Best TfP-verified T-count for hwb6

Outcome

The current Submission is sub_ngii04aokv (art_bni5wt83lb, hwb6_t67.qasm): 67 T, 308 CX, 642 gates, verified with score 67 (vrn_7ua4sgxrbp). Earlier verified Submissions: 70 (sub_l4nzgxvmyx), 69 (sub_wj0c9za3lf) and 68 (sub_3d1di9ab90). The scripts are in art_9yi1fwshex (hwb6_scripts.tar.gz). Every verifier run reported "qcec fallback: all: signal 11; decided by alternating": the deployed QCEC crashes in its all-checkers mode and the alternating checker settles the result.

What worked, in order

1. PyZX teleport_reduce + full_optimize (TODD) with random restarts. TODD shuffles columns at random, and restarts went from 75 to 70.

2. Pauli-rotation layers (rot.py, built on stim):

   • Write U = C · ∏ π/4 Pauli rotations.
   • Merge equal Paulis and push Clifford rotations into C.
   • Split the rotations into mutually commuting layers in a random order.
   • Diagonalise each layer with a greedy Clifford and run todd_simp on its 7-variable phase polynomial.

   Alternating this with PyZX full_optimize gave 69.

3. Exact layer decoding (exact.py). For a layer of rank r ≤ 6, search all codewords of the punctured RM(r−4, r) code for the closest one (2^22 words at r=6, in numpy). This showed that the layers of the 69/67 circuits were already optimal. Gains have to come from changing the structure between layers.

4. Randomly reordering the reference's commuting classical gates (search6.py; first drop the cancelling H pairs) before teleport_reduce, then steps 1–3. This was the biggest lever: 68, then 67 (2 hits in about 1,100 reorderings; typical results were 70–71).

Dead ends

• Relative-phase Toffoli (RCCX/RCCX†) for the only compute/uncompute pair, ccx(2,4→6)…ccx(2,4→6): no gain, because phase teleportation already captures it.
• full_optimize directly on the reference: 78.
• An hour-long plateau random walk over 56 distinct 67-T circuits, plus about 800 more reorderings, found nothing below 67.

Checking

Every candidate was checked locally as a full 128x128 unitary up to global phase, and with MQT QCEC 3.11 with run_zx_checker=False. Default QCEC often returns no_information on these circuits because its ZX checker disagrees with simulation. I did not run the tfp verify CLI preflight.

Environment issues

• The pinned Skill could not be read: resources/read fails client-side schema validation (ttlMs/cacheScope). Reference inputs were fetched through download_url and their hashes checked.
• No Claude Code Remote get_session tool was available, so usage was not reported.

Advice for the next agent

• Start from search6.py with more seeds and CPU time; 67 came up at roughly 1 in 500 reorderings.
• Ideas I didn't try: TOpt/FastTODD for rank-7 layers; a smarter (non-greedy) choice of layers across anticommuting boundaries; reordering combined with alternative Toffoli decompositions.
• Keep CX in view: rotation resynthesis gave 300–400 CX, against a limit of 464.

sub_ngii04aokv was verified at 67 (vrn_7ua4sgxrbp). A second, distinct 67 (seed 42, 324 CX) also turned up. Distribution of final T-counts from search6, about 1,100 reorderings in total: mostly 70–71, a few 68, and 67 twice. Over 05:00–05:30 UTC, neither search6 (2 seeds, about 400 iterations each) nor the plateau walk from the 67 pool (search5, 2 seeds, about 270 iterations, 56 distinct 67-T circuits) found anything below 67. I'll let them run a little longer, then wrap up.

sub_3d1di9ab90 was verified at 68 (vrn_oxh883hw2z). A few minutes later search6 seed 41 (iteration 73) found 67 T with 308 CX. It passes the exact unitary check and QCEC with the ZX checker off, and is submitted as sub_ngii04aokv (art_bni5wt83lb, run vrn_7ua4sgxrbp). All 4 CPUs now run search6 with different seeds: random reordering of the reference's commuting gates, then teleport_reduce, full_optimize and rotation-layer decoding.

The plateau walk (search5, about 350 iterations × 2 seeds, 46 distinct 69-T circuits) found nothing below 69.

search6.py changes the starting point instead:

1. Drop the reference's cancelling H pairs.
2. Randomly topologically reorder the classical gates (X/CX/CCX); two gates commute when neither one's target is a control of the other.
3. Run teleport_reduce and full_optimize, then 2–6 rounds of rotation-layer optimisation (exact RM decoding where rank ≤ 6) interleaved with full_optimize.

Both seeds hit 68 T within about 35 reorderings (CX 381 and 391; the two-qubit limit is 464). Both pass the exact unitary check and QCEC with the ZX checker off. Submitted s6_42_35 as sub_3d1di9ab90 (art_l3nn7iu87c), verification vrn_oxh883hw2z. Lesson: PyZX's phase teleportation depends heavily on the order of the input gates, and that choice matters more than the decoding inside layers. The search continues with search6 and with search5 seeded from the 68s.

exact.py swaps PyZX's todd_iter for an exact minimiser whenever a layer's parity rank is r ≤ 6. It searches all codewords of the punctured Reed–Muller code RM(r-4, r)* (2^22 words at r=6, done in numpy) for the closest one to the odd-parity set, following Amy & Mosca. Rank-7 layers still use best-of-4 TODD. On the 69-T circuit, every layering tried (ASAP, ALAP and 4 random) still gives 69: each commuting layer already has its minimal T-count. Any further gain has to come from changing the rotation structure between layers, not from better decoding within one. Now running search5.py: a plateau random walk. It keeps a pool of distinct equal-T (69) circuits, deduplicated by their rotation multiset, and mixes in PyZX teleport_reduce/full_optimize re-extraction, in the hope of reaching a structure that drops lower.

In search2.py the random-layering rotation TODD alternates with PyZX full_optimize. All four seeds reached 69 T within about 35 iterations, each time with the same circuit shape (260 CX, 524 gates). One of these files was submitted as sub_wj0c9za3lf (art_ukxn9scr0r) and verified with score 69 (run vrn_i116nko7z2; again decided by the alternating checker after "all" crashed).

The 69-T circuit has an odd T-count, so it is not just two T gates cancelling in the 70. The gain comes from TODD across a commuting layer after a PyZX re-extraction moved the rotations. Searches restarted from the 69 pool (search3) have stayed at 69 for 80+ iterations. Now trying best-of-6 TODD per layer (search4).

Submission sub_l4nzgxvmyx (artifact art_wnpdo0f3mw) was verified with score 70 (run vrn_bbzg6gqtcv). Evidence line: "qcec fallback: all: signal 11; decided by alternating", so the deployed checker crashed in its all-checkers mode and fell back to the alternating checker.

Next approach (rot.py): rewrite the circuit as U = C · R_m…R_1, where each R_i is a π/4 Pauli rotation; this uses stim 1.16 tableaux. Then:

1. Merge equal Paulis that commute up to each other, and push even-k Clifford rotations into C.
2. Split the rotations into mutually commuting layers (ASAP, ALAP or random order).
3. Diagonalise each layer with a greedy Clifford, run PyZX todd_simp on the resulting 7-variable phase polynomial, and resynthesise.

The rewrite and resynthesis match the reference unitary exactly. Starting from the 70-T circuit, ASAP layers have sizes 10,4,5,6,7,3,4,6,4,8,3,4,6 and ALAP layers 6..11. TODD leaves every layer at its original T-count: they are too small to reduce. Now running a random-layering search that alternates with PyZX full_optimize (search2.py, 4 seeds).

PyZX's TODD (todd_iter) shuffles columns at random. The search (search1.py, 4 processes, Python random seeded per run) does: teleport_reduce → basic_optimization → full_optimize repeated 1–3 times, then keeps the lowest T-count. Early results: about one seed in two lands at 72, and many reach 70 T (246 CX, 457 gates). All four 70-T files match the reference as exact 128x128 unitaries. Default QCEC returned equivalent for seed 2001 and no_information for the other three, where the ZX checker disagrees with simulation. With the ZX checker off, all four come back equivalent. I'll submit the seed-2001 circuit, since default QCEC accepts it.

Separately, a dead end: replacing the compute/uncompute pair ccx(2,4→6)…ccx(2,4→6) with relative-phase Toffolis (RCCX and its inverse; only one such pair exists) does not lower the PyZX result (still 75/74). Phase teleportation already captures that saving.

Setup: PyZX 0.10.7, mqt.qcec 3.11.0 and qiskit 2.5.2 on Python 3.13. Results, emitted in the allowed gate set (h,s,sdg,t,tdg,x,z,cx; CZ as H-CX-H):

• full_reduce + extract + basic_optimization: 75 T, 143 CX
• full_optimize on the reference (TODD inside): 78 T
• teleport_reduce + basic_optimization: 75 T, 104 CX
• teleport_reduce, then full_optimize: 74 T, 159 CX (two-qubit limit is 464)

Checks: the full 128x128 unitaries agree up to global phase. QCEC with default settings returned no_information on the 74-T circuit because its ZX checker disagreed with simulation. With run_zx_checker=False, QCEC gives equivalent_up_to_global_phase. Whether the deployed checker reaches a conclusion on this circuit is uncertain. Next: random restarts of the TODD stage, and relative-phase Toffoli rewrites applied before optimisation.

No other Attempts or Notes on this revision yet (list_attempts, search_notes on the campaign both came back empty).

Blocker: the client rejected resources/read for the pinned Skill (tfp://skill-versions/skv_zqmrsnnedj/...) and for the reference resources. The error was a schema validation problem, ttlMs / cacheScope missing from the resource result. I fetched the reference inputs through their signed download_url with curl, and their sha256 matched content_hash (reference_circuit d25c21…, spec f5dcd3…). spec: gates h,s,sdg,t,tdg,x,z,cx; ccx at 7T; no extra qubits; 2q ratio ≤4; QCEC timeout 420 s.

Plan: start from the 75-T baseline (PyZX full_reduce). Then try to go lower by optimising the phase polynomial / TODD on the Hadamard-free parts (PyZX phase_block_optimize, if it is available), plus local relative-phase pairing of repeated Toffolis such as ccx(2,4,6)…ccx(2,4,6). Check every candidate locally with MQT QCEC before submitting.