Problem
Problem
AI output can sound deployment-ready before anyone has separated its claims, assumptions, supporting evidence, and missing checks.
Independent public demonstration
Independent public demonstration
Reliance Compiler and Repository Review Gate are shown together as a public fixture. The scenario is illustrative: it makes evidence gaps visible; it does not claim a combined production system.
Scenario
Scenario
Illustrative fixture only: an AI says a repository is ready to deploy, but the available evidence does not cover a destructive migration or a rollback path. The demonstration asks what a bounded review should return; it does not claim that this combined system ran in production.
Constraints
Constraints
- Use public repositories and a synthetic fixture only.
- Keep observed evidence separate from interpretation and open assumptions.
- Keep deployment authority with a named human decision owner.
- Do not infer readiness from checks that were not exercised.
Design decisions
Design decisions
- Treat a claim as the unit of review instead of assigning confidence to a whole answer.
- Bind each material claim to the evidence that can support or disconfirm it.
- Escalate when destructive migration or rollback evidence is missing.
- Return an explicit bounded decision instead of implying that automation has approval authority.
05 / System map
System map
- Unsafe AI outputStart with the deployment-ready statement as an unverified input.
- Claims and assumptionsSeparate what the statement asserts from the assumptions it leaves unstated.
- Evidence bindingLink each material claim to observed repository evidence or mark the evidence gap.
- Failure modesCheck the destructive migration and rollback boundaries that could change the decision.
- Human decision gateRoute the unresolved reliance decision to the person who owns deployment authority.
- VerificationName the smallest checks that could close the material evidence gaps.
- Reliance receiptRecord the bounded output, evidence scope, and remaining uncertainty for review.
INPUT / FIXTURE
Deployment-readiness claim
- AI-generated claim: “Repository is ready to deploy.”
- Observed evidence: a destructive database migration is referenced, but its safety checks are not shown.
- Observed gap: no verified rollback procedure or test result is attached.
DO NOT RELY / HUMAN_DECISION_REQUIRED — human review is required before any deployment decision.
OUTPUT / BOUNDED
Bounded decision
DO NOT RELY / HUMAN_DECISION_REQUIRED
- Claims and assumptions are listed separately.
- The destructive-migration evidence gap is visible.
- The rollback evidence gap is visible.
- No deployment authority is granted by the fixture.
06 / Tests and public checks
Tests and public checks
- Reliance Compiler fixtures check claim, assumption, and evidence references.
- Repository Review Gate routes unverified findings to a human decision.
- This combined fixture emits the bounded no-reliance decision when migration and rollback evidence are missing.
- The demonstration links to public source repositories and does not execute a deployment.
Limitations
Limitations
- This is an independent public demonstration, not client work.
- The fixture is bounded and illustrative; it does not represent a production deployment.
- The public source links show methods and fixtures, not every context needed for a real release decision.
Remaining risks
Remaining risks
- A different repository may contain migration or rollback behaviour that this fixture does not represent.
- Public source availability and implementation versions can change.
- A human owner still needs to assess context and approve or decline the decision.
What this does not prove
What this does not prove
- That every repository is safe to deploy.
- Production reliability or universal correctness.
- A client engagement or client outcome.
- An exhaustive security certification.
A useful first conversation
Start with non-sensitive context. We can identify the claims, review burden, decision owner, and evidence that would change the next step.