Exact-artifact closure evidence¶

The local release has one current certification chain for gates G-01, G-02, G-04, G-05, G-08, G-09, G-14, G-15, G-17, G-26, G-30, G-32, G-34, G-35, and G-37. The chain accepts no editable install, artifact discovery, legacy manifest, fallback binary, skipped case, or mock live evidence.

Generate the campaign manifest¶

First promote the exact local release and independently check its signed promotion evidence. Set EXACT_LOCAL_EVIDENCE_VERIFIER_COMMAND to a JSON argv array naming the deployment-owned verifier. Then generate the campaign input:

generate-exact-local-gates-manifest \
  --release-id "${RELEASE_ID}" \
  --spec "${RELEASE_SPEC}" \
  --promotion-evidence "${PROMOTION_EVIDENCE}" \
  --source-root "${AGENT_UTILITIES_SOURCE_ROOT}" \
  --output "${PRIVATE_EVIDENCE_ROOT}/exact-local-gates-manifest.json"

The source root is used only to hash the current certification harness and its fixed test catalog. It is never copied into the manifest. The generator checks the promotion signature and release specification, requires successful zero-process pre/post gates plus passing canary and doctor checks, cross-checks the engine identity against the native artifact digest, and emits the strict deploy/release/exact-local-gates-manifest.schema.json shape. The output binds the promotion evidence, release specification, installed Agent Utilities tree, installed distribution closure, release Python, GraphOS launcher, full engine, campaign harness, and test catalog.

Run scripts/certification/exact_local_gates.py with that manifest and its independently calculated SHA-256. The campaign evidence now carries the promotion-evidence and release-specification bindings in addition to the exact runtime artifacts.

Produce the exact engine campaign set¶

Deployment resolves the following explicit values from AgentConfig and the exact release inputs. The producer does not inspect ambient configuration, discover a binary, build a binary, or fall back to another Python or source tree.

python scripts/certification/run_exact_engine_campaigns.py \
  --release-id "${RELEASE_ID}" \
  --engine-binary "${ENGINE_BINARY}" \
  --engine-sha256 "${ENGINE_SHA256}" \
  --campaign-python "${RELEASE_PYTHON}" \
  --campaign-python-sha256 "${RELEASE_PYTHON_SHA256}" \
  --epistemic-graph-root "${EPISTEMIC_GRAPH_SOURCE_ROOT}" \
  --source-freeze-evidence "${SOURCE_FREEZE_EVIDENCE}" \
  --source-freeze-sha256 "${SOURCE_FREEZE_EVIDENCE_SHA256}" \
  --authority-config "${G37_AUTHORITY_CONFIG}" \
  --work-root "${PRIVATE_LINUX_WORK_ROOT}" \
  --output-dir "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns"

The work root, authority file, and output parent must be caller-owned private Linux locations. The output directory must not exist. The producer checks and stages the one digest-pinned full engine once, checks the release Python and source-frozen Epistemic Graph producer tree, and invokes every child with argv and a minimal isolated environment. It runs G-37 performance first, followed by G-02/G-05 fault and restart, G-01/G-04 protocol authorization, G-14 multimodal, G-15 KnowledgeBatch, and G-17 reasoning and repair. G-14 receives the SHA-256 of the exact performance.json bytes produced by G-37.

A successful final directory contains exactly these closure-ready documents:

performance.json
fault-restart.json
protocol-authorization.json
multimodal.json
knowledge-batch.json
reasoning-repair.json

Each document is semantically validated by the closure validators before the directory is atomically published. Timeouts, non-zero exits, output overflow, artifact or source mutation, malformed evidence, extra output, and any failed campaign remove the unpublished staging directory and fail closed. Child output is drained but not retained. Temporary engine, environment, and Markdown artifacts are removed; the six aggregate JSON documents contain no local paths, environment values, payloads, or personal identifiers.

Run the manifest-bound exact-local campaign separately, using the same exact release engine identity, before binding the complete closure. Then set these two environment variables to deployment-owned JSON argv arrays:

EXACT_ARTIFACT_CLOSURE_SIGNER_COMMAND
EXACT_ARTIFACT_CLOSURE_VERIFIER_COMMAND

Bind every current live campaign¶

The signer and verifier command variables are the only environment-provided inputs to the binder.

The signer receives bounded canonical unsigned JSON on standard input and returns algorithm, opaque keyId, encoded signature, and subjectDigest. The verifier receives the complete signed document and must return exactly verified=true, the same subjectDigest, and the same opaque keyId. Commands, keys, and credentials never enter arguments or evidence.

bind-exact-local-release-evidence \
  --release-id "${RELEASE_ID}" \
  --spec "${RELEASE_SPEC}" \
  --promotion-evidence "${PROMOTION_EVIDENCE}" \
  --source-root "${AGENT_UTILITIES_SOURCE_ROOT}" \
  --campaign-manifest "${EXACT_LOCAL_GATES_MANIFEST}" \
  --fault-restart-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/fault-restart.json" \
  --protocol-authorization-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/protocol-authorization.json" \
  --performance-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/performance.json" \
  --multimodal-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/multimodal.json" \
  --knowledge-batch-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/knowledge-batch.json" \
  --reasoning-repair-evidence "${PRIVATE_EVIDENCE_ROOT}/engine-campaigns/reasoning-repair.json" \
  --exact-local-evidence "${EXACT_LOCAL_EVIDENCE}" \
  --output "${PRIVATE_EVIDENCE_ROOT}/exact-artifact-closure.json"

The binder independently re-checks promotion evidence and regenerates the expected campaign manifest. It rejects any release, specification, harness, catalog, Agent Utilities, GraphOS, engine, or campaign digest mismatch. It also requires the complete 60-case G-02/G-05 crash matrix and both G-05 restart proofs; the exact G-01/G-04 identity and authorization slice, including denial across all 14 served data paths and ten enabled wire protocols; 30 G-37 scenario families and all 54 ledger rows; four G-14 modalities, 12 behavior dimensions, 16 fault cases, and the exact G-37 digest; all seven G-15 KnowledgeBatch families, requirements, and snapshot cases; all nine G-17 restart, retraction, causal, and fenced-repair cases; and passing G-08 WorkItem lifecycle, local G-09 inbox-to-WorkItem crash replay, G-26/G-30/ G-32/G-34, and G-35 permission-governance details from cryptographically checked exact-local evidence. The closure exposes those shared exact-local bytes as the separately digest-bound workItemAgentBus, exactLocal, and permissionGovernance campaign summaries.

The resulting document conforms to deploy/release/exact-artifact-closure-evidence.schema.json. It contains only opaque release identifiers, SHA-256 bindings, fixed aggregate counts, gate status, and external signature material. It contains no paths, endpoints, credentials, personal identifiers, payloads, prompts, traces, logs, or raw campaign results. Both output destinations must be new absolute files under an existing caller-owned private directory.

Run the source-only guard independently:

python scripts/check_exact_artifact_closure.py

This source guard validates the schemas, exact gate and test inventories, console entry points, documentation, and current manifest agreement. It does not run or replace any exact-artifact campaign.