Training Model

README.md

@@ -118,7 +118,18 @@ CMake builds engrave with two Python workers and share a disposable page cache u
 
 Identical notation bodies share layout and body rasterization even when titles or score identities differ; each score retains its own header, pages, music, and catalog identity. A `cold namespace` means no cache directory exists for the current renderer fingerprint. It does not mean that every score must be rendered independently: repeated bodies can still be reused within that build, including with persistent caching disabled. Progress separates full-score hits, body hits, newly rasterized bodies, and composed pages.
 
-The build-time pipeline is Soundslice/MuseScore -> `.xml` -> musical model and rhythm plans -> part generation -> JSON, engraved PNG pages, generated_tone audio, Pico title artwork, and a read-only catalog database. Sources in `data/songs/` retain the seven-section performance form. Songs generate Acoustic, Electric, and Vocal parts in their authored key, without key variants or a key-selection menu. Performance measure numbers run continuously across the complete arrangement, starting at 1 in the lead-in; they do not restart with a new section or chorus. Pickups occupy the preceding measure. Standalone comparison XMLs use local measure numbers and retain the original arrangement numbers in `manifest.solo`.
+Solos use one pooled model trained on the paired scores in `data/training`. Both solo slots contribute equally; there is no learned acoustic/electric style distinction. **Registered Solo 1 is played exactly by acoustic and Solo 2 exactly by electric.** Unregistered song/slot pairs use deterministic model prediction. Exact registered playback is explicit authored reuse, not a claim of model memorization. Only Autumn Leaves and All of Me are included initially.
+
+Add a source and its authored solo without changing generator code:
+
+```sh
+./modulus.py python -m generator.train --add-source /path/to/song.xml --add-solo /path/to/solo1.xml --slot 1
+./modulus.py python -m generator.train --add-source data/songs/song.xml --add-solo /path/to/solo2.xml --slot 2
+```
+
+To retrain existing pairs, run `./modulus.py python -m generator.train --leave-one-song-out`. Ordinary builds use the saved model; they never retrain automatically. The four-example model is a starting point and does not establish musical quality on unfamiliar songs. [Training and playback details](TECHNICAL_GUIDE.md#authored-repertoire-and-learned-solos) cover pairing, timing, pitch convention, data migration and limitations.
+
+The build-time pipeline is Soundslice/MuseScore -> `.xml` -> musical model and independent parts -> part generation -> JSON, engraved PNG pages, generated_tone audio, Pico title artwork, and a read-only catalog database. Sources in `data/songs/` retain the seven-section performance form. Songs generate Acoustic, Electric, and Vocal parts in their authored key, without key variants or a key-selection menu. Performance measure numbers run continuously across the complete arrangement, starting at 1 in the lead-in; they do not restart with a new section or chorus. Pickups occupy the preceding measure. Standalone comparison XMLs use local measure numbers and retain the original arrangement numbers in `manifest.solo`.
 
 All generated monophonic guitar material uses frets 1–12 without open strings. Accompaniment uses all legal compact positions; a chord can extend to fret 20 while retaining at least one sounding anchor at fret 12 or below and the existing three-fret span restriction. Electric source-melody harmonization retains its existing placement policy. Exact source pitches and ties are preserved; unplayable or unsupported input fails explicitly rather than being silently octave-folded.
 
@@ -126,7 +137,7 @@ There are no genre presets or runtime composition. The complete arranged score i
 
 All parts use two music systems per page, with a single system only when the score ends on an odd system. Lines break only between measures and share consistent staff/TAB widths. A measure that cannot fit at readable spacing fails with a diagnostic instead of being split. Tall notation is scaled uniformly to fit the two rows; guitar notation and TAB stay together and aligned. TAB displays frets and ties without duplicate rests; standard notation retains the complete rhythm. The generated `.pages/*.png` files are the same page images displayed by the product.
 
-Electric melody prioritizes four-string chords, then melody on top. Every generated guitar voicing requires the root, the actual fifth, or both somewhere in the voicing, not necessarily in the bass. Defining chord tones remain mandatory. Acoustic chooses four connected string-voice journeys once across the canonical form, minimizing total fret travel across harmonic spans, then reuses them in every accompaniment iteration. Upper strings stay fixed; bass movement counts both fret distance and one step for switching between strings 5 and 6. All bass inversions are eligible. There is no acoustic register, physical-reach, or shape-effort penalty beyond its hard fret constraints. Complete triads are preferred when feasible; doubling is allowed, authored omissions are never restored, and missing optional tones break equal-movement ties. Electric chord effort balances span and distinct frets against position-center movement, without an absolute fret-height penalty. Unharmonized melody placement retains its existing register and movement costs. This is a fretboard-feasibility approximation, not a finger-assignment model. Exact melody pitches and rhythms remain unchanged.
+Electric source-melody harmonization retains its existing chord and exact-pitch behavior. Accompaniment uses compact source-derived voicings and sparse rhythm figures, reused at corresponding form positions. Neither accompaniment reads a generated solo or the other guitar's realization. Solos use all six strings; fixed scale maps and the former handwritten solo search/catalogs have been removed. New predictions retain chord, color and resolution checks. No dynamics, accents, velocity shaping or random timing are added.
 
 Write all music in reading order: repeat bars, numbered endings, navigation jumps, and pass-specific ties are rejected. The seven-section performance arrangements remain fully written out. Chord labels use `M`, `m`, `aug`, `dim`, `sus`, and `mM7`, including explicit major triads such as `CM`. Recognized Soundslice text-only harmonies are normalized into structured chords; parenthesized chords mark optional turnarounds omitted from the final arrangement ending.