Valdris Proof And Handoff

1. Validate the goal stopping conditions and latest checkpoint.
2. Run the always-required goal, context, skill-registry, provenance, neutrality, privacy, schema-compatibility, run-packet, independent-review, and proof gates, plus classification, foundation, code-intelligence, production, AI-assurance, eval, trajectory, smoke, RCA, domain-pack, and authoritative-assurance gates when the adapter, route, and claimed assurance level make them applicable. For clean-room convergence or migration, also run restricted-residue-gate.mjs with the commissioned external manifest across the repository, release archive, generated packs, knowledge vault, and installed-skill manifest; never commit or print restricted inputs. Verify every non-applicability decision rather than assuming every possible gate is unconditional. The executable finish-line commands are run-packet-gate.mjs, review-gate.mjs, and, for semantic or authoritative claims, authoritative-assurance-gate.mjs; bug, regression, incident, or self-heal routes also run rca-gate.mjs. Before independent review, generate the canonical digest with run-create.mjs --print-evidence-bundle; the Ed25519 review must sign it and verify against an active key at the exact committed runtime-derived trust path (controls/review-trust.v1.json in the harness itself or .valdris-harness/controls/review-trust.v1.json in a commissioned target). Never let an agent add or trust its own review or authority key during the run.
3. When classification is required, verify run/workload-classification.json is bound to the authorized intake and route. When foundation is required, verify foundation/assessment.json proves the required capabilities at the effective assurance tier for the selected workload and stakes profiles; a rewritten or downgraded classification cannot make foundation proof disappear.
4. Before closure, use run-create.mjs --print-accepted-gate-set and bind that digest into the executor and authoritative closure. Run assurance-readiness.mjs, then verify artifact paths, hashes, immutable route-derived tiers/applicability, commit/deployment identity, freshness, typed requirement traceability, eval/trajectory/smoke plan coverage, conditional model-judge calibration, tool/call receipts, durable memory heads, runtime-driver and implementation receipts, economics reconciliation, declared interop transcripts, dependency provenance, trace-v2 bytes/trajectory binding, signed runtime conformance, and semantic pass criteria. New packets use valdris.run-packet.v3. Non-HEAD and genuine pre-v3 packets are integrity-inspected as structural evidence only; full revalidation requires an isolated checkout at the exact packet commit. Authoritative claims additionally require operator-pinned approval; fully bound model-routing, trace, usage, memory, and implementation receipts; raw-object OCI execution with commissioned Git/runtime/daemon and secure output-root identities; and a provider-backed monotonic bridge head. For GitHub, require the canonical commissioned hostname on every call, explicit expected sequence/head/history CAS, cumulative checkpoint plus commissioned full-replay attestation, pre-proposal/pre-merge/post-merge protection observations, deterministic resume evidence, recorded cleanup outcome, and a stable operator-controlled receipt root.
5. Reject unjustified skips, stale evidence, unsupported status claims, failed dependencies, agent-granted approvals, and fake self-heal events.
6. Confirm every Red Zone approval was granted by the scoped human owner. Before live completion, require token-gated human approval with scope route and artifact run/route.json; the bridge binds approval to the route digest.
7. For TestFlight/App Store release, require the token-gated grant to name domain/assurance.json, use the predeclared bridgeEventId, and remain bound to the current domain packet digest/build identity.
8. Treat CI/provider trust fields as contracts, not cryptographic proof by themselves. Release claims require imported provider receipts or signed attestations where the commissioned policy demands them.
9. Write handoff/final.md with bottom line, why, proof paths, risks, rollback, skipped controls, self-heal state, and the next human decision.

No proof means not done. A failed required gate means blocked, even when the implementation appears to work.