Repository navigation
Enforce simple scatter supports at opt-in checker and exporter boundaries - #2992
Merged
Mister-Meeseeks merged 1 commit intoOct 16, 2026
Merged
Conversation
Mister-Meeseeks
pushed a commit
that referenced
this pull request
Oct 17, 2026
…ew kappa) Answers PR #3004's third open question for one full-size word. Row 0 [4,5,6,7,8] of the selected PR #589 bit word is extended four ways and each word runs through the unchanged symbolic checker, record exporter, both entry modes of PR #2992's simple-scatter-v1 guard and the native legality and five-stage column checkers. Appending [4,4] triples an existing slot: the symbolic and both native checkers accept it, and all three exporter artifacts and both native receipts equal the baseline byte for byte (27,234 records, 24,305 ADDs, rank mass 5129, nine COPY/ERASE pairs). Appending [0,0,0] triples an absent slot: symbolic and exported actions agree, but both miss the required endpoint at output 111 (98 columns; dirty input 472 is a one-bit witness), and the full symbolic and native column checkers reject it. PR #2977's paired [0,0] reproduces symbolic acceptance with native rejection. The guard rejects all three repeated rows before creating output, as its stricter schema intends. Native legality accepts all four words, so legality does not certify the endpoint. Occurrence counts are not emitted ADD counts: the accepted triple has 529 occurrences and 527 scatter ADDs. This is a finite, author-checked audit with a separately written harness. Native builds use the pinned checked 128-bit cpp_int stand-in. No schema decision, supplier adoption, cold ancestry, system-Boost build, physical hosting or kappa change is claimed. Adds research/scatter-odd-native-audit/ only. Submitted by @commdata2338 with substantial AI assistance (OpenAI Codex research and implementation; packaged with Claude Code). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01X86sCSXdoDd61SDFfC1PKG
Mister-Meeseeks
pushed a commit
that referenced
this pull request
Oct 20, 2026
…export boundary Answer the supplier-boundary cascade's remaining open question (PR3696 OQ1, following #2977/#2992/#3004 and scatter-parity-contract-audit): for one gate (d, srcs), the pinned symbolic checker operator (accumulate-then-write) and the pinned exporter operator (sequential per-source ADD) agree as linear maps over F2^n iff d is absent from srcs, or the destination multiplicity is odd and every register occurring before the destination's last occurrence occurs an even number of times there. Measured on 9330 (d, srcs) pairs (6 registers, source lists length 0..4, the published census domain, extended to 55986 pairs at length <= 5): destination- free 4686 all agree; destination-in 4644 split into 3492 divergent and 1152 agreeing; simulation and the independent coefficient prediction agree on 100% of pairs. Guard census: the checker's own contract and distinct-only and dest-free-only are sound with 4608 / 4602 / 1152 agreeing rejections (smallest nonempty destination-free witness [a,a]); a transplanted scatter simple-support guard is unsound, admitting 1350 divergent gates; the exact minimal guard is sound with zero over-rejections. The pinned selected tuned word has 5472 D/A/B/C gates, all with distinct nonempty operands, so every sound guard preserves its export; its 111 scatter rows are simple supports. No new kappa (selected 0.001672660197076918 unchanged); no production adoption; frozen packages unchanged. Reproduce: python3.12 -B scripts/research_package.py --package research/gate-operand-parity-contract
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR2977 demonstrated that
scatter[0] += [0,0]passes the selected full symbolic checker but exports an extra ADD. This opt-in successor adoptssimple-scatter-v1and enforces the same simple-support validator before both checker and exporter execution. It snapshots the validated word and authenticated delegate script before running the unchanged source, so replacing the original input path cannot bypass the guard.The selected word produces all three original artifacts byte for byte (27,234 records); a separate scalar F2 interpreter checks all 661 arbitrary-input columns. Both entry modes reject the paired mutant before output creation, while existing-output and semantic-failure controls retain the promised boundary. Export mode validates the schema only; it does not replace full symbolic/native checking.
Validation passed locally:
-I -O -Werror -B.Sources and task are pinned at canonical main
836b19bfa309ceafbc8ee32a15e8c4b10dde7e11, priority revision 6. The accepted draw was bit mechanisms; TASK.json explicitly records the change frombit-helper-contractto the revision-6/PR2977 boundary handoff. Final overlap coverage: 234 open PRs, 2,440 closed PRs and 1,703 discussion records, complete pagination and no warnings; sparse local-note/diff coverage remains a limitation.This completes a local successor boundary module, not selected-supplier adoption. Frozen sources, manifests and the selected frontier are unchanged. Native compilation, complete selected cold replay, hosting/precision and all-size admission were not run; hosted CI is separate.
Credit Nick Petalas / PR2977 (
cd23d242dfe0e7a71cafba0ad5590dc4f2fbf4f2) for the counterexample, candidate guard and handoff; eumemic, DreamingOfClouds, Dylan Crow and Douglas Colkitt retain the source-lineage credit detailed in the README. Prepared for Arham Tahir with substantial OpenAI Codex assistance; Apache-2.0, no independent human review claimed. Canonical investigation and claim.