Source project: private groxaxo/llm-polisher.
Outcome: all three preregistered learning rates failed the public retention floor at the first 8-step evaluation and were rolled back exactly. No candidate was selected. Reserved fact-confirmation and fresh-transfer sets remained sealed and are intentionally absent from this archive.
Parent:
- Qwen3.5-4B BF16 + rank-8 LoRA
- regression: 446/446
- adaptation: 139/240
- proxy: 90/144
- fact DEV: 0/24
Observed pre-rollback regression:
- 2e-5: 346/446
- 1e-5: 429/446
- 3e-5: 386/446
Every proposal restored adapter + AdamW + RNG to the immutable parent. A fresh process then reproduced 55 parent outputs exactly with zero cache substitution.
The full explanation is in REPORT.md. evidence/status.json is the compact machine-readable summary.
Exclusions
This archive deliberately does not contain:
- model weights or LoRA adapter binaries;
- optimizer/RNG checkpoint binaries;
- teacher top-k tensor cache .safetensors;
- credentials, API keys or raw private reviewer reasoning;
- the sealed 48-question fact confirmation payload;
- the sealed 144-case fresh-transfer payload.
This is a failed-strategy research record with passing implementation/audit checks, not a promoted checkpoint.