idkmesh

Sequential evidence reproducibility hardening

Date: 2026-08-28

After PR #206 merged, the new Sequential Evidence Kernel passed on canonical main under both Python 3.11 and 3.13. The GitHub Actions artifact archives, however, had different archive-level digests even though the workflow itself was green.

That exposed an evidence-quality issue: an artifact archive digest can reflect packaging/container details and should not be used as a direct proof that the mathematical payload itself is byte-identical across interpreters.

Hardening decision

Replace the two independent matrix jobs with one read-only job that performs both interpreter runs and then compares the actual generated JSON payloads directly:

Python 3.11 -> compile -> tests -> demo-3.11.json
Python 3.13 -> compile -> tests -> demo-3.13.json
                         |
                         v
                   byte-for-byte cmp
                         |
                         v
                 explicit SHA-256 content hash

The job fails if the two JSON files differ at all. It then writes a SHA-256 over the content itself and retains both payloads plus the hash in one artifact.

This makes the reproducibility claim about the mathematical output, not about GitHub’s archive packaging.

Authority boundary

This hardening does not change the algorithm, confidence thresholds, governance decisions, repository permissions, or integration authority. The workflow remains:

permissions: contents: read

main remains unprotected and that continues to be a separate hard governance boundary.