code/rubric_gen/compound_patch.py

#!/usr/bin/env python3
"""Compound causal test: apply ALL fired rubrics' fixes to one program at once.

    python3 compound_patch.py --decisions data/canon_pr_decisions_lib38.jsonl \
        --library data/canon_lib38_frozen.jsonl --out data/canon_compound.json

WHY THIS EXISTS. The single-patch experiment (counterfactual.py) asks whether
ONE rubric's fix moves the score. That is the right attribution standard for
per-rubric precision, but it structurally under-credits two real situations:
fixes that compound (each individually below the noise floor, jointly above
it) and diagnostic fixes whose value is enabling the NEXT change. This runs
the complementary experiment: for each held-out program, one editor pass
applies every fix the judge fired on it, then one re-run and one official
re-grade. Improvement above the floor means the fired set collectively
explains a real part of the gap even where no single member does.

Per-program, not per-rubric: a positive here cannot say WHICH rubric earned
it. Report alongside, never instead of, the single-patch numbers.
"""

from __future__ import annotations

import argparse
import json
import shutil
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent))

import collect as collect_mod
import split as split_mod
from common import append_jsonl, read_jsonl, write_json
from counterfactual import patched_program


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--decisions", default="data/canon_pr_decisions_lib38.jsonl")
    ap.add_argument("--pool", default="data/canon_eval_scored.jsonl")
    ap.add_argument("--library", default="data/canon_lib38_frozen.jsonl")
    ap.add_argument("--edit-model", default="claude-opus-5")
    ap.add_argument("--python", default=".venv/bin/python")
    ap.add_argument("--timeout", type=int, default=1500)
    ap.add_argument("--root", default="/tmp/rubric_gen_compound")
    ap.add_argument("--noise", default="data/noise_floor.json")
    ap.add_argument("--log", default="data/canon_compound_runs.jsonl")
    ap.add_argument("--out", default="data/canon_compound.json")
    ap.add_argument("--data-dir", default="/workspace/edwardosunny/mle-bench/data")
    args = ap.parse_args()

    import os
    python = os.path.abspath(shutil.which(args.python) or args.python)
    from mlebench.registry import registry as _registry
    registry = _registry.set_data_dir(Path(args.data_dir))
    sp = split_mod.get()

    states = {s["state_id"]: s for s in read_jsonl(args.pool)
              if (s.get("grading") or {}).get("graded")}
    for comp in sorted({s["competition"] for s in states.values()}):
        sp.assert_evaluation_only(comp)
    library = {r["id"]: r for r in read_jsonl(args.library)}

    from eval_precision_recall import JUDGE_VERSION
    fires_by_state: dict[str, list[dict]] = {}
    for d in read_jsonl(args.decisions):
        if (d.get("fires") and d.get("judge_version") == JUDGE_VERSION
                and d.get("family") == "conceptual"
                and not d["rubric_id"].startswith("ctrl_")
                and d["state_id"] in states):
            cand = library.get(d["rubric_id"].rsplit("__", 1)[-1])
            if cand and (cand.get("proposed_patch") or "").strip():
                fires_by_state.setdefault(d["state_id"], []).append(cand)

    try:
        floor = json.loads(Path(args.noise).read_text())["global_rel_floor_p95"]
    except Exception:
        floor = 0.0154

    done = {r["state_id"] for r in read_jsonl(args.log)}
    todo = [(sid, cands) for sid, cands in sorted(fires_by_state.items())
            if sid not in done]
    print(f"{len(fires_by_state)} programs with fires; {len(todo)} to run")

    for sid, cands in todo:
        state = states[sid]
        criterion = (f"The following {len(cands)} issues were all identified "
                     f"in this program:\n" +
                     "\n".join(f"{i+1}. {c['criterion']}"
                               for i, c in enumerate(cands)))
        patch = ("Apply ALL of the following fixes together:\n" +
                 "\n".join(f"{i+1}. {c['proposed_patch']}"
                           for i, c in enumerate(cands)))
        source = collect_mod.relocalize_submission(state["code"], state)
        new_code = patched_program(source, criterion, patch,
                                   model=args.edit_model, code_cap=16000)
        if new_code:
            new_code = collect_mod.relocalize_submission(new_code, state)
        row = {"state_id": sid, "competition": state["competition"],
               "n_fixes": len(cands),
               "rubric_ids": [c["id"] for c in cands]}
        if not new_code or new_code.strip() == source.strip():
            row.update({"outcome": "edit_failed"})
            append_jsonl(args.log, row); print(sid, "edit_failed"); continue
        workdir = Path(args.root) / state["competition"] / sid[:12]
        shutil.rmtree(workdir, ignore_errors=True)
        ex = collect_mod.execute(new_code, workdir, python=python,
                                 timeout=args.timeout)
        row.update({"seconds": ex.get("seconds"),
                    "stderr_tail": (ex.get("stderr") or "")[-1200:]})
        if not ex["clean"]:
            row.update({"outcome": "patched_run_failed",
                        "exception": ex.get("exception_class")})
            append_jsonl(args.log, row); print(sid, "run failed"); continue
        from mlebench.grade import grade_csv
        try:
            rep = grade_csv(workdir / "submission.csv",
                            registry.get_competition(state["competition"]))
        except Exception as exc:
            row.update({"outcome": "grade_failed", "reason": str(exc)[:120]})
            append_jsonl(args.log, row); continue
        before = state["grading"]["score"]
        lb = bool(rep.is_lower_better)
        delta = (before - rep.score) if lb else (rep.score - before)
        rel = delta / max(abs(before), 1e-9)
        row.update({"outcome": "measured", "score_before": before,
                    "score_after": rep.score, "rel_improvement": round(rel, 6),
                    "passes": rel > floor})
        append_jsonl(args.log, row)
        print(f"{sid} {state['competition'][:22]} n={len(cands)} "
              f"{before:.4f}->{rep.score:.4f} rel={rel:+.4f} "
              f"{'PASS' if row['passes'] else ''}", flush=True)

    rows = read_jsonl(args.log)
    measured = [r for r in rows if r.get("outcome") == "measured"]
    passed = [r for r in measured if r["passes"]]
    report = {"programs_with_fires": len(fires_by_state),
              "measured": len(measured), "improved_above_floor": len(passed),
              "noise_floor": floor,
              "per_program": [{k: r.get(k) for k in
                               ("state_id", "competition", "n_fixes",
                                "outcome", "rel_improvement", "passes")}
                              for r in rows],
              "note": "Compound test: all fired fixes applied at once, one "
                      "re-run, official re-grade. Complements the per-rubric "
                      "single-patch numbers; cannot attribute credit to "
                      "individual rubrics."}
    write_json(args.out, report)
    print(json.dumps({k: v for k, v in report.items() if k != "per_program"},
                     indent=1))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())