Repository navigation
Reduce H25 side-verification time by 29% with exact native GF(2) - #5929
Conversation
|
FINDING / DONE / RELEASED — complete frozen H25 side verification passes the 15% wall/CPU and noise gates; broader scopes remain unclaimed.
|
|
CI CONFIRMATION — exact publication head a456c24 passed run40514711386. The plan job passed676 shared tests with22 skips and selected only research/gf2-side-verification. The scoped replay then passed all three commands on CPython3.12.15: fresh native finite/source/binding/fallback checks in normal and optimized modes, and complete canonical witness/rank transport. The final tracked-input and index diff checks were clean. Raw exact-head diagnostics are retained locally; the plan log excerpt is explicitly filtered/truncated by the diagnostics service. This supersedes the pending-CI status in the PR body and seven-field handoff. Hosted CI did not rerun full H25 timings or establish a hosted performance factor. The29.2746%wall/29.2773%CPU claim remains the declared six local cold processes with identical full scientific observations and all added costs paid. No measured source/sample or package commit changed after submission. PR remains open/unmerged; no frontier or kappa claim. The separate further15%PR5864 target remains unmet. |
|
DONE / RELEASED — the complete-caller follow-up is published as PR5995, head The full frozen historical H25 CLI with retained observations passes the declared six cold BCCBBC comparison: median wall 39.686181760→28.636256246s (27.8433% reduction), CPU 39.58→28.46s (28.0950%). Both exceed 15% and repeat-spread gates. All samples, including slower C3, remain. Every candidate pays cold compile/link/load, independent dense/finite/source/type/fallback controls and identical full output work. The unchanged full caller executes all original arithmetic/assertions, real 19-file source verification, certificate write and both stdout lines. Exact certificate, typed return, ordered side witness/rank and complement observations match. Fresh finite normal/O, complete read-only Git-free C correctness, canonical transport, metadata, committed bytes/modes/source ancestry and 125-file export/all 42 inputs pass. Entry-import and publication-cleanup failures are preserved with repairs; measured source and samples are unchanged. See PR5995 for current hosted exact-head correctness status; CI does not rerun or establish the local performance factor. This resolves the specific next question from PR5929. Its separate 29.27% side result remains unchanged. No bare-CLI, selected-frontier, mathematical/kappa or PR5864 percentage transfer is claimed. The further 15% PR5864 goal remains blocked/incomplete. No extra optimization or timing reroll is nominated. Kappa/Work1/Work2 mathematics, unrelated feeds and automations stay paused. The original bounded claim began 2026-10-27 19:48UTC at canonical main 19fdb08 (priority revision 8); this result supersedes its pending state and releases the scoped claim. Original authors/notices and PR5929/native precedents remain credited; Astra xhigh source design/static reviews and one parent serial runner. |
Change and mathematical scope
The complete frozen H=25 complex-side verification segment takes 24.807329s median wall / 24.76s CPU with checked native GF(2) complement and Gram kernels, versus 35.075541s / 35.01s for unchanged Python: 29.2746% wall and 29.2773% CPU reduction across the declared six cold B,C,C,B,B,C processes. Both savings exceed 15%, 0.50s, and the maximum within-arm repeat spread. Every run pays fresh compile/link/load, added finite/binding/fallback checks, and identical complete witness serialization and complement-stream hashing.
This opt-in tooling package preserves source assertions, pivot/vector order, the exact GF(2) field, every verification stage in this side segment, and all four complete scientific identities. It does not establish a full complex-network certificate speedup, a further PR5864 improvement, or any mathematical/kappa/frontier claim. The separate PR5864 integer/formatting and rational findings are retained in #5908; their further 15% target remains unmet.
Task and overlap
Task replay-runtime, refined scope frozenH25/complete-side-verification/native-GF2-complement-Gram; see TASK.json. Pinned original source is de5ec2e, with unchanged copies introduced by ancestor b47911b. Current main a3e1e3f was fetched immediately before submission: both original source files remain byte-identical. The branch retains Ali Romman's PR5761 routing-test repair 01014cb; it is now in main, and the current compare contains only this package's 113 files. Submitted head: a456c24.
The unchanged caller's order is construction, Checks initialization, physical-role compilation, all verify_map, verify_labels and verify_role_frames. Only Label.complement_in and Label.nondegenerate gain opt-in checked uint64 native paths; unsupported widths/types fall back to the original Python. A bounded open-PR H25 search returned zero matches; this is an overlap check, not an exhaustive novelty claim. Adjacent native precedents are #5168, #5382, #2853 and #3143. The contribution is a measured source-bound verifier implementation, not new linear algebra.
Validation
All six complete runs match result SHA 342e99212b12e518902ae9598613f5484113bacee8747c27a13b58a4b1d3986c, full witness SHA 4b170c1dbfc33c7cb141a2284c6fd69eac7b6d100971d3e9ee625502b5893f82, ordered Gram-rank SHA 85006ca332bbc57c2ba9fd06f61a0281d4cee600c9f3b449342379bf2aedec8f, and ordered complement SHA 7dec2bf33303c37e1c699609e83768d61df7e19104d34a26a148f66d5ccc4838. That covers 59,959 rank calls and 238,685 complements, including 55,839,507 canonical complement bytes. The full 58,143,622-byte witness and 2,674,272-byte rank transcript are retained in lossless canonical gzip files; hashes of decompressed bytes are authoritative.
Finite checks pass in normal and optimized mode: 5,329 exhaustive ordered small complement pairs, 585 Gram cases, a separate dense-list oracle, 64/65-bit and augmentation boundaries, fallback/type behavior and source/live-environment/proxy drift rejection. Independent static review found helper-freshness and missing-complement-observation gaps in v1; v2 fixes both, preserving all earlier records without a v1 candidate timing. One declared literal anchor fixed the complement identity before the six comparisons. Complete verification requires normal assertion mode; optimized full verification is rejected.
Local environment: Linux x86-64, CPython 3.12.3, GCC 13.3. C source SHA e9c3c9194a08718bd1575c1cf08c6c6772d3b644c2e274c25a84a1f9249d06cc. Seven scientific core files exactly match the measured stage. Metadata, committed bytes/modes, Git-free export/inputs, normal/O finite checks, canonical transport and a complete readonly-source C replay pass. The readonly run preserves source bytes/modes and is a publication check, not another timing sample. Hosted exact-head correctness run40514711386 passed at a456c24:676 shared tests/22 skips, selection of only this package, both fresh native finite modes and canonical witness/rank transport, with clean tracked-input/index diffs. CI confirmation. Full H25 timings were not repeated by hosted CI; local timings are not hosted performance evidence.
Registered finite reproduction:
python3.12 -B scripts/research_package.py --package research/gf2-side-verification
Complete fresh guarded replay (new external directory):
Arm B selects unchanged Python. One serial runner, 180s per command, 1GiB sampled process-group RSS, zero automatic retries. Recorded round: 600s initial feasibility and 1,500s shared cap. No dependency installation or compiled binary is included. See README.md, VALIDATION.json, evidence/MATRIX-COMPLETE-GATE.json and retained raw receipts. The entire original complex-network caller was not replayed and remains unpriced.
Credit
Original CrocSwap source, licenses, contributor notices and predecessor history remain unchanged. Credit to Ali Romman for the retained routing-test repair and source authors, and to adjacent native/GMP/FLINT work linked above. Prepared by Chris R / Clear1nsight with substantial OpenAI Codex implementation and user-requested GPT-6-astra xhigh design plus separate static review. Only the parent executed compiler/numerical work, serially. This is assisted engineering evidence, not human or proof-kernel acceptance of a theorem.
Open next questions
TASK.json records the smallest useful follow-up: a bounded profile of the complete original complex-network caller before proposing integration or a whole-certificate speed factor. Other stages, dependencies and resource costs are unpriced. Broader basis kernels would require canonical pivot/vector and full witness contracts. No shared helper, selected result, mathematical stream or research frontier is changed.