Benchmark DeepSeek V4 Flash high on all 12 TensorCircuit tasks

[QingyunQian]

Summary

Adds a complete single-trial ORBIT-Q benchmark of deepseek-v4-flash at high reasoning effort, with gpt-5.6-sol/high as the independent verifier.

Result

• Final validity: 5 / 12
• Passed: Tasks 03, 04, 05, 10, 12
• Functional/static passes: 7 / 12
• Audit passes: 5 / 12
• Five tasks reached the 1,800-second Agent limit without a candidate; Tasks 06 and 07 passed functional checks but failed source audit.

Included

• concise human-readable README and per-task table
• P/F comparison figure covering Sol high/ultra, Terra high, Luna high, and DeepSeek V4 Flash high
• solver wall-time, token-use, and configuration-efficiency figure
• selected solutions, rewards, audit/functional outputs, solver and trial logs, configs, metadata, and task-copy provenance
• scripts to prepare, archive, regenerate figures, and verify the archive

Protocol and validation

• TensorCircuit Docker tasks, sequential execution, one valid non-infrastructure outcome per task
• frozen task base commit 0201238ec2983907e2891f5319f5fff2d00844d5
• aggregate task-copy SHA-256 19fe27b83eaf668b3df32d1a68902b08cbe28585f189290769018eb16d927895
• solver recorded as deepseek-v4-flash, reasoning_effort=high
• auditor recorded as gpt-5.6-sol, model_reasoning_effort=high
• archive verifier checks all 12 outcomes, configuration stamps, task and solution hashes, and figure dimensions

Validation command:

python3 results/deepseek-v4-flash-high/tools/verify_archive.py

Result figures

Final task outcomes

Agent-side resource use