tempscripts/story_remediation/unbundle/run_auth_c1c3.py

"""Run C1 (ground-truth validity) and C3 (v2 prompt) on the AUTH-PROBLEM rows,
judging both v5 (before) and v6 (after the deterministic auth fixes) so the only
variable is the authored NL (calls are byte-identical v5==v6).

Auth-problem set = union of
  (a) auth-primary CAUGHT rows  -> caught_reasons_v5.jsonl primary==D_auth_verification
  (b) rows the deterministic lint FIXED in v6  -> NL differs v5 vs v6

Reports, for this set:
  * C3 v2 caught rate  v5 vs v6   (drop = the auth fixes fooled the judge)
  * C1 fail/borderline v5 vs v6   (should not regress; calls unchanged)

Run: python -u temp/story_remediation/unbundle/run_auth_c1c3.py
"""
from __future__ import annotations
import json, logging, sys
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
import yaml

for _n in ("azure", "azure.identity", "azure.core.pipeline.policies.http_logging_policy"):
    logging.getLogger(_n).setLevel(logging.WARNING)

HERE = Path(__file__).resolve().parent
ROOT = HERE.parents[2]
sys.path.insert(0, str(ROOT))
from datasetreview import pipelines as P            # noqa: E402
from datasetreview import judge_prompts as J        # noqa: E402
from datasetreview.llm_client import make_judge      # noqa: E402

OUT = HERE / "out"
cfg = yaml.safe_load(open(ROOT / "datasetreview" / "config.yaml", encoding="utf-8"))
BASELINE = ROOT / "datasetreview" / "results" / "new" / "C3.jsonl"
SAMPLES = 3
WORKERS = max(8, cfg["run"].get("workers", 4))

v5 = {e["example_id"]: e for e in
      (json.loads(l) for l in open(OUT / "trajectories_all_v5.jsonl", encoding="utf-8") if l.strip())}
v6 = {e["example_id"]: e for e in
      (json.loads(l) for l in open(OUT / "trajectories_all_v6.jsonl", encoding="utf-8") if l.strip())}

# (a) auth-primary caught rows
auth_caught = {json.loads(l)["item_id"]
               for l in open(OUT / "caught_reasons_v5.jsonl", encoding="utf-8") if l.strip()
               and json.loads(l).get("primary") == "D_auth_verification"}
# (b) lint-fixed rows (NL differs)
fixed = {eid for eid in v5
         if json.dumps(v5[eid].get("history")) != json.dumps(v6[eid].get("history"))
         or v5[eid].get("query") != v6[eid].get("query")}

AUTH_IDS = sorted((auth_caught | fixed) & set(v5) & set(v6))
print(f"auth-primary caught: {len(auth_caught)} | lint-fixed: {len(fixed)} "
      f"| union present in both: {len(AUTH_IDS)}")

base = {}
for line in open(BASELINE, encoding="utf-8"):
    if line.strip():
        d = json.loads(line); base[d["item_id"]] = d

reals = P.real_trajectories()
pairer = P.make_pairer(reals)
judge = make_judge(cfg["model"])


def c3_pair(fake, real, swap):
    system, user = J.load_prompt("C3v2")
    fr = J.render_trajectory(fake, blind_tools=True, include_metadata=False, conversation_only=True)
    rr = J.render_trajectory(real, blind_tools=True, include_metadata=False, conversation_only=True)
    a, b, key = (rr, fr, "B") if swap else (fr, rr, "A")
    user = user.replace("{{CONV_A}}", a).replace("{{CONV_B}}", b)
    return {"system": system, "user": user, "answer_key": key}


def maj(msgs, field):
    vals = []
    for _ in range(SAMPLES):
        try:
            vals.append(judge.judge({"system": msgs["system"], "user": msgs["user"]}).get(field))
        except Exception:  # noqa: BLE001
            pass
    if not vals:
        return None, 0.0
    m = Counter(vals).most_common(1)[0][0]
    return m, vals.count(m) / len(vals)


def one(eid):
    real = pairer(v5[eid])
    swap = (base.get(eid) or base.get(v5[eid].get("metadata", {}).get("orig_eid")) or {}
            ).get("answer_key") == "B"
    out = {"item_id": eid, "role": v5[eid].get("metadata", {}).get("unbundle_role")}
    for tag, ds in (("v5", v5), ("v6", v6)):
        fake = ds[eid]
        c3 = c3_pair(fake, real, swap); key = c3["answer_key"]
        g, ga = maj(c3, "guess")
        c1 = J.build_c1(fake)
        vd, vda = maj(c1, "verdict")
        out[f"{tag}_c3_guess"] = g
        out[f"{tag}_c3_caught"] = (g == key)
        out[f"{tag}_c1_verdict"] = vd
    return out


def main():
    print(f"judging {len(AUTH_IDS)} auth rows x2 versions x(C1+C3v2) "
          f"(samples={SAMPLES}, workers={WORKERS})")
    with ThreadPoolExecutor(max_workers=WORKERS) as ex:
        rows = list(ex.map(one, AUTH_IDS))
    (OUT / "auth_c1c3.jsonl").write_text("\n".join(json.dumps(r) for r in rows) + "\n", encoding="utf-8")

    n = len(rows)
    v5c = sum(1 for r in rows if r["v5_c3_caught"])
    v6c = sum(1 for r in rows if r["v6_c3_caught"])
    flipped = [r for r in rows if r["v5_c3_caught"] and not r["v6_c3_caught"]]
    regress = [r for r in rows if not r["v5_c3_caught"] and r["v6_c3_caught"]]

    def c1fail(tag):
        return sum(1 for r in rows if r[f"{tag}_c1_verdict"] in ("fail", "borderline"))

    print("\n" + "=" * 74)
    print(f"AUTH-PROBLEM ROWS: {n}")
    print(f"  C3 (v2 prompt) caught  v5: {v5c}/{n} = {v5c/n:.0%}   "
          f"v6: {v6c}/{n} = {v6c/n:.0%}")
    print(f"  -> flipped caught->fooled by the fixes: {len(flipped)}/{max(v5c,1)}"
          f" = {len(flipped)/max(v5c,1):.0%}")
    print(f"  -> regressions (fooled->caught):        {len(regress)}")
    print(f"  C1 fail/borderline     v5: {c1fail('v5')}/{n}   v6: {c1fail('v6')}/{n}  "
          f"(calls unchanged; expect ~equal)")
    print("=" * 74)
    print("\nflipped (auth fix fooled the C3 v2 judge):")
    for r in flipped[:20]:
        print(f"  {r['item_id']:44s} role={r['role']}  v5={r['v5_c3_guess']}->v6={r['v6_c3_guess']}")
    if regress:
        print("\nREGRESSIONS to inspect:")
        for r in regress:
            print(f"  {r['item_id']:44s} role={r['role']}")
    print(f"\nsaved -> {(OUT/'auth_c1c3.jsonl').relative_to(ROOT)}")


if __name__ == "__main__":
    main()