| { |
| "benchmarks": [ |
| { |
| "id": "imo_proof", |
| "label": "IMO-ProofBench" |
| }, |
| { |
| "id": "proofbench", |
| "label": "ProofBench" |
| } |
| ], |
| "experiments": [ |
| { |
| "id": "q9_base", |
| "label": "Qwen3.5-9B · harness methods", |
| "model": "Qwen/Qwen3.5-9B", |
| "seeds": [ |
| 0, |
| 1, |
| 2, |
| 3, |
| 4, |
| 5, |
| 6, |
| 7 |
| ], |
| "description": "Nine proof harnesses, 8 seeds, 8.4M-token trajectory cap, thinking enabled.", |
| "artifact": "qwen35_9b_harness_large_budget_v1", |
| "arms": [ |
| { |
| "id": "direct", |
| "label": "Direct", |
| "description": "One solver call, no verifier." |
| }, |
| { |
| "id": "gvr", |
| "label": "GVR", |
| "description": "Generator → Verifier → Reviser/Generator, up to 3 candidates; the verifier returns a verdict and a ≤600-char critique." |
| }, |
| { |
| "id": "gvr_subagents", |
| "label": "GVR + subagents", |
| "description": "GVR where the generator may spawn up to 3 fresh-context research subagents." |
| }, |
| { |
| "id": "gvr_reference", |
| "label": "GVR + reference", |
| "description": "GVR whose verifier also sees the benchmark reference proof (copied reference spans are filtered from feedback)." |
| }, |
| { |
| "id": "gvr_rationale_score", |
| "label": "GVR · rationale+score", |
| "description": "GVR whose verifier returns a success probability and a ≤1200-char rationale." |
| }, |
| { |
| "id": "gvr_reference_rationale_score", |
| "label": "GVR + reference · rationale+score", |
| "description": "Reference-assisted verifier with the rationale+score feedback." |
| }, |
| { |
| "id": "value_tool", |
| "label": "Value tool", |
| "description": "The solver may call query_success_probability up to 3 times; a fresh-context verifier sees the partial trace and returns only a probability." |
| }, |
| { |
| "id": "value_tool_rationale_score", |
| "label": "Value tool · rationale+score", |
| "description": "Value tool that also returns a rationale." |
| }, |
| { |
| "id": "cch_plan_work_review", |
| "label": "Plan → Work → Review", |
| "description": "Claude-Code-Harness-style plan, work, fresh review and bounded retakes (blind)." |
| } |
| ] |
| }, |
| { |
| "id": "q9_attempt", |
| "label": "Qwen3.5-9B · attempt-conditioned verifiers", |
| "model": "Qwen/Qwen3.5-9B", |
| "seeds": [ |
| 8, |
| 9, |
| 10, |
| 11, |
| 12, |
| 13, |
| 14, |
| 15 |
| ], |
| "description": "24 arms: prior-attempt evidence × GVR/value tool × feedback × reference access; fresh seeds 8–15.", |
| "artifact": "qwen35_9b_attempt_conditioning_v1", |
| "arms": [ |
| { |
| "id": "attempt_solutions_gvr_legacy_no_gold", |
| "label": "GVR · Full solutions · legacy", |
| "description": "GVR whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_solutions_gvr_legacy_gold", |
| "label": "GVR · Full solutions · legacy · gold", |
| "description": "GVR whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_solutions_gvr_rationale_no_gold", |
| "label": "GVR · Full solutions · rationale", |
| "description": "GVR whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_solutions_gvr_rationale_gold", |
| "label": "GVR · Full solutions · rationale · gold", |
| "description": "GVR whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_gvr_legacy_no_gold", |
| "label": "GVR · Solution summaries · legacy", |
| "description": "GVR whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_gvr_legacy_gold", |
| "label": "GVR · Solution summaries · legacy · gold", |
| "description": "GVR whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_gvr_rationale_no_gold", |
| "label": "GVR · Solution summaries · rationale", |
| "description": "GVR whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_gvr_rationale_gold", |
| "label": "GVR · Solution summaries · rationale · gold", |
| "description": "GVR whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_gvr_legacy_no_gold", |
| "label": "GVR · Solution+thinking summaries · legacy", |
| "description": "GVR whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_gvr_legacy_gold", |
| "label": "GVR · Solution+thinking summaries · legacy · gold", |
| "description": "GVR whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_gvr_rationale_no_gold", |
| "label": "GVR · Solution+thinking summaries · rationale", |
| "description": "GVR whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_gvr_rationale_gold", |
| "label": "GVR · Solution+thinking summaries · rationale · gold", |
| "description": "GVR whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| }, |
| { |
| "id": "attempt_solutions_value_tool_legacy_no_gold", |
| "label": "Value tool · Full solutions · legacy", |
| "description": "Value tool whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_solutions_value_tool_legacy_gold", |
| "label": "Value tool · Full solutions · legacy · gold", |
| "description": "Value tool whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_solutions_value_tool_rationale_no_gold", |
| "label": "Value tool · Full solutions · rationale", |
| "description": "Value tool whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_solutions_value_tool_rationale_gold", |
| "label": "Value tool · Full solutions · rationale · gold", |
| "description": "Value tool whose verifier sees full solutions of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_value_tool_legacy_no_gold", |
| "label": "Value tool · Solution summaries · legacy", |
| "description": "Value tool whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_value_tool_legacy_gold", |
| "label": "Value tool · Solution summaries · legacy · gold", |
| "description": "Value tool whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_value_tool_rationale_no_gold", |
| "label": "Value tool · Solution summaries · rationale", |
| "description": "Value tool whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_solution_summary_value_tool_rationale_gold", |
| "label": "Value tool · Solution summaries · rationale · gold", |
| "description": "Value tool whose verifier sees solution summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_value_tool_legacy_no_gold", |
| "label": "Value tool · Solution+thinking summaries · legacy", |
| "description": "Value tool whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); legacy feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_value_tool_legacy_gold", |
| "label": "Value tool · Solution+thinking summaries · legacy · gold", |
| "description": "Value tool whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; legacy feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_value_tool_rationale_no_gold", |
| "label": "Value tool · Solution+thinking summaries · rationale", |
| "description": "Value tool whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run); rationale feedback." |
| }, |
| { |
| "id": "attempt_thinking_summary_value_tool_rationale_gold", |
| "label": "Value tool · Solution+thinking summaries · rationale · gold", |
| "description": "Value tool whose verifier sees solution+thinking summaries of 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) plus the reference proof; rationale feedback." |
| } |
| ] |
| }, |
| { |
| "id": "q27_base", |
| "label": "Qwen3.8-27B · harness methods", |
| "model": "Qwen/Qwen3.8-27B", |
| "seeds": [ |
| 0, |
| 1, |
| 2 |
| ], |
| "description": "Nine proof harnesses, 3 seeds, 229k-token trajectory cap.", |
| "artifact": "qwen38_27b_harness_v4", |
| "arms": [ |
| { |
| "id": "direct", |
| "label": "Direct", |
| "description": "One solver call, no verifier." |
| }, |
| { |
| "id": "gvr", |
| "label": "GVR", |
| "description": "Generator → Verifier → Reviser/Generator, up to 3 candidates; the verifier returns a verdict and a ≤600-char critique." |
| }, |
| { |
| "id": "gvr_subagents", |
| "label": "GVR + subagents", |
| "description": "GVR where the generator may spawn up to 3 fresh-context research subagents." |
| }, |
| { |
| "id": "gvr_reference", |
| "label": "GVR + reference", |
| "description": "GVR whose verifier also sees the benchmark reference proof (copied reference spans are filtered from feedback)." |
| }, |
| { |
| "id": "gvr_rationale_score", |
| "label": "GVR · rationale+score", |
| "description": "GVR whose verifier returns a success probability and a ≤1200-char rationale." |
| }, |
| { |
| "id": "gvr_reference_rationale_score", |
| "label": "GVR + reference · rationale+score", |
| "description": "Reference-assisted verifier with the rationale+score feedback." |
| }, |
| { |
| "id": "value_tool", |
| "label": "Value tool", |
| "description": "The solver may call query_success_probability up to 3 times; a fresh-context verifier sees the partial trace and returns only a probability." |
| }, |
| { |
| "id": "value_tool_rationale_score", |
| "label": "Value tool · rationale+score", |
| "description": "Value tool that also returns a rationale." |
| }, |
| { |
| "id": "cch_plan_work_review", |
| "label": "Plan → Work → Review", |
| "description": "Claude-Code-Harness-style plan, work, fresh review and bounded retakes (blind)." |
| } |
| ] |
| } |
| ], |
| "status_codes": { |
| "A": "accepted", |
| "L": "cycle_limit", |
| "C": "completed", |
| "P": "protocol_error", |
| "B": "budget_exhausted", |
| "X": "context_exhausted", |
| "F": "failed" |
| }, |
| "verdict_codes": { |
| "c": "correct", |
| "m": "minor_fix", |
| "x": "critical_flaw" |
| }, |
| "cell_fields": [ |
| "seed", |
| "score", |
| "status", |
| "verdicts", |
| "generated_tokens", |
| "calls" |
| ] |
| } |
