean. The

degree-resolution bug put the entire functional harmony two octaves too high for four passes. Round

4 is the first render without it — and Josh still graded it below round 2. That does not rescue the

architecture; if anything it removes the best available excuse for rounds 2 and 3.

• OBJECTION 4 — this audit trusts the same instruments it says cannot judge. Fair, and the

boundary matters. Every number quoted here is used in the one direction the instruments' own

validation licenses: as a defect detector or as a round-over-round comparison on identical code.

No claim here rests on an axis ranking quality, and the one claim that could be read that way —

the random-melody null — is the single axis in the bank that needs no corpus at all.

• OBJECTION 5 — retiring the composition engine discards 88,000 lines and four rounds of work.

§8.1 is deliberately longer than §8.2 because most of it survives. What retires is the **melodic

source and the objective function**, not the form engine, the care mechanism, the registry, the

mix policy or the measurement discipline.

---

10. Honest limits of this post-mortem

• Nobody in this lane has heard any of the four rounds. This audit reads scores, code and

measurements. Every apportionment in §6 is inference from numbers plus Josh's words, and where his

words and our numbers disagree — *"how slow the melody is"* against an IN-BAND density reading —

his words are treated as the ground truth and the number as the suspect. That is the correct

ordering and it is also an assumption.

• It picks no replacement and prices nothing. The comparator lane owns that. §8.3 is a

requirements list derived from our own failures and should be read as constraints on that

lane's choice, not as a pre-empted answer.

• The external URLs in §1.3 were read at search-result depth only. They ground what exists

outside our stack; no number in this document comes from any of them, and any claim about a named

model's real capability needs its own verified read before it decides anything.

• The four-grade sample is small and it is one listener. He is also the only listener who

matters for this product, he has been consistent across six verdicts, and three of his four round

grades located a defect our own instruments later confirmed in our own files. Treating that record

as reliable is a judgement, and it is the judgement this document makes.

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