| """Capture C3 reasoning for the STILL-CAUGHT remediated rows, then bucket the
|
| tells so we can see what C3 is still catching them on.
|
|
|
| Rebuilds the exact same prompt/orientation used in the scored run (baseline
|
| answer_key), calls the judge once per row for its reasoning, writes
|
| out/C3_remediated_caught_reasons.jsonl, and prints a tell-bucket histogram.
|
|
|
| Run: python -u temp/story_remediation/analyze_caught.py
|
| """
|
| from __future__ import annotations
|
| import json, logging, re, 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[1]
|
| sys.path.insert(0, str(ROOT))
|
| from datasetreview import pipelines as P
|
| from datasetreview import judge_prompts as J
|
| from datasetreview.llm_client import make_judge
|
|
|
| OUT = HERE / "out"
|
| cfg = yaml.safe_load(open(ROOT / "datasetreview" / "config.yaml", encoding="utf-8"))
|
| base = {json.loads(l)["item_id"]: json.loads(l)
|
| for l in open(ROOT / "datasetreview" / "results" / "new" / "C3.jsonl", encoding="utf-8") if l.strip()}
|
| remed = {r["example_id"]: r for r in
|
| (json.loads(l) for l in open(OUT / "trajectories_remediated.jsonl", encoding="utf-8") if l.strip())}
|
| scored = {json.loads(l)["item_id"]: json.loads(l)
|
| for l in open(OUT / "C3_remediated.jsonl", encoding="utf-8") if l.strip()}
|
| caught_ids = [iid for iid, r in scored.items() if r.get("caught")]
|
|
|
| reals = P.real_trajectories()
|
| pairer = P.make_pairer(reals)
|
| judge = make_judge(cfg["model"])
|
|
|
|
|
| BUCKETS = {
|
| "bundled_multitask": r"multi[- ]?task|multiple (distinct|separate) |several (distinct|requests)|checklist|enumerat|packs|laundry list|bundl",
|
| "too_precise_ids": r"exact (order|id|item)|precise|order number|conveniently|front[- ]?load|all the details|specific ids|recites|recite",
|
| "no_backforth_dense": r"no back[- ]?and[- ]?forth|dense|rapid|single (turn|message)|one (turn|message|go)|without (any )?clarif|immediately|right away|opening (message|turn)",
|
| "invented_capability": r"unusual (request|task|service)|would not (typically|normally)|no real|not (a )?(typical|standard|common)|niche|obscure|rare (request|service)|atypical",
|
| "agent_overconfident": r"agent (volunteer|claims|asserts|states)|policy|without (verif|confirm|authenticat)|overconfident|too confident|proactively",
|
| "scripted_synthetic": r"synthetic|constructed|engineered|scripted|test scenario|designed to|artificial|contrived|crafted|feels? (fake|off|unnatural)|too (tidy|neat|clean|smooth|polished)|stilted|robotic",
|
| "verification_flow": r"verif|authenticat|identity|name and zip|otp|one[- ]?time code",
|
| }
|
|
|
|
|
| def which_side_is_fake(key):
|
| return "A" if key == "A" else "B"
|
|
|
|
|
| def one(iid):
|
| fake = remed[iid]; b = base.get(iid)
|
| real = pairer(fake)
|
| swap = (b or {}).get("answer_key") == "B"
|
| msgs = J.build_c3(fake, real, swap=swap)
|
| try:
|
| r = judge.judge({"system": msgs["system"], "user": msgs["user"]})
|
| except Exception as e:
|
| return {"item_id": iid, "error": str(e)}
|
| return {"item_id": iid, "answer_key": msgs["answer_key"], "guess": r.get("guess"),
|
| "confidence": r.get("confidence"), "reasoning": r.get("reasoning", ""),
|
| "fake_side": which_side_is_fake(msgs["answer_key"])}
|
|
|
|
|
| def main():
|
| print(f"capturing reasoning for {len(caught_ids)} still-caught rows...")
|
| with ThreadPoolExecutor(max_workers=8) as ex:
|
| res = list(ex.map(one, caught_ids))
|
| res = [r for r in res if not r.get("error")]
|
| (OUT / "C3_remediated_caught_reasons.jsonl").write_text(
|
| "\n".join(json.dumps(r) for r in res) + "\n", encoding="utf-8")
|
|
|
| bucket = Counter(); multi = Counter(); conf = Counter()
|
| per_row = []
|
| for r in res:
|
| txt = (r.get("reasoning") or "").lower()
|
| conf[r.get("confidence")] += 1
|
| hits = [name for name, pat in BUCKETS.items() if re.search(pat, txt)]
|
| if not hits:
|
| hits = ["other_unbucketed"]
|
| for h in hits:
|
| bucket[h] += 1
|
| multi[len(hits)] += 1
|
| per_row.append((r["item_id"], r.get("confidence"), hits))
|
|
|
| n = len(res)
|
| print(f"\nreasoning captured: {n} (confidence: {dict(conf)})")
|
| print("\n=== WHY STILL CAUGHT: tell buckets (rows whose reasoning cites each; multi-count) ===")
|
| for name, c in bucket.most_common():
|
| print(f" {name:22s} {c:3d} ({c/n:.0%})")
|
| print("\n=== tells per row ===", dict(sorted(multi.items())))
|
|
|
| print("\n=== sample HIGH-confidence catches (verbatim reasoning) ===")
|
| hi = [r for r in res if r.get("confidence") == "high"][:6]
|
| for r in hi:
|
| print(f"\n[{r['item_id']}] fake={r['fake_side']} guess={r['guess']}")
|
| print(" " + (r.get("reasoning") or "")[:400])
|
|
|
|
|
| if __name__ == "__main__":
|
| main()
|
|
|
