Codeseys/composer-replication-framework

1# Deep Work Loop Log — Composer 2.5 Replication Framework2 3Started: 2026-05-264Operator: Codeseys (Hermes Agent autonomous loop)5Skill: `deep-work-loop` v1.0.06 7## Vision8 9> Take any HuggingFace model → further RL train it using:10> 1. RLVR (tests-pass reward),11> 2. SDPO/hint-distillation (Composer 2.5's "targeted RL with textual feedback"),12> 3. multi-teacher trace-replay DPO,13> integrated against TRL/VeRL/OpenEnv with DiLoCo-style outer loop sync.14>15> Output: a published, reproducible framework — the "Composer 2.5 replication" the open ecosystem is missing.16 17## Starting state18 19- HEAD: `040eff8` (Wave 6: vision validation self-audit, 5/10 scorecard)20- Tests: 38/38 green in `spikes/005-integrated-trainer-skeleton/`21- Working tree: clean22 23## Phase ledger24 25| Phase | Description | Status | Started | Done |26|---|---|---|---|---|27| 1 | commit-state | ✅ | 2026-05-26 | 2026-05-26 |28| 2 | backlog-audit (BACKLOG.md from VISION_VALIDATION) | ✅ | 2026-05-26 | 2026-05-26 |29| 3 | parallel-research (3 subagents) | 🟡 | 2026-05-26 |  |30| 4 | architect with ADRs (ADR-001..003) | ⏳ |  |  |31| 5 | plan in waves (W7–W10) | ⏳ |  |  |32| 6 | execute W7 — Spike 006 (real HF model smoke) | ⏳ |  |  |33| 7 | execute W8 — Spike 007 (real trace ingestion) | ⏳ |  |  |34| 8 | execute W9 — Spike 008 (DiLoCo smoke) | ⏳ |  |  |35| 9 | execute W10 — packaging | ⏳ |  |  |36| 10 | (Modal-gated) Spike 002a-mini real GPU smoke | ⏳ |  |  |37| 11 | cross-model-final-review | ⏳ |  |  |38| 12 | update scorecard + push | ⏳ |  |  |39 40## Constraints41 42- Verify ALL claims against primary sources (Wave 2 lesson — subagent synthesis is not evidence).43- Tests must pass before commit.44- Memory L1 is at 99% — write to L2 wiki + L3 fact_store, not L1.45- Modal budget: $20 hard cap for this loop. Anything more goes to user for approval.46- No `upload_file` mixing with `git push` — `git push hf master:main` only.47- Commit messages via `-F /tmp/<wave>-commit-msg.txt`.48