vinhable/able_e6_400_rollouts

able_e6 vs CXXV King — four complete scored rounds

Public offline-evaluation artifacts arranged as 4 rounds × 100 sample duels. Every duel contains two independent Challenger trajectories and the two corresponding stored CXXV King trajectories. The canonical checklist scores are embedded directly in the rollout rows and duel manifests.

Directory layout

rounds/
  round-001/                 # samples 001–100
  round-002/                 # samples 101–200
  round-003/                 # samples 201–300
  round-004/                 # samples 301–400
    challenger/
      able_e6-r1.jsonl       # our checkpoint, replica 1
      able_e6-r2.jsonl       # our checkpoint, replica 2
    king/
      cxxv-king-r1.jsonl     # corresponding King, replica 1
      cxxv-king-r2.jsonl     # corresponding King, replica 2
    duel-manifest.jsonl      # pairs all four trajectories and their scores
evaluation/
  summary.json               # complete aggregate evaluation output
  trajectory-scores.jsonl    # 1,600 final majority-vote trajectory verdicts
  judge-votes.jsonl          # 4,695 individual raw judge votes
  retry-errors.jsonl         # 137 superseded failed attempts retained for audit
duel-summary.json             # layout, counts, and headline evaluation metrics

The same five duel files exist inside every round directory. Each JSONL contains exactly 100 rows.

Sides

• Challenger / ours: vuhaian/able_e6; duel_side: "challenger" and participant: "able_e6".
• King: historical CXXV production King for the same task; duel_side: "king" and participant: "cxxv_king".

The King trajectories came from the frozen production-evaluation artifacts and were not regenerated.

Score fields

Every trajectory row exposes the most useful score values directly as top-level columns so they are immediately visible and sortable in the Hugging Face viewer:

• trajectory_score and trajectory_yes_rate;
• judge_votes_held and judge_disputed_questions;
• score_provider.

The same row also retains a complete evaluation object containing the final score, per-question majority answers and explanations, vote identifiers, provider, repeat counts, disagreement count, and elapsed time. No lookup in the separate evaluation files is required to interpret a rollout.

Every duel-manifest.jsonl row has an evaluation object containing:

• trajectory_scores: Challenger r1/r2 and King r1/r2;
• challenger_mean and king_mean: average of the two corresponding replicas;
• delta: challenger_mean - king_mean;
• result: win, tie, or loss from the sign of delta.

For convenient browsing, the manifest also exposes these top-level columns: challenger_r1_score, challenger_r2_score, king_r1_score, king_r2_score, challenger_mean_score, king_mean_score, score_delta, and duel_result.

The judge protocol used three votes per non-deterministic-zero trajectory and aggregated the majority independently for every checklist question. The 35 deterministic-zero trajectories did not require API judge calls.

Counts and aggregate result

• 4 rounds, 400 duels, 1,600 trajectories total;
• 800 Challenger trajectories and 800 King trajectories;
• 1,600/1,600 final trajectory records scored, 0 parse failures;
• Challenger mean: 0.43017491;
• King mean: 0.6509097725;
• Mean delta: -0.2207348625;
• Challenger W/T/L: 72 / 37 / 291 (win rate 0.18).

The score journals were append-only. retry-errors.jsonl contains 137 initial failures that were superseded by successful retries; they are preserved solely for provenance and are not used in the published final scores.