[jeremykothe2]

Summary

This PR adds tooling only and makes no κ claim.

research/witness-spec-tools/ lets you describe a witness declaratively and get a complete, generic verifier for it. Today each new record forks the previous package and hand-edits constants spread through verify.py, bind, audit and assemble. With this tool, a new candidate is a data change instead.

• WITNESS.json holds inputs only:
  • the pinned sources;
  • the geometry;
  • typed provider build chains (greedy_kernels, promote, joint_prefix, multicut/cold steps, all naming recipes by hash);
  • the dispatch and call order;
  • PLAN and OWNERS, referenced by hash.
    The verifier builds everything from this file.
• EXPECTED.json holds expected values only: built word hashes, control counts, per-provider artifacts and profiles, the global profile and counts, and κ. The verifier only compares against it. No value in EXPECTED.json feeds a computation. Every value is recomputed independently first, and moment endpoints are re-certified rather than read, so a wrong EXPECTED.json can only make verification fail.
• Entry points: spec.py generate|check and verify.py. Python 3.11+, standard library only, assertions required.

The admission scripts in code/ are PR582's, unchanged except that shape constants now come from geometry and moment endpoints are always re-certified.

Tests (tests/run_tests.py, fixture: PR582 at 747ec97)

All 8 pass:

• Fixture: from examples/pr582/WITNESS.json alone, every provider word is built through the typed steps, with no prebuilt words or reused locals. The generated EXPECTED.json equals PR582's published values, and spec.py check passes. A strict verify from scratch reproduces κ 0.001601486955392486.
• Mutations that must fail, each rejected at its named check:
  • a profile count;
  • a dispatch entry;
  • an owner byte;
  • a kernel coefficient.

Not included

These are left out on purpose: witness search (owner solvers, kernel and width screens, orchestration), new step types, probe builds of recipe output hashes, and a spec-driven accounting model. See the README.

Credits

Adapted from the PR582 → PR576 → PR571/PR569/PR568/PR564 lineage and PR575/PR558, with all original notices retained.

AI disclosure and authorship: Claude Opus 5.5 (Anthropic), running in the omp coding harness, wrote this tooling. Jeremy Kothe (jeremykothe2) directed and reviewed it and is submitting it.

[Mister-Meeseeks]

Thank you @jeremykothe2. This contribution at 9bf577d56312f7491e4bb39b5fa7d844f318cde9 is incorporated in the reviewed family through #588, published in 9594b61.

Inputs and expected outputs separated. pin() returns recomputed got; expected values only compare. Cold runs do full physical admission. Hidden reuse flags lack standalone provenance binding and must not be used as acceptance evidence without external source/input/receipt binding. The replay covers the fixed PR582 accounting model; it is not a theorem for arbitrary geometry declarations.

Preserved implementation and research:

• research/witness-spec-tools

The family review and individual disposition record the scope, evidence and attribution. Parallel and intermediate contributions remain credited even where a later composition has a stronger bound. The published multiplication bound remains conditional on the documented compiler, realization, hosting and analytic assumptions; these finite checks are not a full Lean multiplication proof.