dfthc_source/BEST.md

Verified result: 31.183% below baseline

The30% objective is achieved. Ready submission: submission/.

Quantity	Original baseline	Frozen result
Score, lower is better	3,355,599,588.3155365	2,309,230,097.5865555
Improvement	—	31.182788%
Total Toffolis	93,237	65,095.562940
Total qubits	730	741
R / B / C	65 / 60 / 60	65 / 60 / 28
brot / J / K	12 / 8 / 10	10 / 11 / 7
Effective normalization	49.30136925057439	47.873768372534656

Both default verification and independent-salt verification PASS. The independent score is2,309,230,043.6354394, a31.182789% improvement. Each run covers 254,464 address-shots. Small score differences reflect fully corrected measurement-dependent phase-cleanup operations.

Two-body error is19.311558mHa, one-body error12.935649mHa; raw Lambda294.02612134123535, gap3.923618193151924. All acceptance gates pass. Every circuit has exactly128 qubits:12 selection,110 word and six workspace. ANGLES cost56,615.562940 Toffolis; heuristic costs total8480 (outer606 + inner5024 + Givens2850). Shared preparation workspace is543.

Canonical digest: 7b79a311bfb66d9e9573cb8b12023ad1b852ed07ae7dae75f79037fde98171ea. certificate.json audits frozen file hashes, both results, the30% threshold, original scoring-source integrity, reference data, circuit widths, numerical invariants and branch-check evidence. The original verifier, cost model, baseline submission, reference integrals and fixed shifts remain unchanged.

A fresh rebuild reproduced every frozen circuit and NPZ file hash exactly; see reproduction_check.json.

The numerical method alternates whitened coefficient fits with coupled integer word moves. Accounting for interactions between moves permits hundreds of simultaneous angle updates, making ten-bit accuracy feasible. Compression to C28, gap-targeted coefficient fitting and exact orthogonal copy rotations then recover the required normalization. One-body words are regenerated from the decoded two-body representation. The frozen numerical checkpoint is experiments/precision_30/C28_b10_final.npz.

The lookup uses a high-bit Möbius/ANF decomposition, omits invalid-address monomials, and merges deferred measurement phases using six quadratic features and at most two executed CCZs. An exact proof and independent review cover arbitrary word inputs and every measurement branch. Tests include132,512 exhaustive small-instance cases,43,648 phase-synthesis cases,524,288 independent kickmix shots on the prototype, and222,656 arbitrary-word/inverse cases on these final emitted circuits. Finite tests supplement the proof.

Rebuild the exact submission from the frozen checkpoint:

cd /home/gwen/Quantum
OPENBLAS_NUM_THREADS=4 PYTHONPATH=kickmix:verifiers .venv313/bin/python research/qrom_30/anf_merged.py experiments/coordinator_verified/merged_C28_b10_30/submission/submission.npz /tmp/dfthcsos-30-rebuild --brot 10 --batch-size 11 --workspace 6
OPENBLAS_NUM_THREADS=4 PYTHONPATH=kickmix:verifiers .venv313/bin/python verifiers/dfthcsos/verify.py /tmp/dfthcsos-30-rebuild --salt independent-reproduction
OPENBLAS_NUM_THREADS=1 .venv313/bin/python research/shape_30/certify30.py

Compact reports: numerical fitting, lookup and Claude work, compression and Qwen assessment, shape experiments. Claude supplied a useful tool-enabled ANF prototype; native review corrected its test-coverage overclaim. Qwen's initial cost analysis was rejected, then a narrowly specified correction was independently checked. Native workers completed fitting, synthesis and verification. Previous verified submissions and research checkpoints are preserved. No Modal compute was used.