run/notes/audit_code_report.md

audit-code report (2026-07-12) — digest

• Source-of-truth trainer: src/grpo_train_efficient.py (eff/eff3/mtword = same script, env-var diffs)
• Rewards eff-track: format 0.5 / correctness 2.0 / token_efficiency ramp 40→150 payout 1.5 (correct-gated) / mtword (Round D only) / logger w=0. reward_weights=[1,1,1,1,0]
• Eval: 5 parallel scripts, temp 0.3 top_p 0.9 max_new 512, seed-1 shuffle n=70; tokens = reasoning span only
• AIME eval: aime2026_eval_mtword.py, loads MathArena/aime_2026 from HF, max_new 2048, needs internet
• Round E chain: generate_corpus (400 GSM8K w/ eff3) → mine_supertokens (2-4-gram doc-freq, top-K 48, 238-example corpus in practice) → extend_vocab (add , mean-init embeddings) → sft_seed (r16 SFT, merge) → GRPO polish. Swap points: mining corpus, TOP_K/MIN_FREQ
• No requirements manifest; API scripts need OPENROUTER_API_KEY in ../.env (absent)
• GEPA side uses /usr/share/dict/words; train side uses data/dict_words.txt
• Extension points confirmed: reward shape :141-160, weights :221, SYSTEM_PROMPT :87-100 (must keep tags), max_completion_length :228