Targeted implementation tests
10 / 10 PASS
The current public-repair tranche passed its targeted cases on the implementation branch.
TARGETED REGRESSION
False authority / proof custody / bounded repair
AI Integrity, Repair & Reentry
A protection layer for separating what an AI system claimed from what can be proved, repairing bounded failures, and controlling reentry after a bad state.
IMPLEMENTATION BRANCH + FULL CI EVIDENCE / RELEASE HELD
01 / PROBLEM
A convincing repair is not enough if the repair can mint its own authority.
The system is designed around a stricter question: can a failure be detected, documented, repaired and re-entered without letting the repair layer silently become the owner of truth, permission or the next consequential action?
What the system does
The architecture preserves claim state, source/proof demand, blocked use, repair provenance, regression evidence and human-required boundaries as separate objects.
The public repair satellite is intentionally advisory: it can expose a mismatch and propose bounded repair without granting itself the authority whose misuse it is diagnosing.
02 / CURRENT EVIDENCE
What has actually been demonstrated.
Each item below is deliberately narrower than a product-readiness or superiority claim.
Post-commit validation
Full repository + Docker green
A later CI run completed successfully with full-repository pytest and Docker validation green, security gates green, and the artifact bound to the run.
CI / DOCKER / SECURITY / ARTIFACT
Process integrity
Sequence defect retained
A required full-suite-before-commit ordering step was missed. Green CI did not erase that process defect; it remains part of the evidence state.
PROCESS RECEIPT / NO GREEN-LAUNDERING
Human-required boundary
Repair cannot self-promote
The design keeps authority and consequential release outside the repair satellite even when detection/repair checks pass.
HUMAN AUTHORITY / BOUNDED ADVISORY
03 / CLAIM CEILING
What this page does not claim.
It does not claim universal hallucination prevention, prompt-injection immunity, compliance, provider superiority, finished production enforcement, or product-ready release.
Current development
The current path is independent PR-state/patch review, correct stage promotion, and later family/runtime validation without letting successful repair tests substitute for release authority.