Elfsong/Mystery-Benchmark-Full

Mystery Benchmark full evaluations and partial GPT-6 Sol snapshot

This gated dataset contains complete Kimi K2.5, GPT-OSS 120B, and GPT-6 Luna evaluations, the GPT-OSS 120B guarded ablations, and a partial GPT-6 Sol snapshot. The index has 9,688 validated episodes: 9,000 from the previously published complete runs and ablations, plus 688 of 2,000 planned Sol episodes. The Sol run was stopped at the user's request; its missing 1,312 episodes are not scored as model failures. The Sol snapshot passed completion and structural validation; full world-state replay was not performed. All runs share the same 1,000 fixed nominal benchmark cases.

The ablation policies are full_minus_accuse, delay_control, and prompt_ledger. See run/gpt_oss_120b/ablations_20261003/RESULTS.md for the results and comparison with the original GPT-OSS vanilla and guarded baselines. The baselines were not rerun in the ablation job.

Contents

• data/episodes.csv: previewable combined index with per-episode metrics and raw/shard paths.
• data/cases.jsonl.gz: all 1,000 source case files used across these runs, including ground truth.
• data/trajectories/: 55 shards, including GPT-OSS ablations, GPT-6 Luna, and partial GPT-6 Sol; one row per episode with every ordered trajectory record (header, steps, footer).
• raw/: original JSONL trajectory archives for all runs, including codex_gpt6_luna/trajectories.tar.gz, codex_gpt6_sol/trajectories.tar.gz, and labeled interrupted-attempt archives.
• raw/*/executions/, raw/*/failed_attempts/, and ablation execution/failed-attempt archives: execution ledgers and failed retry artifacts. Kimi also includes the original queue archive.
• run/*/: summaries, validation, configurations and per-episode metrics. run/codex_gpt6_luna/episode_metrics.csv has all 2,000 Luna records; run/codex_gpt6_sol/episode_metrics.csv has the 688 validated Sol records.
• SHA256SUMS: SHA-256 checksums for all published files.

Join episodes to source cases by instance_id and verify using source_sha256. The trajectory value is the path inside trajectory_archive; the corresponding sharded row is in trajectory_shard and can be matched by model, policy, and instance ID.

Interpretation

The nominal 1,000 case files comprise 874 byte-distinct worlds because the benchmark generator can retry seeds. Do not treat all nominal cases as independent world samples. See each run's metadata under run/*/ for scoring details. Model access and decoding settings differ across runs. GPT-6 Luna and GPT-6 Sol used the ChatGPT-authenticated Codex CLI and a JSON Schema, not direct model APIs. Paired comparisons should use the supplied fixed case identities and note those settings.

GPT-6 Sol partial snapshot

Snapshot: 2026-10-13T09:58:21+00:00. Validated: 688/2000 episodes across 344 nominal cases. Interrupted attempt files: 77; these are archived separately and excluded from the episode index and metrics.

run/codex_gpt6_sol/snapshot_summary.csv summarizes only observed episodes. Its coverage varies by policy and difficulty, so these preliminary metrics must not be read as full-benchmark scores. See run/codex_gpt6_sol/validation.json for missing episode identities and run/codex_gpt6_sol/progress.json for the stop status.

Access to this dataset is gated because raw trajectories and run metadata are included.