agu18dec/chive-climb-test204

CHIVE results_2026-10-14: transcript arm + judge files

Run: the CHIVE held-out eval climb_test204. That's 204 incidents from the public HF dataset agu18dec/chive-forensics (190 WildChat, 11 SystemChat2, 2 Hermes, 1 ToolACE). The target model is Qwen3.6-27B. Each incident is one conversation plus one behaviour the model shows in k of 30 re-samples, with ground-truth causes established by CHIVE's counterfactual experiments.

Run settings:

• Investigator: Claude Opus 5, effort max, on the frozen detective-joracle free-form harness (variants=['textfirst','absence']).
• 5 independent draws per incident.
• Arms differ only in what the investigator can read. Transcript = no activation readouts. White-box arms: OLens ×4, J-lens, NLA L42 ×4, multi-layer NLA ×4.

Origin: the folder results_2026-10-14/ on Modal volume chive-archive holds chive_points4/ (the point-level judge), chive_mlnla/ (multi-layer NLA agent records only) and chive_claims/ (a separate claims experiment). The transcript arm's own agent records were never copied into that archive. They come from the CHIVE case viewer's data export, and I checked them against the archive: per-incident judge pass rates match judge/judge_pass1.json exactly on 204/204 incidents for transcript, OLens, J-lens, NLA and multi-layer NLA. Transcript pass@1 = 70.1, pass@5 = 82.8 (the published numbers).

Files

path	what	n
transcript_arm/<prompt_id>.json	transcript-only investigator: 5 episodes per incident, judge verdict inline	204 incidents, 1,020 episodes, all stopped_by=finish
judge/judge_pass1.json	free-form judge: per-incident pass rate (passes / 5 draws), for 5 arms: transcript, olens_rl_B, nla_B, mlnla_B, jlens_B	204 per arm
judge/per_incident.json	point-eval metrics per incident (claims, disproved rate, recall, rivals, pass@1) for transcript, olens_rl_B, mlnla_B, jlens_B	198 per arm
judge/points4_points/<pid>.json	the atomic ground-truth points per incident (Opus split of CHIVE's causes)	204
judge/points4_tags/<pid>.json	primary[i]: is point i part of the primary cause T1	203
judge/points4_grades/transcript/<pid>.d<k>.json	point-level grade of each transcript answer	1,015 (203 incidents × 5)
judge/points4_grades_rep/transcript/…	independent re-grade (reliability subset)	80 (67 incidents)
judge/points4_rivals/transcript/…	does the answer assert / reject each refuted rival	925 (186 incidents with rivals)
judge/points4_report.md, points4_summary.json	point-eval report, all 4 arms, 3 strata
judge/judge_rubric_detective_joracle_chive_judge.py	the free-form judge code + rubric
incidents_ground_truth.jsonl	the 204 incidents' behaviour, rate, true causes (targets), refuted rivals (ruled_out); subset of the public HF dataset	204
FAILURES.md, failures.json	the transcript-failure slice, with white-box arms on it

Schemas

transcript_arm/<pid>.json: {dataset, prompt_id, arm, episodes: [...]}. Each episode has:

• draw (0–4), stopped_by, seconds, output_tokens;
• system, seed: the investigator's system prompt and incident brief;
• steps: tool calls {tool, args, out, out_len, pos}, median 8;
• notebook: the agent's own notes;
• visited / cited: token positions;
• result: the final answer {hypothesis, evidence, rival, rival_ruled_out_by};
• judge: {score, pass, model}, plus replicates, a list of 3 {score, best_target, matched_exclusion, rationale}.

Free-form judge:

• Opus 5 (claude-opus-5) grades the final hypothesis 1–10 against the incident's ground truth.
• 3 replicates, pass = mean ≥ 7.
• best_target names the ground-truth cause it matched (T1, T2, …). matched_exclusion names a refuted rival it matched, if any.

Point eval (points4_*, Opus 5):

• grades: {prompt_id, arm, claims: [{claim, label}], point_scores: [...]}.
  • Claim labels: supported / disproved / uncheckable / rejected.
  • point_scores[i] is 1–10 for points4_points/<pid>.json point i.
• rivals: one of asserted / rejected / absent per refuted rival.
• Points scoring ≥7 count as stated.
• Re-grade reliability: point-score correlation 0.92, same stated / not-stated call on 95% of points (n = 297 answers, all arms).

Not included (available on request)

• The white-box arms' agent records (same format, 1,020 episodes each). Only their per-incident judge pass rates are here, in judge/judge_pass1.json.
• The point grades of the white-box arms.
• The chive_claims experiment.
• The activation readout banks.

Hygiene

• Scanned for API keys and tokens (Anthropic, OpenAI, HF, GitHub, AWS, Google, Slack, JWT, private keys, OAuth strings), local paths and personal names or emails. Nothing from our side was found.
• One redaction: incident prompt_003934 is a WildChat user's Python script with a hardcoded, real-format OpenAI key (sk-proj-…). It appeared in that file's tool outputs: in full once, in phrases, and token by token in show_tokens dumps. I replaced it with sk-proj-[REDACTED-KEY] / # characters everywhere. A side effect: the agent's notes quoting the key's T3Blbk… marker read [REDACTED-KEY].
  • The original conversation, key included, is still in the public CHIVE / WildChat data.
• The rest is CHIVE / WildChat conversation content plus model-generated investigator text and judge rationales.

Where the transcript-only investigator fails

See FAILURES.md: the transcript arm fails all 5 draws on 35/204 incidents (17%), and on those OLens ×4 reaches 30.3% pass@1 vs 19.4% for black-box resampling (+10.9 [+1.7, +20.6], paired bootstrap, re-run draws).

Related

• Incidents and ground truth: agu18dec/chive-forensics
• Conversation content comes from WildChat (ODC-BY), SystemChat2, Hermes and ToolACE; the investigator text and judge rationales are model-generated.