Exact

Exact Union RegExp

Repository files (latest commit first)

Filename	Latest commit message	Latest commit date
Transaction Science 4346fc0931 wai-quantum: the compilation flow's renaming test holds each exchange it names, and the benchmark note says what was measured ... An independent check of ba1cbb557 found its test and change note saying more than was measured. No verdict or receipt changes. engine::miter::flow::tests::the_renaming_finds_each_exchange_and_changes_nothing said it held an exchange written, one in three CXs, one in three CZs between Hadamards and one that absorbed a CX; its one program held a SWAP written and one exchange found (three CZs between Hadamards after a CX on the pair). It now holds four programs, each first checked to be its source's compile amplitude for amplitude: an exchange in three CXs, one in three CZs between Hadamards, one in two CXs that absorbed the CX before it, and the first program; in each the exchange is found, no closing exchange is left, every source run is matched and the plan is the compiled program amplitude for amplitude. The 2026-11-13 note, corrected from the run's receipts (the aarch64 build host's before and after runs; the after run's are the receipts in the repository): - the equal pairs refuted are 59 in both runs, the 49 trapped-ion compilations (not the suite's circuits) besides the 10 superconducting ones it named; - peak memory is now over the 896 checks both runs finished (Grover 129 -> 127 MB, exact QPE 1080 -> 1009 MB): a stopped check records none, and the after figures it gave (140, 2174 MB) came from checks stopped before; - of the check-seconds' fall, 6595 -> 3907, 2259 s is the 16 checks stopped before (counted at the cap) and 429 s the 896 checks both runs finished (2755 -> 2326 s); over the checks each run finished the time rose, 2755 s (896) to 3187 s (909); - the ten pairs the change closed are proved in doubles, none exactly: five on the dense route alone (every entry within 1e-9, largest differences 9.8e-14 to 4.8e-13), one on the near-identity rule alone, four on a diagram in doubles and the dense route; a receipt's brief form keeps no tolerance or near-identity statement, and brief()'s doc and the note say how to read it; - DENSE_WORK_PER_STEP (256) was chosen with the benchmark in hand: the walk at 9 on the 54-qubit device takes 50.4% of the bound and would be past 128 by 0.9%; at the run's median rates (2.3 ns an update, 2.2 us a diagram step) the bound is about a quarter of the diagram's time; its doc says so; - near the cap the outcome is the host's: the QPE at 22 and the QFT at 16 (36 checks) again on the Linux x86_64 build host (32 cores, load 81 -> 42, built there from cd7abe66c, same files, cap and budget): five of the seven checks that finished undecided at 181-216 s on the aarch64 host were stopped at the cap, two finished undecided at 183 and 184 s, and the other 31 gave the same verdict bytes on both hosts. Flow test and clippy -D warnings in five of the crate's feature sets (two on wasm32-wasip2) on the macOS arm64 build host. Change note: wai/changes/2026-11-13-equivalence-compilation-flow-corrections.md. Co-Authored-By: Claude Opus 5.5 <[email protected]>		2026-11-13 18:03:08 -05:00
..
common.rs	wai-quantum: the compilation flow's renaming test holds each exchange it names, and the benchmark note says what was measured	2026-11-13 18:03:08 -05:00