Task 05: TensorCircuit-native cooling with stable OMECo paths

[QingyunQian]

Take-home insight

For the fully TensorCircuit-native tensor network, the measured acceleration comes from giving OMECo enough search budget to avoid bad contraction paths. Exact pair fusion makes the graph smaller, but its isolated five-pair effect is unresolved and is not claimed.

This supersedes the earlier custom no-QR MPS prototype: the new solution keeps the quantum process inside TensorCircuit APIs (tc.Circuit, unitary, expectation, QuOperator, and mpo_expectation).

Measured result

Direct comparison	Mean paired speedup (95% t-CI)	Outcome
OMECo 4×4 vs 1×1, same fused circuit	1.420x [1.181, 1.659]	5/5 wins; retain
Fused vs unfused, both OMECo 1×1	1.198x [0.916, 1.480]	unresolved; do not claim

Across five cold-process pairs, the final candidate ran in 60.114 ± 2.092 s and beat the public expert 5/5 times. Median paired speedup was 1.668x; mean paired speedup was 1.939x [1.105, 2.772]. Excluding one visible 180.22 s reference long tail only as a sensitivity check gives 1.639x.

Scope and validation

• only the concise report, optimized answer, and factor-ablation SVG;
• official functional evaluator: PASS, 600/600 updates;
• static policy: 1.0, 117 effective lines, no raw-simulator bypass;
• original module docstring unchanged;
• full per-pair data and interpretation are in the report.

The local automated LLM-audit stage was unavailable because its verifier Codex credential returned 401; functional and static checks completed normally.