khodo/product-vamp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Cleanup

Browse Source
This commit is contained in:
khodowany
2026-10-05 08:51:03 -05:00
parent ef62af892b
commit 654299decd
25 changed files with 1695 additions and 351 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/composer/implementation_report.md
+95 -46
View File
@@ -1,76 +1,125 @@
# Exact-target composer repair — release status
# Complete-phrase selection and shared-rule learning
**Not released. Exact production reproduction of the four targets has not been achieved.** Valid derivations exist for all four complete event streams, but a target-constrained derivation is not an unrestricted production winner. The existing published assets remain intact. The previous release's report is [archived here](baseline_0e68f53.md); its tests and performance numbers do not apply to this rewrite.
**Exact reproduction is not achieved. The published baseline is unchanged.** The four complete authored solos remain the acceptance criterion. Legal construction, matching prefixes, learned rankings and passing tests do not satisfy it.
Starting revision: `0e68f5364bb0196ab1b7087ecc186f5d7446c42b`. Preserved baseline: `build/solo_reference_comparison/repair/baseline/`. Its `inputs.json` records source/data/reference hashes and initial cache state. The sources remain `data/songs/all-of-me.xml` and `data/songs/autumn_leaves.xml`; no archive source has replaced either.
Starting revision: `ef62af892bce168e2e9846dc6035b414f55d8959`. The preceding report is [archived](baseline_ef62af8.md). Current evidence is consolidated in `build/solo_reference_comparison/structured/`; `baseline/` preserves the previous complete comparison output. Saved output is never a production input.
## Acceptance and target identity
## Active implementation
The exact target files are `build/all-of-me-solo-1.xml`, `build/all-of-me-solo-2.xml`, `build/autumn-leaves-solo-1.xml`, and `build/autumn-leaves-solo-2.xml`. Solo 1 is acoustic; Solo 2 is electric. They are untouched. The optional adapter applies the confirmed guitar-written octave convention and merges ties while preserving repeated attacks. Autumn aligns at the first Gm7 in measure 2, including the entire preceding measure at negative chorus-relative times. All of Me aligns at measure 1.
The production path is `production.generate_role → solo.composer.compose → construction.Builder → placement → validation/export`. The former builder committed the cheapest next action and backtracked only after failure. It now retains histories through common exact-time frontiers and selects among completed event streams. Limits are 16 histories, 24 successors per history and eight completed exports. A route, sounding tail and pending harmonic obligation survive a computational frontier.
The latest approved development scope allows target-guided general-rule work and shared-coefficient fitting, but prohibits stored tunes, per-song recipes and runtime target reads. No fitted model is promoted automatically. Optional development tools and their artifacts remain outside production imports and cache identities.
`Selector.complete` is the authoritative `phrase-selection-2` evaluator. It merges explicit ties, reconstructs silence and reports the contributions actually used for selection. Action labels no longer earn numerical rewards. `scale_thirds` means alternate scale-degree walks; `arpeggio` walks written chord members. A completed opening owns the home identity, separately from the eight-note recent-history buffer.
The baseline exact comparison returned failure: All of Me acoustic/electric had 0/3 exact matching attacks; Autumn acoustic/electric had 0/0. Differing onset positions were 205, 201, 211 and 178 respectively, and rests differed in all four. These are exact-event diagnostics, not musical-quality scores.
A controlled test retains a locally more expensive opening and selects it when it produces the better completed phrase. Tests also cover partition invariance, pending continuation, arpeggios, home identity, deferred validation and exact incremental scoring. These establish selector behavior, not aesthetic quality or global optimality.
## Changes in this working tree
No source song, scale family, rhythm catalog, accompaniment vocabulary, fret domain, runtime schema or performance expression changed. Production data changes remove action-family bonuses and replace the backtrack budget. Missing/no-collection harmony now rejects optional colors before parsing `N.C.`. Fast retrieval uses canonical evaluation for tied fragments and learned contextual rules.
The hash-preselected phrase constructor and separate route solver are replaced by enumerable joint musical actions, one interpretable rational evaluator and bounded deterministic backtracking. Actions include ordinary notes/rests, complete degree routes, realized-theme returns/answers, colors, approaches, enclosures and anticipations. Pending treatments carry their actual future targets and deadlines across actions. Whole-line guitar placement follows selection; it cannot change the tune or inspect another generated part.
## Complete default-build result
General expressive rules now cover delayed exact-target approaches, separated enclosures, reattacked anticipations and anticipated colors. Color octave enumeration no longer loses alternating F-sharp registers through rounding. The base scale vocabulary remains the same 21 collections; exceptions have typed, bounded contracts. Full target endings and source-derived pickup allocations are supported. Electric may overlap the returning vocal pickup as authorized; source vocal events remain unchanged.
The staged build passed `--check` and produced six arrangement scores, complete solos and independent accompaniment for both songs. It is in `structured/staged`, not the published asset directory.
Performance changes remove redundant parent catalog parsing, add content-addressed non-executable catalog JSON, memoize pure chord/color/spelling checks, and avoid building full rational contribution reports for thousands of rejected candidates. The lightweight ranking total and detailed evaluation are equal and tested. Existing independent process workers remain available.
| Output | Generated / target attacks | Matching pitch/onset/release attacks | Full exact result |
|---|---:|---:|---|
| All of Me acoustic | 116 / 137 | 0 | Fail |
| All of Me electric | 333 / 141 | 2 | Fail |
| Autumn Leaves acoustic | 164 / 121 | 1 | Fail |
| Autumn Leaves electric | 311 / 115 | 6 | Fail |
The development fitter uses only shared whitelisted musical features and integer updates. An initial trial exposed two defects: unordered source-interval evidence and per-attack rewards that favored subdivisions. Ordered source relationships and duration-normalized rewards repaired them. A later trace exposed indistinguishable pending targets; their relative delay, step and hold now enter selection explicitly. Interrupted diagnostic trials are not reported as completed fitting or benchmark results.
Rests differ in every case. These are substantial failures, especially the electric parts' continuous activity. Counts are diagnostics, not substitute acceptance thresholds. The exact comparison command returned **1**.
## Evidence and limitations
The [comparison page](../../build/solo_reference_comparison/structured/comparison/index.html) includes baseline, candidate and unchanged reference XML; complete solo-only and fixed-backing WAVs; opening excerpts; final independent backing; and exact differences. Rendering uses consistent neutral treatment and constant gains. **No ear-based audition was performed.**
The staged build completed with `--check`: six performance scores, complete solo XML, notation/TAB, playback assets, and independent accompaniment for both current sources. It is under `build/solo_reference_comparison/repair/staged`, not the published asset directory. The exact gate returned exit code 1:
Reparsed exports agree with all four selected event streams. Every TAB position is legal; ties retain position. Harmonic validation reports no unresolved treatments. All 32 completed-candidate totals equal a fresh canonical evaluation (`asset-verification.json`, `evaluation-recheck.json`). The mismatch is in composition/search/selection, not export.
| Output | Generated / target attacks | Exactly matching attacks | Differing onset positions | Exact result |
|---|---:|---:|---:|---|
| All of Me acoustic | 118 / 137 | 2 | 189 | Fail |
| All of Me electric | 128 / 141 | 0 | 232 | Fail |
| Autumn Leaves acoustic | 106 / 121 | 1 | 177 | Fail |
| Autumn Leaves electric | 143 / 115 | 5 | 195 | Fail |
## Failure diagnosis
Rests differ in every case. These results are not close reproduction, and attack counts are not a substitute for matching the requested music. The full comparison page is `build/solo_reference_comparison/repair/comparison/index.html`. It includes untouched references, the baseline and staged XML, complete solo-only and fixed-backing WAVs, opening excerpts, and exact note differences. Rendering used one neutral treatment and constant shared gains; no ear-based audition was performed.
All four targets are playable on strings 1–6, frets 1–12 without open strings under the confirmed guitar-written convention. All of Me aligns at measure 1; Autumn Leaves at the first Gm7 in measure 2, retaining the preceding measure at negative form-relative time. Sources and references are unchanged.
The actual selected events and TAB were checked against all four reparsed exports: all agree. Thus these discrepancies arise in composition/selection, not from an exporter shifting the selected melody. Source vocal and electric head documents are unchanged from the baseline. Accompaniment documents changed with the shared difficulty configuration and the source-derived participation allocation; they are still independently generated. Source and reference SHA-256 values, and the previously published solo exports, remain unchanged. Evidence: `repair/asset-verification.json`.
Stateful constrained derivation reaches all four complete streams with valid harmonic contracts and zero diagnostic history cuts. This demonstrates coverage only. The ordinary first successor cutoff drops every whole-action-compatible target opening:
The fitting trial was stopped after **40 completed passes of its configured 200-pass maximum**, without convergence. Every pass still had four incorrect rollouts; the best had only eight matching prefix attacks in total across the four cases. This is an interrupted diagnostic trial, not a completed 200-pass result and not a proof that every shared model is infeasible. Its best coefficients remain development-only under `repair/fitting`; none were copied to production data.
| Case | Proposed actions | Exact compatible openings | Retained |
|---|---:|---:|---:|
| All of Me acoustic | 9,853 | 1 | 0 |
| All of Me electric | 9,853 | 2 | 0 |
| Autumn Leaves acoustic | 8,492 | 2 | 0 |
| Autumn Leaves electric | 8,492 | 9 | 0 |
A further normal-production check of that best checkpoint completed All of Me acoustic: **5 matching attacks, 223 differing onset positions, and different rests**; it also failed the exact gate. The other fitted-model jobs were interrupted without completed exports. This is not a four-case fitted-model release test. Its completed XML and comparison are in `repair/fitted` and `repair/fitted-comparison`; the interruption is recorded in `repair/fitted-build-status.json`. The complete four-case results above use the unfitted production defaults.
Evidence: `exact-opening-pruning.json`. An earlier clipped-prefix diagnostic counted a shortened Autumn acoustic hold as surviving; `opening-pruning.json` records that weaker historical observation. The observer now checks complete committed events, with a regression preventing this mistake. It never supplies candidates or changes scores.
The default selection trace explains an important failure. All of Me acoustic selected 29 arpeggio actions and six returns; electric selected 32 scalar actions, nine returns and one ordinary tone. No standalone rest won in either output. Many legal expressive candidates were proposed and rejected by score. The role's action-kind preferences overwhelm much of the vocabulary, and the local policy has no complete-phrase destination comparison. Adding legal actions has not repaired that selection limitation.
Final ranking is also insufficient. Complete targets would lose under default weights even if retained:
### Commands actually validated
| Case | Selected total | Complete target total |
|---|---:|---:|
| All of Me acoustic | −2787/64 | 945/64 |
| All of Me electric | −4551/128 | 1287/128 |
| Autumn Leaves acoustic | −1420/33 | 14/11 |
| Autumn Leaves electric | −420/11 | 239/22 |
All commands use `build/build_common/python/bin/python3 build/build_tools/python_runner.py build/build_common/modulus` as their prefix. Exact invocations are in [commands.md](commands.md).
Lower wins. `evaluation-diagnosis-all.json` records component differences. Source interval-pair resemblance, directed motion, register cost and duration preferences still favor continuous movement. The current features do not establish the longer ordered relationships needed to specify these authored tunes. This is not a proof that every generic model must fail.
- `-m unittest discover -s tests -p 'test_*.py'`: **446 tests passed in 348.905 seconds** (`repair/verified-tests.log`).
- `-m unittest tests.test_solo_comparison`: **9 tests passed in 0.083 seconds** after removing an invalid optional report link (`repair/comparison-tests.log`).
- `-m audition.dependency_check --output .../dependency-final`: passed all six roles, including an operational electric register change that changed that role's document and left the other roles intact.
- Full staged `-m generator ... --check --jobs 2 --composition-jobs 2 --export-solo-xml`: passed. The first attempt exposed an unfillable final quarter-beat; a tempo-gated sixteenth note/rest primitive and a regression test repaired that failure before the successful build.
- `-m audition.solo_compare ... --require-exact`: failed all four outputs, as shown above. This is the release-blocking musical result, not a passing check.
- Full-score selected-event/TAB verification and source/reference/published-baseline hash checks: passed (`repair/asset-verification.json`).
The source-only four-bar diagnostic exports eight completed alternatives and identifies the automatic winner. Seven first-bar realizations differ, but all share the first four attacks. Their complete events, home identities, contracts, actions, frontier counts and scores are in `completed-phrases/candidates.json`. Eight retained alternatives do not by themselves demonstrate useful opening diversity.
The staged asset run took **430.402 seconds** with two composition workers and warm generic catalog/non-solo caches. Tests and development fitting were running concurrently, so that time is not an isolated speed benchmark. Solo role times were 156.117, 228.292, 182.301 and 174.160 seconds in that run. No cold-generation speedup is claimed from these numbers. The full recorded stage timings and actual read traces are in `repair/build-profile.json`.
## Development learning
After other generation/rendering jobs stopped, `-m audition.benchmark --source data/songs/all-of-me.xml --role acoustic_solo --cache .../performance/cache --output .../performance/cold` took **134.722 seconds** for the role, plus 0.046 seconds for source analysis, with one worker and an empty concrete/catalog cache. Repeating it with `--output .../performance/warm` took **2.068 seconds** with a role-cache hit. Hardware/runtime: 12 reported CPUs, Linux 7.2.3 x86-64, repository Python environment; cProfile disabled. Cold, warm and parallel-full-build documents have the same hash `adcc372315aa10cae8a956cc5002700859c81cdcf4eeaa59b4ae6f8b41161baf`. Evidence: `repair/performance/summary.json` and the two benchmark records.
The prefix perceptron is removed. The replacement forms exact rational inequalities between complete ordinary outputs and complete examples. It learns shared bounded integer coefficients per role. Allowed features are relative musical categories and bounded two/three-condition rules, not song names, hashes, absolute measure indices, event arrays or per-song recipes. New conjunctions require demonstrated feature collisions. Production does not import the solver or read examples.
The requested cold-generation reduction is **not achieved**. Warm memoization avoids recomposition, but the wider action enumeration is still expensive. The historical baseline's different music, algorithm and benchmark conditions do not support a controlled speedup ratio. Candidate construction/evaluation remains the dominant recorded stage; no further beam expansion is justified by these results.
The optional dependency is pinned to `z3-solver==4.13.3.0`. Exact zero/one-coefficient interval enumeration handles sparse cases; integer optimization handles the remainder. Feasibility and optimization have deterministic work limits. Incompatibility or exhaustion yields a diagnosis, not a substituted tune.
- [Active path, reproduced findings and search limits](active_path_audit.md)
- [Data migration and provenance](data_migration.md)
- [Authoritative evaluation and fitting contract](evaluation_contract.md)
- [Capabilities and unimplemented behavior](musical_capability_matrix.md)
- [Independent-input checks](dependency_proof.md)
- [Reproduction and exact-release commands](commands.md)
The first four-example round compared each target with eight saved complete alternatives. It found acoustic `duration/1 = 9` and electric `metric/eighth/1_2 = 15`. These sparse coefficients separate the sampled alternatives; **they are not a learned jazz grammar or evidence of reproduction**. They remain in `structured/round-1/data`, with explicit source/target/artifact hashes and training provenance. Nothing was promoted to production `data/`.
Detailed winning/runner-up contributions, selected note groups and proposal/valid/retained/selected counts are in `build/solo_reference_comparison/repair/decisions.md`. Superseded diagnostic logs are consolidated in `repair/diagnostic-history.zip`; current evidence, the baseline and targets remain intact.
Successive rounds can accumulate full-output counterexamples using `--previous-round`. Source/example identities, evaluator, fixed priors and musical policies must agree; only learned weights may differ. Incompatible histories are rejected and duplicate inequalities removed. Later fitting cannot silently forget earlier failed complete outputs. No target-prefix update is used.
The search is a local action policy, not a global complete-phrase optimizer. Dedicated suspension/blues constructors and two-sided phrase revision are not present in this replacement; validators remain. Source influence and the short realized home theme are limited models of melodic identity. A catalog containing thousands of paths does not mean all those paths currently compete as complete musical ideas. No new held-out source-standard evaluation set was added. Existing synthetic transposition/meter/harmony tests do not establish generalization to an unseen standard or subjective musical quality.
A fresh ordinary build of that proposal completed under `learned-stage`, using four workers and no target arguments. It passed technical checks but failed the exact gate in all four cases:
The earliest demonstrated remaining failure is selection: legal target-compatible actions lose even at the opening, and later action choices collapse toward a few favored families. The next justified investigation is whether a shared context model can distinguish those conflicting choices over complete phrases, before adding more rules, expanding the beam, or running another long fit. The finite trial here does not settle that question. There is no automatic fallback to reference notes or previous output.
| Learned output | Generated / target attacks | Exact matching attacks | Exact result |
|---|---:|---:|---|
| All of Me acoustic | 214 / 137 | 0 | Fail |
| All of Me electric | 244 / 141 | 0 | Fail |
| Autumn Leaves acoustic | 315 / 121 | 0 | Fail |
| Autumn Leaves electric | 280 / 115 | 0 | Fail |
Technical completion, exact target equality and listener preference are separate. No ear-based audition is claimed. The release remains blocked unless all four unrestricted exported comparisons pass; a valid or stylistically plausible mismatch is not substituted for that requirement.
The [learned-model comparison](../../build/solo_reference_comparison/structured/learned-comparison/index.html) contains the complete exports, consistent auditions and exact differences. All 32 saved learned-candidate scores and four selected event/TAB streams passed independent rechecking. All eight complete non-solo role documents remained byte-canonically identical to the default stage despite the changed solo weights (`full-song-role-independence.json`). No inference about musical preference follows from that technical result.
The second fit accumulated **64 complete-output inequalities** from both ordinary builds. It found acoustic `space/after_arrival = -37` and electric `duration/1_4 = 19`. Both acoustic targets are still excluded at the first retrieval cutoff. The two electric openings survive that check, which proves neither full retention nor reproduction. This proposal was not staged or published, and no third full rollout is claimed. The result demonstrates how a sparse fit can separate sampled full outputs through coarse rhythmic statistics without recovering the ordered melody. Repeating the same fit or widening the beam is not a demonstrated remedy. Evidence is in `round-2/training.json`.
Learning explicitly depends on the authored examples. Runtime target-free is **not training independence**. These repeatedly studied examples are not held-out data. Transfer to separate standards has not been demonstrated here and is not substituted for exact reproduction.
## Validation
Commands use `build/build_common/python/bin/python3 build/build_tools/python_runner.py build/build_common/modulus`; [commands.md](commands.md) records arguments.
- `-m unittest discover -s tests -p 'test_*.py'`: **458 run: 456 passed, two optional solver skips, 1,139.453 seconds**. A preceding run had one `N.C.` lookahead error; it was repaired before the full rerun.
- With `PYTHONPATH=build/development_packages`, `-m unittest tests.test_joint_actions tests.test_structured_selection`: **35 passed in 24.256 seconds** (`successor-cache-tests.log`). This covers later exact shortcuts, solver changes, accumulated-round compatibility, strict diagnostic matching and incoming-state cache invalidation. The full suite was not rerun after every performance edit.
- `-m audition.dependency_check --output .../dependency-cached`: passed all six roles alone/cold, reversed/warm, two-process/cold, with invalid targets present/absent, and after an operational electric-only register change. All case hashes match the preceding `dependency-final` and `dependency` runs. Actual reads are recorded.
- Full `-m generator ... --check --composition-jobs 2 --export-solo-xml`: passed for the default stage. The separate learned-data run with `--composition-jobs 4` also passed. Actual read traces contain no reference or other part's concrete output.
- `-m audition.solo_compare ... --require-exact`: failed all four in **both** full builds, preventing a success claim.
- `structured/verify_completed.py`, through the Python wrapper: passed full selected-event, TAB, tie and canonical-score checks for both stages (64 completed candidates total).
- `git diff --check`: passed at the latest review.
The dependency exercise uses a synthetic one-bar source and real catalogs, not four complete-song cold/warm repetitions. [dependency_proof.md](dependency_proof.md) gives hashes and scope.
## Performance and cleanup
Hardware reports Intel Core i7-1355U, 12 logical CPUs, Linux 7.2.3 x86-64. The two-worker default stage took **4,832.443 seconds**: 4,817.592 in performance generation, 1.418 playback compilation and 12.655 engraving. Solo role times were 2,126.676 / 2,711.550 seconds for All of Me and 1,958.730 / 2,101.732 for Autumn Leaves. Other roles took about 0.2–0.5 seconds each. Tests/diagnostics ran concurrently; this is not an isolated benchmark.
This is a substantial regression from the prior one-action policy. Wider comparison is not a generation speedup. A 20-second live sample found retrieval, state advancement and frontier scoring dominant; engraving is not the bottleneck.
Pure-operation caching, incremental evaluation, short-circuit theme equality and direct retrieval costs remove repeated work. All eight four-bar candidates and every frontier were identical across 227.751-, 126.421- and 92.633-second runs with different concurrent loads; these do **not** establish a controlled whole-build speedup. A paired 9,853-action microbenchmark measured median 0.274 seconds with full feature reports versus 0.125 with direct costs, every cost equal. That claim is limited to warm-cache retrieval.
Successor-ranking memoization also preserves all eight candidates and every frontier. Its four-bar run took 154.177 seconds while the four learned workers ran, with 90 cache hits and 645 misses. The different load prohibits interpreting that elapsed time as a cache speedup or slowdown. The cache distinguishes catalog identity, the three incoming pitches, release time, sustained/rest context, pending obligations, weights and budget; complete home-aware evaluation still occurs per history. It removes repeated ranking work without changing search breadth.
The learned four-worker full build took **2,007.666 seconds**, including 1,994 seconds of performance generation and approximately 13 seconds of engraving. Solo role times were 1,708.084 / 1,920.536 seconds for All of Me and 1,989.394 / 1,945.428 for Autumn Leaves. Its music, weights, optimizations, worker count and concurrent load differ from the default run. The shorter wall time cannot be attributed to parallelism alone. Cold iteration remains expensive; no arbitrary performance target has been met or promised.
Default-build workers started before the final performance fixes. `frozen-inputs.json` and `post-freeze-repair.json` preserve that distinction. The no-chord guard does not apply to these sources; equality and export checks cover the shortcuts. The learned build has its own `learned-frozen-inputs.json` and predates only the final successor-cache optimization, recorded in `learned-post-freeze-repair.json`. Neither full run is presented as a cold benchmark of every final code byte.
Incomplete task-local stages and superseded caches totaling about 12.8 MB were removed; `cleanup.json` records paths. References, sources, baseline, completed comparisons and useful verified caches remain. Temporary work stays in the consolidated ignored build directory.
## Remaining boundary
The selector repair and learning mechanism are implemented. **The exact musical objective remains unmet.** No learned parameters or staged assets have been promoted. Independent parts, harmonic/physical safeguards and runtime representation remain intact.
The next bottleneck is specific: complete-phrase discrimination and early retrieval disagree, and the current event features permit degenerate rhythmic policies to beat the examples. More beam width, more unconstrained vocabulary, another prefix fit or hidden target retrieval would not establish the requested solution. Any next model change must face accumulated complete ordinary counterexamples and the four full-output target-free gate.
There is also a narrower theme-construction limitation: `related()` transposes or inverts the entire realized opening through strictly mapped pitches. It does not adapt the opening's expressive contracts. A legal chromatic/color-bearing home therefore need not have a legal generated return under this constructor. All four default outputs selected zero returns/answers. This is a construction/representation limitation, not evidence that the scorer has learned to avoid unmusical returns. Full operation counters and winner/runner-up contributions are in `structured/decisions.md`.
See [active_path_audit.md](active_path_audit.md), [data_migration.md](data_migration.md), [evaluation_contract.md](evaluation_contract.md), and [musical_capability_matrix.md](musical_capability_matrix.md).