Benchmark DeepSeek V4 Flash max on all 12 TensorCircuit tasks

[QingyunQian]

Summary

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

Result

• Final validity: 5 / 12
• Passed: Tasks 02, 05, 09, 10, and 12
• Functional/static passes: 8 / 12
• Audit passes: 5 / 12
• DeepSeek high and max both reach 5/12, but with different accepted task sets.
• Task 08 is a failure for every displayed solver setting under final expert adjudication.

Result figures

Final task outcomes

Agent-side resource use

Included

• concise README and complete per-task result table
• outcome and agent-resource figures in PNG, SVG, and PDF formats
• selected solutions, rewards, audits, functional outputs, logs, configs, metadata, and task-copy provenance
• scripts to archive, regenerate figures, and verify the result bundle

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 deepseek-v4-flash, reasoning_effort=max
• auditor gpt-5.6-sol, model_reasoning_effort=high

Validation command:

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

[QingyunQian]

For Benchmark results, I think Deepseekv4-flash high is enough, thinking effort max is not required