Skip to content

ĀML hardening: independent verification, release discipline, evidence corpus - #162

Closed
aruintelligence wants to merge 10 commits into
mainfrom
hardening/independent-release-20260914
Closed

aruintelligence wants to merge 10 commits into
mainfrom
hardening/independent-release-20260914

Conversation

@aruintelligence

@aruintelligence aruintelligence commented Sep 14, 2026

Copy link
Copy Markdown
Owner

Purpose

Consolidate the next ĀML hardening wave around independent verification, immutable release identity, and disciplined real-screen evidence.

Changes

  • Moves verifier challenge cases into conformance/verifier-challenge-cases.json, a language-neutral corpus using a narrow RFC 6901 JSON Pointer replacement profile.
  • Updates the black-box harness to consume that corpus while preserving the newer exact challenge/witness-vector SHA-256 bindings.
  • Adds the case corpus and its SHA-256 to the reference-code-free External Verifier Kit without removing Snapshot 2/catalog/migration protections already on main.
  • Updates regression tests so archived conformance evidence still verifies against the exact challenge contract while case definitions remain readable without inspecting canonical harness source.
  • Adds docs/RELEASE_DISCIPLINE.md to separate stable, preview, and experimental states and define minimum release evidence.
  • Adds docs/REAL_SCREEN_EVIDENCE_CONTRACT.md to standardize capture provenance, labels, policy binding, receipts, reruns, before/after evidence, and claim boundaries.
  • Adds publications/INDEPENDENT_IMPLEMENTATION_CHALLENGE.md to invite outside clean-room implementations with PASS / FAIL / MIXED reporting.

Evidence boundary

A project-authored test or GitHub Actions result remains project-controlled evidence. A black-box PASS is compatibility evidence for the exact tested contract, not certification, proof of implementation independence, standards approval, scientific validation, regulatory compliance, safety, or adoption.

The attention/restoration labels used in real-screen experiments remain declared reviewer inputs, not objective measurements of cognition or wellbeing.

Relationship to older PRs

This PR modernizes the useful language-neutral verifier-case idea from #68 against current main rather than merging that stale branch wholesale. The real-screen evidence from #152 is already present in current main; this PR adds a forward evidence contract rather than overwriting those historical packages.

Copy link
Copy Markdown
Owner Author

Superseded by #166. main advanced substantially while this branch was under test, so the hardening changes were replayed as one clean commit on top of the newer main instead of forcing a stale merge. No evidence or intended functionality from this branch is being treated as independently validated merely because it was replayed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant