Skip to content

ACGN v2.11

Choose a tag to compare

View all tags
@AlexandervonWu AlexandervonWu released this 30 Aug 01:36
· 15 commits to aislop since this release
v2.11
63f3b8c

ACGN v2.11

This release publishes the August 29, 2026 clean full-corpus replay after
repairing the Fast Rewrite IR field-ownership and inherited-temporal alpha
alignment boundaries. The Certificate-Integrated pathway and its fail-closed
semantic boundary remain intact.

Provenance

  • Result-producing source commit: fbd9b1497a9036c55780da777f56581bc1c6bcec
  • Clean publication run: df4d8d4c-6265-4fe7-88d5-3aceee60398b
  • Final packaging and tag commit: 63f3b8c2931202e9b3582a8802c558725a67ac26
  • Dataset: 66,080 Alloy files; 61,598 eligible AST-different pairs
  • Imported manifest-bound stage outputs: 5,808 files
  • Configuration: Java 17.0.20, 16 workers, 8 GiB heap, reward pool 100
  • Exact-commit Actions run: https://github.com/AlexandervonWu/ACGN/actions/runs/33281721348

Reviewers must run git lfs pull before verifying the two large JSON result
files. Run ./scripts/verify_imported_publication_snapshot.sh to check every
imported stage artifact against the archived run manifest.

Headline results

  • Paired evaluation: 61,598 successes, 4,482 AST-identical skips, 0 failures,
    and 0 incorrect zeroes under either canonical path.
  • Augmented evaluation: all 42,386 incorrect predicates ranked and rewarded,
    with 0 preparation, ranking, or reward failures and 0 incorrect-to-truth
    zeroes under both Fast Rewrite and Certificate-Integrated distance.
  • Seven-arm ablation: all 431,186 eligible arm rows succeeded. The
    Certificate-Integrated IR certified 4,088 CORRECT zeroes, retained all
    4,074 Fast Rewrite IR zeroes, and added 14.
  • Bounded natural-corpus semantic replay: 4,088 unique claims, 0
    counterexamples, and 0 errors; four targeted negative controls remained
    unmerged.
  • Capability matrix: slotted e-graph, Fast Rewrite IR, and
    Certificate-Integrated IR each recovered 5,500/5,500 generated pairs.
  • Rewards and reward correlations are present. Both rewarded stages used pool
    size 100 and reported 0 reward failures.

Repair in this release

The preceding snapshot exposed ten Fast Rewrite-only incorrect-to-truth
zeroes. Nine erased the owner coordinate of same-named fields, and one allowed
inconsistent alpha mappings for an inherited binder across temporal phases.
Fast Rewrite now retains field-owner identity and minimizes one coherent
inherited-temporal mapping. The full replay reduces both incorrect-zero counts
to zero while preserving all previously observed paired CORRECT Fast Rewrite
zeroes. The Certificate-Integrated pathway was not broadened by this repair.

Certificate and assurance boundary

Certificate coverage remains bounded and fixture-scoped. The exported
certificate census is 1 VERIFIED, 2 UNCHECKABLE, and 0 REJECTED; it is
not a corpus-wide proof of Alloy equivalence. The checked-in Section 3
assurance record remains INCOMPLETE while its documented metatheoretical and
refinement obligations remain open.

Assets

  • acgn-experiments.jar: 2,292,989 bytes; SHA-256
    361b33ef56f6ccb1089a7a6fdda2a92bf621e501166c5a1c73330a0cc1686807
  • SHA256SUMS: SHA-256
    798a0dac47d0a186ca91eab6368adfda63434538895db660dbdc431a25d4a69c

The JAR is the byte-identical archived experiment artifact from the clean run;
it was not rebuilt for this release.

Assets 4
Loading