artifacts/phase3_decisions.json

{
  "recorded_utc": "2026-09-28 12:07 UTC",
  "version": "phase01-v6",
  "holdout_k": 3,
  "amendment": null,
  "decisions": {
    "D3": {
      "choice": "B",
      "rationale": "Fine-tune on s′ with negatives mined from s′ itself - the only option that keeps negative strategy as the independent variable while training on functionality-labelled data (PHASE3_PLAN §3.3). Option (A) is kept as the free secondary reading in §9; (C) is impossible on the CodeXGLUE file."
    },
    "D4": {
      "rule": "max_pair_count_then_lowest_id",
      "rationale": "\"Best-represented\" is measured, not assumed: the census (§5) prints and commits the per-functionality pair counts, takes the k highest and breaks ties by functionality id - before any F1 is computed."
    },
    "D5": {
      "choice": "embeded/",
      "rationale": "PHASE3_PLAN header note: matches the tested runner modules and both call sites (CLOUD.md §4, scripts/run_all.sh); IMPLEMENTATION_PLAN §3 is re-pointed from eval/ + demo/."
    },
    "D6": {
      "choice": "2x2: (C0 decision) x (C_best decision)",
      "rationale": "\"All four outcome modes\" (SCOPE P2-21) is never enumerated in the repo; this is the reading PHASE3_PLAN §5 asks to confirm, and the demo check in §12 requires at least one real held-out pair in the picker for each of the four cells."
    }
  }
}