# Core Architecture v2 Conformance

**Role:** claim-to-evidence boundary map. The normative target is [Core Architecture v2](CORE_ARCHITECTURE_V2.md); current implementation gaps and next gates are in [Core Architecture v2 Status](CORE_ARCHITECTURE_V2_STATUS.md). Historical incidents and checkpoint records remain in the [Engineering Ledger](ENGINEERING_LEDGER.md).

## Verification boundary

Baseline conformance observations came from the read-only architecture audit at commit f1518caf405f44851227cf73ea37b79267b1a098 and were not rerun during that audit. S0 changed documentation only. The deterministic S1, S2, and S3 gates were accepted earlier in the campaign; the S5 context/session gate was independently reviewed and accepted on 2026-09-24 as recorded below. S6 was independently accepted on its deterministic gate on 2026-09-24; the evidence lanes below remain distinct from provider certification. Other audit observations remain at their stated evidence boundary. The Campaign Director independently accepted the S7 deterministic gate on 2026-09-24 after reviewing the schema-condition remediation and rerunning focused, typecheck, build, full non-system, and S7 deterministic system gates. The S8 deterministic slice gate and bounded CLI-surface remediation were independently reviewed and accepted on 2026-09-25. S9 deterministic source, system, packed provider-fault/restart, and prior fixture-browser evidence is recorded below. The R12 bootstrap failure-finalization repair has focused deterministic evidence and a fresh packed inspection; the actual durable-operation browser choice/resume/replay/cancel journey and REAL_PROVIDER lifecycle certification remain pending. S10 adds real rootless bwrap isolation execution and real local OSS Trivy SAST evidence, recorded in its own section below; those are local real-provider executions and do not certify hosted providers, rootless Podman/OCI, OPA, or S13 certification.

The source and test references below identify the evidence inspected by that audit and the S1 revalidation recorded below. CONTRACT, UNIT, and INTEGRATION evidence does not imply REAL_PROVIDER, BROWSER, ADVERSARIAL, or PACKED_E2E certification. Those lanes are reported separately below. Historical test counts in the Engineering Ledger remain dated checkpoint evidence and are not current certification claims.

The PR #64 Core Architecture v2 CI alignment inventory is maintained in [CORE_ARCHITECTURE_V2_CI_ALIGNMENT.md](CORE_ARCHITECTURE_V2_CI_ALIGNMENT.md). Its deterministic, adversarial, human, rootless, provider, full-stack, packaged-consumer, Pact, and supply-chain lanes are assigned to their owning environments; full-stack and supply-chain jobs upload their machine-readable evidence, excluding private signing material. Historical checkpoint 325605/run 368030 passed 11 jobs and failed packaged-consumer/rootless-isolation; pre-amendment run 368171 passed 12/13 with P3 packaged-consumer still failing; intermediate run 368223 passed 12/13 with the F5 deterministic detection fixture timeout. The implementation/workflow checkpoint d21d6b5 passed all 13 jobs in run [36824682741](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36824682741). The later status-doc candidate 963cc also passed all 13 jobs in run [36832161124](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36832161124). Exact-SHA check results for this and any later evidence snapshot are attached to PR #64. These CI results do not alter the normative target. S14 and final closure acceptance are recorded by the independent review artifacts below; fresh independent review (Round 16) accepted status-doc candidate 963cc after its 13/13 result in run 36832161124.


Intermediate PR #64 CI evidence (2026-10-01): head `ff2b9f279fcf6c974f626875dc99248881613a79`, run 36822314115 on PR merge SHA `64e3f963f7bf70e287da13f0f9aefb13e26c0450`, 12/13 jobs passing. The required test/release lane timed out at F5 in `tests/security/isolation.test.ts` while ambient host-provider detection was used; the dedicated rootless and packaged-consumer lanes passed. A local fixture isolated PATH/root and passed 10/10 in 417ms; hosted run 36824682741 later verified that repair on implementation checkpoint d21d6b5. This intermediate CI result did not change TARGET conformance claims.
Last green implementation checkpoint CI evidence for PR #64 (2026-10-01, before this documentation-only amendment): head `d21d6b5db6b36b8387af7687d9488c4f95229de1`; run [36824682741](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36824682741) completed all 13 required jobs successfully. This confirms the deterministic test/release lane and every dedicated environment/provider lane, including human, adversarial, rootless, packaged-consumer, full-stack, supply-chain, Pact, and OCI. CI evidence remains separate from architecture certification. The separate S14 and closure acceptance records are `docs/evidence/s14/review-s14.json` and `docs/evidence/closure/review-closure.json`, both with `SLICE_ACCEPTED`; see the final closure conformance record below. The later status-doc candidate 963cc passed 13/13 in run 36832161124; fresh independent review (Round 16) accepted status-doc candidate 963cc after its 13/13 result in run 36832161124.

On 2026-09-24, the S1 callback-mutation follow-up reran deterministic focused evidence, `npx tsc --noEmit`, `npm run build`, and the full non-system Vitest suite after integration. Candidate-truth and controller-fencing prerequisites passed 2 files / 12 tests; the focused S1 set passed 15 files / 160 tests; TypeScript exited 0; build exited 0; the full non-system suite passed 149 files with 1 skipped, 768 tests with 7 skipped. The full suite emitted missing-source-map notices from its packed build-hygiene test, which passed. These are CONTRACT/UNIT and build evidence only. The launch transports were captured at deterministic adapter boundaries; no REAL_PROVIDER, container, browser, adversarial, or complete packed-journey certification is inferred.

The same-slice execution-semantics/revision remediation then reran `npx vitest run tests/operations.test.ts` — 1 file / 33 tests passed; the focused S1 set — 15 files / 165 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; and `npx vitest run --exclude 'tests/system/**'` — 149 files passed, 1 skipped; 773 tests passed, 7 skipped. The generic future-digest preemption and absent-revision upgrade cases, including the reproduced absent-to-revision-1 patch, now reject without changing durable bytes; an owning changed-digest bind increments execution revision once and invalidates frozen policy and participant bindings. These results are deterministic contract/unit and build evidence only; no REAL_PROVIDER, browser, adversarial, container, or packed end-to-end certification is inferred.

The earlier S2 checkpoint that reported a missing production composition is superseded by the user-approved AgentTopology-to-Paseo architecture decision and the integrated implementation recorded here. On 2026-09-24, S2 deterministic verification reran `npx vitest run tests/auditIntent.test.ts tests/triage.test.ts tests/recovery.test.ts tests/candidateAssembler.test.ts tests/candidateWaveLineage.test.ts tests/candidateRepairProtection.test.ts tests/issueIntake.test.ts tests/issueDeliveryIdentity.test.ts tests/participantsStackDiscovery.test.ts tests/participantsCanonical.test.ts tests/semanticAssessment.test.ts tests/semanticAssessmentRuntime.test.ts tests/semanticAssessorTopology.test.ts tests/runRepairCandidateLifecycle.test.ts` — 14 files / 108 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0; and `npx vitest run --exclude 'tests/system/**'` — 151 files passed, 1 skipped; 788 tests passed, 7 skipped. The full suite emitted missing-source-map notices from build-hygiene checks, which passed. These deterministic fixtures verify the named public paths, typed runtime, and safeguards; they do not establish REAL_PROVIDER execution or certification.

The S2 repository-boundary and INTENT remediation was independently verified on 2026-09-24. The focused S2 command listed above passed 14 files / 123 tests; a separate stack/semantic runtime set passed 5 files / 75 tests; `npx tsc --noEmit` exited 0; `npm run build` exited 0; and the full `npx vitest run --exclude 'tests/system/**'` lane passed 151 files with 1 skipped and 803 tests with 7 skipped. The regressions bind STACK to canonical realpath and current repository content before/after capture and inference, exclude ignored `.env` and `.harness` runtime/cache evidence, reject invalid or escaping proposed project skill roots, and accept bounded HIGH-risk INTENT while retaining deterministic assurance floors. These are deterministic contract/unit/build results only. No REAL_PROVIDER, BROWSER, or ADVERSARIAL certification is claimed.

On 2026-09-24, final S3 deterministic verification after the Lead acknowledgement fencing change ran the focused authority/action command (`npx vitest run tests/security/toolActionGate.test.ts tests/humanDecision.test.ts tests/security/controllerFencing.test.ts tests/operations.test.ts tests/controlCenter.test.ts tests/architectureIntegrationContracts.test.ts tests/security/actionReconciliation.test.ts tests/deliveryFinalize.test.ts tests/operationCompletion.test.ts tests/agentPromptStructuredResultProvenance.test.ts`) — 10 files / 146 tests passed; `npx tsc -p tsconfig.json --noEmit` — exit 0; `git diff --check` — clean; `npm run build` — exit 0 (`release-1790277810343-2032089-c2f3666b`). The final full non-system command (`mapfile -t test_files < <(rg --files tests -g '*.test.ts' | rg -v '^tests/system/'); npx vitest run "${test_files[@]}" --reporter=json --outputFile=/tmp/s3-full-vitest-final.json`) completed with 318/318 suites successful, 810 tests passed, 7 pending, and zero failed. The final `npx vitest run tests/system` ran 4 files: adversarial, concurrency, and scenario-matrix suites passed 45 tests; overall 63 passed and 1 failed (64 total, no skipped). The failure is `aehHumanBlackBox.test.ts:84`: its natural-language `intent` CLI assertion received empty stdout because Paseo preflight exited with `SEMANTIC_ASSESSMENT_UNAVAILABLE`; configured model `opencode/opencode-go/MiMo-V2.6-Flash` is absent from the local catalog. Direct reproduction confirmed this environment/provider mismatch; it is not S3 evidence. The `opa` executable is unavailable, so OPA evidence is source/adapter-level only. These are CONTRACT, UNIT, build, and SYSTEM_DETERMINISTIC results; they do not establish REAL_PROVIDER, BROWSER, packed adversarial, or cancellation/takeover certification.


The preceding S3 entry is a pre-remediation checkpoint and is superseded for deterministic S3 closure by this product-choice continuation remediation. On 2026-09-24, the focused continuation command (npx vitest run tests/controlCenter.test.ts tests/humanDecision.test.ts tests/operations.test.ts tests/changeHandoff.test.ts) passed 4 files / 55 tests. The expanded authority/action/continuation command (npx vitest run tests/security/toolActionGate.test.ts tests/humanDecision.test.ts tests/security/controllerFencing.test.ts tests/operations.test.ts tests/controlCenter.test.ts tests/architectureIntegrationContracts.test.ts tests/security/actionReconciliation.test.ts tests/deliveryFinalize.test.ts tests/operationCompletion.test.ts tests/agentPromptStructuredResultProvenance.test.ts tests/changeHandoff.test.ts) passed 11 files / 153 tests. npx tsc -p tsconfig.json --noEmit exited 0; git diff --check was clean; npm run build exited 0 and produced release-1790282947173-2145162-1fdb29fd with the Control Center UI bundled. The full npx vitest run --exclude 'tests/system/**' --reporter=json --outputFile=/tmp/s3-remediation-final-vitest.json passed all 318 suites, with 814 tests passed, 7 pending, and zero failed; it emitted missing-source-map notices from the packed build-hygiene test. npx vitest run tests/system/aehAdversarialE2E.test.ts tests/system/aehConcurrencyCampaign.test.ts tests/system/aehScenarioMatrix.test.ts passed 3 files / 45 tests. These runs provide deterministic contract, unit, build, and system evidence for the S3 gate. The shared-checkout UI API mock smoke (npx tsx /tmp/s3-shared-ui-smoke.mts) passed pairing, strict DecisionRequest parsing, stale 409 handling, CSRF and submission-field checks, accepted response, and refresh behavior. The initial full-suite attempt found the ignored ui/control-center/dist output populated by the standalone UI build; that generated output was moved outside the checkout and the full suite then passed. No live OPA, browser, real-provider, packed adversarial, or provider cancellation/takeover certification is claimed.

On 2026-09-24, the S5 manifest-to-grant remediation first reproduced the old issuer defect, then passed the same authorization regression suite. The Campaign Director independently reviewed the complete S5 diff, exact manifest/envelope/ref-set checks, focused regressions, and slice gate; reran the focused context/session/provenance command (12 files / 125 tests), `npx tsc -p tsconfig.json --noEmit` (exit 0), `npm run build` (exit 0; `release-1790294205611-17-b4fe25cb`), the full non-system suite (152 files passed, 1 skipped; 835 tests passed, 7 skipped), and the three relevant deterministic system suites (3 files / 45 tests passed). The first full-suite attempt was blocked by sandbox `EPERM` errors on loopback listeners and TSX IPC; the same gate passed when rerun with the required local process/loopback permissions. `git diff --check` was clean, and the normative target had no diff. S5 is accepted on its deterministic gate. These are CONTRACT, UNIT, INTEGRATION, BUILD, and SYSTEM_DETERMINISTIC evidence only; they do not establish REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E certification.

On 2026-09-24, S6 deterministic verification ran the focused acceptance/completion/lifecycle set (`npx vitest run tests/objectiveCompletion.test.ts tests/objectiveCompletionLifecycle.test.ts tests/acceptanceOracle.test.ts tests/managedLeadAcceptance.test.ts tests/candidateAssurance.test.ts tests/runRepairCandidateLifecycle.test.ts tests/reviewLifecycle.test.ts tests/deliveryFinalize.test.ts tests/operationV2Identity.test.ts tests/operationLifecycleRegression.test.ts tests/operationSupervisionArchitecture.test.ts`) — 11 files / 124 tests passed; `npx tsc --noEmit` — exit 0; `npm run build` — exit 0 (`release-1790303504564-2721309-64c0f9cc`); full non-system Vitest — 157 files passed, 1 skipped; 931 tests passed, 7 skipped; relevant lifecycle scenario/concurrency system tests — 2 files / 17 tests passed; and `git diff --check` — clean. Lifecycle fixtures distinguish read-only audit terminalization from managed run/change success, which requires the current persisted AcceptanceOracle disposition and complete objective evidence. These are deterministic contract, unit, integration, build, and system results only; no REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or complete PACKED_E2E certification is claimed. The S6-W3 contract omits both certification disposition and identity when certification is not required; the evaluator rejects supplied values. The normative target was not changed.

The Campaign Director independently reviewed the final S6 source/test/documentation diff, reran the focused S6 set (11 files / 124 tests), `npx tsc --noEmit` (exit 0), `npm run build` (exit 0; `release-1790305046220-17-0bfdef32`), full non-system Vitest (157 files passed, 1 skipped; 931 passed, 7 skipped), and the relevant system scenario/concurrency set (2 files / 17 tests passed). The documentation-only remediation then passed `git diff --check` and confirmed no normative target diff. S6 is accepted on the deterministic gate; no REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or complete PACKED_E2E certification is claimed.

The initial 2026-09-25 S9 attempt ran `npx tsc -p tsconfig.json --noEmit` — exit 0 — and the focused runtime/Control Center command — 10 files / 101 tests passed. A separate source-importing `tests/system/aehConcurrencyCampaign.test.ts` passed 1 file / 9 tests. `npm run ui:build` and `npm run build` exited 0 and produced `release-1790327373114-3350119-66a118bd`. These initial results are retained as historical evidence; the system campaign imports source and is not a packed E2E claim.

The final S9 same-slice remediation reran `./node_modules/.bin/vitest run tests/paseoProviderLeaseLifecycle.test.ts tests/paseoInitialTurnBarrier.test.ts tests/runtimeSupervisorV2.test.ts tests/processCleanup.test.ts tests/security/controllerFencing.test.ts tests/operations.test.ts tests/operationLifecycleRegression.test.ts tests/operationStateDurability.test.ts tests/controlCenterOperationProjection.test.ts tests/controlCenter.test.ts tests/controlCenterApiContracts.test.ts tests/humanDecision.test.ts tests/paseoStart.test.ts` — 13 files / 137 tests passed — followed by `npm run typecheck` (exit 0). `npm run ui:build` and `npm run build` exited 0; direct build produced `release-1790334556192-3490030-97632a32`. The final `npm pack` ran its prepare/build lane successfully. The source-importing `./node_modules/.bin/vitest run tests/system/aehConcurrencyCampaign.test.ts` passed 1 file / 9 tests. These source, system, and build results are separate from the packed lane.

Final packed provider lifecycle evidence used `/tmp/s9-packed-final4/agentic-engineering-harness-0.8.4.tgz`, extracted to `/tmp/s9-packed-fixture-final4/package`; `node tests/packed/s9ProviderLifecycleCampaign.mjs /tmp/s9-packed-fixture-final4/package` passed: a child provider owner was terminated with SIGKILL; epoch 1-to-2 takeover inspected the exact prior session, stopped it, observed idle and resumed; injected stop failure kept cancellation nonterminal with the lease retained; a refreshed scoped cancellation then stopped/re-inspected the exact session, released the old lease, and terminalized CANCELLED. The harness removed its temporary operation roots and verifies clean lease state. Its provider state/inspect/stop calls are deterministic stubs, so this is PACKED_E2E lifecycle-boundary evidence and not REAL_PROVIDER certification. The final tarball digest is included in the S9 handoff report.

The 2026-09-25 browser fixture journey used agent-browser 0.27.0 with Google Chrome for Testing 148.0.7778.97 against the locally built loopback UI. It paired through the single-use URL, showed a current typed choice and blocked stage, submitted the decision to the paired `/api/v1/decisions` route (HTTP 200) and showed the accepted/refreshed state, displayed service and provider-lease rows in Project Control Center, and posted cancellation to `/api/v1/operations/S9-BROWSER-CANCEL/cancel` (HTTP 200). The fixture callback accepted the submitted values; this is BROWSER_FIXTURE evidence, not a durable operation-resume or REAL_PROVIDER result. `paseo --version` returned 0.9.1 and `serena --version` returned Serena 1.6.1, but no AEH provider operation was launched under the S9 checkout restriction. No packed provider-fault/cancellation-takeover certification or REAL_PROVIDER lifecycle claim was run.

The Campaign Director independently reran the S9 focused command — 13 files / 137 tests passed after granting local loopback permission (the sandboxed attempt failed with `listen EPERM`); `npm run typecheck`, `npm run ui:build`, and `npm run build` all exited 0. The new build was release `release-1790335428237-17-96aae8a4`. A fresh `npm pack` produced a package whose `dist/current` is `release-1790335491733-35-b0f6a84b`, tarball SHA-256 `6727f2f0f9bf22339ce28d67aa89a22232c78906d7510ff53e95cc08524adeb3`; the packed lifecycle campaign passed takeover, injected stop-fault fencing, and restart cancellation cleanup. The packed provider inspector/stopper are deterministic stubs and do not establish REAL_PROVIDER evidence. The Campaign Director's Paseo browser bridge returned `No browser automation host is connected.` at that checkpoint, and S9 was then marked **SLICE_BLOCKED**. That blocker is superseded by provisioned real Playwright Chromium; S9 is now open pending the same durable-operation choice, one-time consumption, continuation resume/revalidation, refreshed projection, and paired scoped cancellation journey. No S10 work was started.

At the original remediation checkpoint, `mcp__paseo__browser_list_tabs({})` returned exactly `No browser automation host is connected.` and the Director had prohibited installing agent-browser through npx/network, so no browser tool was installed or invoked then. The persistent elevated shell later provisioned real Playwright Chromium and verified Control Center shell/health; this is readiness evidence only. No durable-operation browser journey has yet been completed; prior BROWSER_FIXTURE results remain fixture-only. Deterministic paired HTTP decision/replay tests, durable continuation-consumption/revalidation tests, and packed scoped cancellation are reported independently and do not substitute for browser evidence.

## Deterministic evidence map

This section records deterministic contracts and source/test references only. It does not promote them into provider or product-journey certification.

| Capability | Deterministic evidence referenced by the audit | Boundary or result |
| --- | --- | --- |
| Candidate identity and assembly | src/candidates/; tests/candidateIdentity.test.ts, tests/candidateWaveLineage.test.ts, tests/candidateRepairProtection.test.ts | Candidate assembly, frozen wave bases, explicit rebases, repair protection, and workspace checks are the strongest completed deterministic subsystem. Downstream impact and assurance consumers remain open. |
| WorkGraph and resource claims | src/architecture/workGraph.ts, src/agents/parallelism.ts; tests/architectureWorkGraph.test.ts, tests/parallelism.test.ts | DAG, scope, logical claims, and scheduling checks are present in the delegated path. Full operation-policy integration remains open. |
| Controller fencing and action gates | src/operations/state.ts, src/operations/controller.ts, src/security/toolActionGate.ts, src/security/actionReconciliation.ts; tests/security/controllerFencing.test.ts, tests/security/toolActionGate.test.ts, tests/security/actionReconciliation.test.ts, tests/operations.test.ts, tests/system/aehConcurrencyCampaign.test.ts | Controller-owned mutations and active terminal transitions require current epoch plus controller token; the explicit Lead acknowledgement binds current Lead actor, epoch, and operation revision. Cancellation consumes a scoped HumanDecision before takeover, rechecks current operation/candidate/policy/revision/epoch before cleanup, and blocks CANCELLED while action intents remain unresolved. Sensitive effects validate frozen policy, actor, scope, lease, candidate, and epoch; required action authorization binds exact effect digest. S9 packed provider-fault/restart lifecycle passed; packed adversarial public-action/effect certification remains open. |
| Runtime service and provider lease lifecycle | src/runtime/supervisorV2.ts, src/runtime/managed.ts, src/runtime/providerLifecycle.ts, src/paseo/runtimeCore.ts, src/operations/controller.ts; tests/runtimeSupervisorV2.test.ts, tests/paseoProviderLeaseLifecycle.test.ts, tests/operationLifecycleRegression.test.ts, tests/packed/s9ProviderLifecycleCampaign.mjs | Shared snapshots reload under a root-scoped lock and persist atomically. Production operation-owned Paseo SDK materialization and participant/Lead turns now consume the managed lease API; current operation, candidate, execution revision, frozen policy, token/epoch, and participant or Lead-generation identity remain checked. Calls renew while pending, preserve uncertain lifecycle state, stop/inspect the exact session, and release only after observed quiescence. Packed release testing killed a child owner, verified epoch takeover, injected a provider stop failure that retained the lease and RUNNING state, then proved exact-session stop/idle observation and lease cleanup before CANCELLED. Provider calls were deterministic stubs; no live provider was used or certified. |
| Control Center operation projection and controls | src/control-center/operationProjection.ts, src/entry.ts, src/control-center/decision.ts, ui/control-center/src/api.ts, ui/control-center/src/App.tsx; tests/controlCenterOperationProjection.test.ts, tests/controlCenter.test.ts, tests/controlCenterApiContracts.test.ts, tests/humanDecision.test.ts, tests/operations.test.ts | Operation, stage, blocker, participant, candidate, service, and provider-lease projections come from durable/current typed sources. Waiting product choices are projected only when request, continuation, candidate, execution revision, frozen policy digest, and epoch all match. Product decisions and cancellation stay server-routed through paired HumanDecision/current-controller checks; UI does not mutate operation state. Deterministic paired-route, persistence, replay, continuation-consumption, and revalidation tests pass. The earlier BROWSER_FIXTURE journey covered fixture callback choice and cancellation; durable-operation browser resume/revalidation remains open because no Paseo browser host was connected during remediation. Packed provider cancellation does not certify packed product-choice interaction. |
| HumanDecision contract | src/security/humanDecision.ts, src/control-center/decision.ts, src/security/toolActionGate.ts, src/operations/state.ts, src/operations/change.ts; tests/humanDecision.test.ts, tests/controlCenter.test.ts, tests/changeHandoff.test.ts, tests/operations.test.ts, tests/security/toolActionGate.test.ts, tests/security/controllerFencing.test.ts | V2 decisions bind the current operation/candidate/execution-revision/policy/epoch identity, paired external actor, typed purpose/scope, expiry, and one-time consumption. Product choice, action authorization, and operation control are disjoint. The production Spec Manager consumer persists a typed request and continuation/checkpoint before HUMAN_REQUIRED, consumes only the exact request-bound choice, revalidates the continuation, and resumes the saved SPEC_AUTHORING stage. The consumed continuation retains its original full decision binding; the checkpoint envelope must match both the continuation and current operation across full candidate, execution revision, policy digest, and controller epoch. Revised requirements permit only one execution-revision advance and a same-epoch current-policy checkpoint rewrite; stale consumed choices are never reused: after exact old-receipt and unchanged candidate/revision checks, the controller opens a fresh current-binding HUMAN_REQUIRED request; intact unanswered requests with unchanged candidate/revision can be reissued after policy/epoch drift. A product choice grants no unrelated authority. Deterministic paired-route tests prove actor derivation and replay rejection, and operation tests prove continuation consumption/revalidation. Durable browser revalidation is still open; no Paseo browser host was connected. |
| OPA canonical role matching | policies/core/trust-boundary.rego; tests/architectureIntegrationContracts.test.ts | Trust-boundary predicates match exact canonical `Implementer` and `Reviewer` spellings; adapter tests compare every policy role literal with the canonical role catalog. `opa` is unavailable, so no live Rego evaluation or provider certification is claimed. |
| Rootless isolation provider | src/security/isolation.ts; tests/security/isolation.test.ts, tests/security/validatorIsolationFailClosed.test.ts, tests/security/providerFailClosed.test.ts, tests/system/aehIsolationCampaign.test.ts | Deterministic capability detection and hardened bwrap argument construction; real execution of user/mount/PID/UTS/IPC/network namespaces; minimal read-only host root with explicit toolchain binds, workspace-only writes, and `--clearenv` allowlist; network denied by default with a real `ENETUNREACH` connect failure; validator commands run isolated with repository mutation and egress denied. Missing/unsupported providers return explicit `ISOLATION_PROVIDER_UNAVAILABLE`/`ISOLATION_PROVIDER_UNSUPPORTED` blockers and never run the validator or report a silent SKIP/PASS. Rootless Podman/OCI execution is not provisioned here. |
| Candidate-bound SAST evidence | src/security/sastEvidence.ts, src/validators/external.ts, src/core/run.ts; tests/security/sastEvidence.test.ts, tests/packed/s10RealSastCampaign.mjs | `SastEvidenceV1` binds the exact candidate identity and workspace source digest, adapter, tool version, command digest, raw-output digest, findings, and isolation evidence under a candidate-scoped path. Digest verification rejects tampering; wrong-revision evidence is stale; missing evidence is required-blocking; workspace drift refuses persistence. The S4 security impact path requires candidate-bound evidence for opengrep/trivy. Real local OSS Trivy v0.70.0 execution produced a finding-bound FAIL for r1 and a clean PASS for r2. |
| Execution identity and result gateway | src/architecture/executionIdentity.ts, src/architecture/participantPlan.ts, src/operations/state.ts, src/workers/resultGateway.ts, src/workers/agentPrompt.ts, src/workers/runtimeSessions.ts, src/distributed/worker.ts; tests/executionIdentity.test.ts, tests/architectureWorkGraph.test.ts, tests/operations.test.ts, tests/structuredResultGateway.test.ts, tests/agentPromptStructuredResultProvenance.test.ts, tests/runtimeSessions.test.ts, tests/paseoPendingExecutionIdentity.test.ts, tests/workerExecutionAuthority.test.ts, tests/distributedExecutionIdentity.test.ts | S1 deterministically compiles versioned frozen policy, per-role ceilings, the deeply immutable expanded blueprint, and participant binding. The operation-state guard compares complete canonical policy values against a pre-callback snapshot, validates v1 on `operation.policy.bound`, allows only creation/rebinding of an unchanged policy there, and permits candidate assembly, controller takeover, or a changed previously bound semantics event to clear identity only through its owning lifecycle event. Generic add/replace/mutate/clear changes fail closed. Regressions prove same-declared-digest and in-place callback mutations fail without changing the durable record. Before any mutation callback, the guard snapshots the participant-ID set and each optional binding’s complete canonical body. It compares the union of pre-callback and post-callback participant IDs, preserving absent-versus-present binding state, so in-place mutation and participant deletion/re-keying cannot alias both sides. Binding add/replace is permitted only through the explicit validated bind event; candidate/takeover/semantics invalidation events may only clear bindings. Regressions mutate `runtime.sessionId` while retaining its stale declared digest, then separately delete/re-key the participant; each rejection preserves durable operation JSON byte-for-byte. Current result acceptance requires the full binding and checks durable candidate, execution revision, epoch, policy, output, runtime/session, and skill/context/prompt identity. OpenCode, Codex, Podman/OpenCode, Paseo, and distributed launch paths materialize or reserve an actual session before dispatch, then check the returned session against the binding. Captured valid and rejected cases are adapter-boundary evidence, not provider certification. The operation-state transition guard also snapshots executionSemanticsDigest and operationExecutionRevision presence/value before callbacks. Generic attempts to add, replace, clear, preempt the semantics digest, or synthesize an absent legacy revision are rejected; the owning event binds the first digest without advancing execution revision, treats the same digest as an identity no-op, and advances once while clearing policy and all participant bindings when replacing previously bound semantics. Candidate/takeover invalidations remain lifecycle-owned. The five new operation-state regressions assert durable-byte preservation. |
| Semantic decision surfaces | src/audit/intent.ts, src/core/triage.ts, src/entry.ts, src/operations/change.ts, src/cli.ts, src/issues/intake.ts, src/core/run.ts, src/candidates/assembler.ts, src/architecture/candidateAssurance.ts; tests/auditIntent.test.ts, tests/triage.test.ts, tests/issueIntake.test.ts, tests/recovery.test.ts, tests/candidateAssembler.test.ts, tests/candidateAssurance.test.ts | INTENT, ROUTE, STACK, ISSUE, FAILURE, and CANDIDATE_IMPACT remain MODEL assessments through the single AgentTopology-resolved AEH Semantic Assessor executed via Paseo; VALIDATION_NEED remains typed without broadening S2 scope. S4 consumes only actual post-assembly CANDIDATE_IMPACT whose deterministic candidate/base/patch bindings and digest verify against the current CandidateRevision. The deterministic compiler maps impact dimensions through frozen policy into assurance floors, actual configured Reviewer assignments, validation requirements, candidate-bound assertions, and evidence strength. Model output cannot choose identities, actions, tools, capabilities, policy, or acceptance. Explicit intent and hard assurance floors remain deterministic. S4 focused tests pass 7 files / 41 tests; full non-system verification passes 153 files / 848 tests with 7 skipped. This is deterministic fixture evidence; REAL_PROVIDER is pending. |
| Semantic Assessor topology, Paseo execution, and safety | src/semantic/assessment.ts, src/semantic/runtime.ts, src/agents/permissions.ts, src/agents/outputContracts.ts, src/workers/agentPrompt.ts; tests/semanticAssessorTopology.test.ts, tests/semanticAssessmentRuntime.test.ts, tests/semanticAssessment.test.ts | One general Semantic Assessor is resolved from AgentTopology and semantic capability policy, runs as an AEH-managed Paseo agent/session, and returns typed structured data to deterministic schema, evidence, binding, policy, receipt, provenance, digest, and cache checks. It is not a WorkGraph Participant; assessment permissions deny tools, reads from unbounded repository interfaces, writes, shell, network, delegation, acceptance, policy mutation, and capability grants. Deterministic tests verify topology, session provenance, stale/replayed evidence rejection, unknown retention, and fail-closed behavior. They do not establish REAL_PROVIDER behavior. |
| Stack discovery and detector retirement | src/participants/stack.ts, src/participants/index.ts, src/core/run.ts, src/core/git.ts, src/semantic/runtime.ts, src/architecture/validationRequirements.ts; tests/participantsStackDiscovery.test.ts, tests/participantsCanonical.test.ts, tests/semanticAssessmentRuntime.test.ts | Production stack meaning comes from the typed model judgment over bounded generic repository reads with verifiable receipts. STACK requires canonical realpath and current content digests before capture, after capture, after assessor execution, and before profile return. Its evidence paths are drawn from the same tracked/non-ignored-untracked inventory hashed by `computeWorktreeDigest`, or a stricter generic subset; ignored `.env` and `.harness` runtime/cache files are absent and cannot be cited. Deterministic code validates receipt paths and proposed project skill roots for normalization, existence, and realpath containment; these descriptive paths grant no capability. Technology-specific detector implementations, registries, APIs, fast paths, and exports were removed with no compatibility fallback. The refreshed stack suite passes 24 tests; the full focused S2 set passes 14 files / 123 tests. REAL_PROVIDER interpretation remains pending. |
| Knowledge and SkillTrust | src/knowledge/index.ts, src/agents/outputContracts.ts, src/agents/waveExecutor.ts, src/architecture/executionIdentity.ts; tests/knowledgeGate.test.ts, tests/agentPromptStructuredResultProvenance.test.ts, tests/distributedExecutionIdentity.test.ts | S1 requires grounded per-step claim/source references, retains exact accepted procedure bytes in scoped SkillManifest, projects them verbatim only to the assigned authorized participant, and binds the manifest digest at launch/result boundaries. Provider-backed Librarian research and REAL_PROVIDER evidence remain open. |
| Context Runtime v2 | src/context/, src/workers/agentPrompt.ts, src/workers/resultGateway.ts; tests/contextAuthorizationV2.test.ts, tests/contextRuntimeV2.test.ts, tests/contextEfficiency.test.ts, tests/agentPromptStructuredResultProvenance.test.ts, tests/structuredResultGateway.test.ts | S5 implements controller-fenced, read-lease-gated ContextRefAuthorizationV1 issuance; exact operation/candidate/participant-generation/blueprint/policy/epoch/session/ref/source/expiry identity; source-byte revalidation; durable candidate/session-bound retrieval receipts and budget; durable continuation sequencing and replay rejection; and result-bound ProgressiveContextManifestV1 evidence. S1 launch ContextManifest and PromptManifest digests remain immutable in ExecutionBinding. Before issuing a grant, the controller verifies the canonical launch-manifest digest against that binding, the manifest envelopeDigest against verified durable envelope provenance, and exact unique `(refId, artifactPath, sourceDigest)` set equality. Regression tests reject an extra ref in a self-consistent envelope, changed envelope digest with identical refs, and artifact-path/source-digest mismatches; exact match issues only the exact refs. The S5 deterministic gate was independently accepted after Director source/diff review and reruns: focused context/session/provenance 12 files / 125 tests; full non-system 152 files passed, 1 skipped; 835 tests passed, 7 skipped; TypeScript/build passed; relevant deterministic system suites 3 files / 45 tests. REAL_PROVIDER and live Paseo/MCP behavior remain pending. |
| Paseo session binding | src/paseo/sessionBinding.ts, src/paseo/launchSpec.ts, src/paseo/sdk.ts, src/paseo/runtimeCore.ts, src/workers/agentPrompt.ts; tests/paseo/sessionBinding.test.ts, tests/agentPromptStructuredResultProvenance.test.ts, tests/paseoPendingExecutionIdentity.test.ts, tests/paseoInitialTurnBarrier.test.ts | Production prompt preparation persists strict canonical PaseoSessionBindingV1 identity over operation/revision, participant/generation, candidate, blueprint, policy, context/prompt, controller epoch, and actual provider session. Reuse requires ACTIVE full equality; session drift rotates to a new actual session or rejects reuse, and the StructuredResult retains the returned actual session. Focused binding and launch/result tests pass; full non-system lane passes 320 files / 830 tests with 7 pending. Deterministic fixtures do not certify REAL_PROVIDER session behavior. |
| CandidateImpact and assurance | src/candidates/assembler.ts, src/candidates/wave.ts, src/candidates/repair.ts, src/core/run.ts, src/architecture/candidateAssurance.ts, src/architecture/acceptanceOracle.ts; tests/candidateAssembler.test.ts, tests/candidateAssurance.test.ts, tests/candidateWaveLineage.test.ts, tests/candidateRepairProtection.test.ts, tests/runRepairCandidateLifecycle.test.ts, tests/acceptanceOracle.test.ts | S4 consumes exact post-assembly impact and compiles current assurance assertions; S6 consumes these unchanged candidate/impact/policy/evidence-strength bindings and rejects stale evidence. Deterministic S4/S6 coverage is present; MODEL interpretation and live provider/session behavior remain separate. |
| Validation | src/core/verify.ts, src/core/run.ts, src/architecture/validationRequirements.ts, src/architecture/candidateAssurance.ts, src/validators/external.ts; tests/validationProviders.test.ts, tests/candidateAssurance.test.ts, tests/candidateAssembler.test.ts, tests/runRepairCandidateLifecycle.test.ts | Typed impact requirements resolve under frozen policy and are checked against resolver digest, exact requirement/action kind, scope, and evidence needs. The approved project/configured action is executed; missing, disallowed, unsupported, or non-PASS required evidence fails closed and records candidate/impact/policy identity. Playwright/OpenGrep/Trivy provider names use matching external adapters; unknown specialized providers cannot fall through to generic unit tests. Validation PASS is not product acceptance. Focused S4: 7 files / 41 tests; full non-system: 153 files passed, 1 skipped; 848 passed, 7 skipped. Provider/browser lanes remain pending at S4; S11 adds the candidate-bound lanes below. |
| Contract/integration lane evidence (S11) | src/validation/laneEvidence.ts, src/providers/validation/registry.ts, src/providers/validation/integrationEnvironment.ts, src/providers/validation/types.ts, src/validators/registry.ts, src/validators/openapi.ts, src/core/run.ts, src/architecture/validationRequirements.ts, src/architecture/candidateAssurance.ts; tests/validation/laneEvidence.test.ts, tests/validation/failClosedProviders.test.ts, tests/validation/requiredLaneEvidence.test.ts, tests/system/aehContractIntegrationCampaign.test.ts, tests/fixtures/integration/http-service.mjs | `ProviderLaneEvidenceV1` binds lane, check, candidate, workspace source digest, provider identity, command digest, status, findings, per-artifact digests, timestamps, and blockers under a candidate-scoped path; persist/load/verify/require reject workspace drift, stale candidates, tampered JSON/raw/artifacts, and lane cross-use. Declared specialized capabilities resolve to matching providers; missing Pact/OCI providers return explicit blockers rather than SKIP, and unsupported capabilities fail with `UNSUPPORTED_VALIDATION_CAPABILITY`. Review remediation: the S4 impact path no longer synthesizes lane evidence from an arbitrary passing check — `ensureProviderLaneEvidenceFromCheckV1` is deleted and `requireProviderLaneEvidenceForActionV1` fails closed with `PROVIDER_LANE_EVIDENCE_REQUIRED` unless the matching provider/adapter execution persisted the evidence. The OpenAPI adapter (`openapi → contract-test`) persists its own CONTRACT evidence when a candidate is bound. The integration lifecycle requires provision/readiness/test/cleanup, always attempts cleanup, and reports `cleanupRequired`/`INTEGRATION_CLEANUP_REQUIRED` only when resources were actually provisioned (no resources created never yields a cleanup blocker). Adversarial `requiredLaneEvidence.test.ts` proves project-script-only browser/visual/contract/integration routes FAIL closed while the OpenAPI, isolated-service, and Playwright-provider routes PASS. R2 found the declared `bdd`/`gherkin` adapter still bypassed the capability provider; `src/validators/registry.ts` now routes `gherkin`/`bdd` through `runCapabilityValidator` (configured BDD provider commands honored, missing runner → typed `CONTRACT_PROVIDER_UNAVAILABLE`, PASS → candidate-bound CONTRACT evidence) and the superseded `src/validators/gherkin.ts` is deleted, with `maturity/components.yaml` and `docs/component-maturity.json` migrated to `src/validators/registry.ts`. Focused S11 set 9 files / 57 tests; four deterministic system suites 4 files / 23 tests. Pact remains an unprovisioned fail-closed provider lane. |
| Browser/visual lane evidence (S11) | src/validation/laneEvidence.ts, src/validators/external.ts, src/architecture/candidateAssurance.ts, src/architecture/validationRequirements.ts, src/core/run.ts, scripts/s11BrowserVisualCampaign.ts, tests/browser/s11-control-center-browser-visual.e2e.ts, tests/browser/fixture/controlCenterJourney.ts; .harness/evidence/s11/browser-visual-campaign.json | A `visual-test` kind, `visual` adapter, and `UI/visual` impact dimension join the existing `browser-test`/Playwright mapping. Required browser/visual impact validations require candidate-bound BROWSER/VISUAL evidence; the pinned candidate Playwright binary is the only default provider, and missing binary/browser/candidate binding fails with explicit `*_PROVIDER_UNAVAILABLE`/`PROVIDER_LANE_CANDIDATE_BINDING_REQUIRED` blockers. Review remediation: each lane owns a separate durable Playwright output directory, so the VISUAL run cannot delete the BROWSER screenshot; the campaign verifies both lanes through `verifyProviderLaneEvidenceV1` after the full run and fails unless both verify with all artifacts present; VISUAL evidence binds the committed reference baseline digest and the comparison configuration (schema extension: `baseline` artifact kind plus `comparison`), with `VISUAL_REFERENCE_BASELINE_REQUIRED`/`VISUAL_COMPARISON_CONFIG_REQUIRED` blockers and no baseline auto-generation; and the sanitized machine summary moved into the ignored evidence directory so the campaign no longer rewrites a source-tree file and its candidate/workspace digest recomputes MATCH after the run. R2's LOW durability finding is repaired too: each lane's Playwright output directory is candidate-scoped (`.harness/evidence/playwright-s11-output/<lane>/<candidateId>-r<revision>-<identity12>`), so a later campaign run against a different candidate cannot delete a prior candidate's screenshots; the earlier `a20fd933b163` candidate evidence still verifies after the newer `7ad7d70b2ef1` run. Boundary: scripted deterministic Paseo provider conversation, not REAL_PROVIDER or packed certification; visual evidence is rendered artifacts plus in-run stability and a committed host screenshot baseline. |
| Review and convergence | src/agents/reviewLifecycle.ts, src/core/run.ts; tests/candidateAssurance.test.ts, tests/reviewLifecycle.test.ts, tests/reviewRepairCandidateLifecycle.test.ts, tests/runRepairCandidateLifecycle.test.ts | S4 assignments are actual configured canonical read-only Reviewer invocations bound to candidate, impact, policy, dimensions, and exact returned session identity. S6 requires a matching current participant ExecutionBinding and persisted result artifact; reviewer agreement remains evidence and never acceptance. Repair/rollback recompiles for the resulting candidate. Focused S6: 11 files / 124 tests; full non-system: 157 files passed, 1 skipped; 931 passed, 7 skipped. REAL_PROVIDER pending. |
| Acceptance | src/architecture/candidateAssurance.ts, src/architecture/acceptanceOracle.ts, src/architecture/objectiveCompletion.ts, src/agents/managedLeadAcceptance.ts, src/core/run.ts, src/operations/state.ts, src/cli.ts; tests/acceptanceOracle.test.ts, tests/managedLeadAcceptance.test.ts, tests/objectiveCompletion.test.ts, tests/objectiveCompletionLifecycle.test.ts, tests/runRepairCandidateLifecycle.test.ts, tests/deliveryFinalize.test.ts, tests/cliS8Handoff.test.ts | S4 assertions remain candidate/impact/policy-bound; S6 persists the current ACCEPTED/REJECTED disposition and recomputes objective completion. S8 consumes that unchanged current AcceptanceOracle before any requested public issue/branch/push/PR effect; controller-owned local execution workspace setup remains permitted before acceptance and outside delivery effects. The S8 default issue-path regression covers pre-acceptance workspace action and post-acceptance receipted delivery. The standalone `aeh sdd handoff` CLI path reports unsupported before project loading because it has no managed accepted-delivery context. Live managed provider and packed journeys remain pending. |
| Delivery and reconciliation | src/issues/workflow.ts, src/delivery/handoff.ts, src/delivery/finalize.ts, src/cli.ts, src/architecture/acceptanceOracle.ts, src/operations/state.ts, src/security/actionPolicy.ts, src/security/gatedAction.ts, src/security/toolActionGate.ts, src/security/actionReconciliation.ts, src/provenance/generate.ts; tests/deliveryFinalize.test.ts, tests/cliS8Handoff.test.ts, tests/security/actionPolicy.test.ts, tests/security/gatedAction.test.ts, tests/security/toolActionGate.test.ts, tests/security/actionReconciliation.test.ts, tests/system/aehScenarioMatrix.test.ts, tests/system/aehConcurrencyCampaign.test.ts | S8 routes default issue execution through a managed run: the controller provisions the isolated local workspace before acceptance, while public GitHub issue/branch/push/PR effects require current AcceptanceOracle acceptance and policy-required S7 provenance. paseo.workspace.create is LOCAL_RESOURCE_CREATION, excluded from configured delivery effects, and remains controller-fenced with durable intent/receipt and conservative HUMAN_REQUIRED reconciliation when identity is uncertain. Sensitive effects use ToolActionGate, bound ActionIntents and ActionReceipts; executor uncertainty is inspected before retry. Normal `aeh sdd handoff` invocation is explicitly unsupported before project loading and does not enter direct delivery code; this adds no lifecycle semantics. The deterministic S8 gate was independently accepted by the Campaign Director on 2026-09-25. Mocked GitHub requests/local Git remotes do not establish provider certification. |
| Control Center | src/control-center/, src/entry.ts, ui/control-center/src/api.ts, ui/control-center/src/App.tsx; tests/controlCenter.test.ts, tests/controlCenterApiContracts.test.ts, tests/humanDecision.test.ts, tests/operations.test.ts | The authenticated UI projects the current bounded DecisionRequest, displays evidence/options/consequences and the saved resume target, and submits only operation/request/choice identifiers plus optional reason. The paired server derives actor and current operation/candidate/policy/revision/epoch and rejects stale, expired, malformed, mismatched, or replayed choices. UI typecheck/build and deterministic API boundary tests pass; fixture-backed browser choice, blocker, runtime-lease, and cancellation journeys passed, while durable-operation browser certification remains pending. |
| Supply chain | src/provenance/generate.ts, src/build/identity.ts, src/core/config.ts, src/core/types.ts, schemas/project.schema.json, src/delivery/finalize.ts, src/cli.ts, src/core/doctor.ts; tests/provenance.test.ts, tests/provenanceS7.test.ts, tests/projectSchemaS7.test.ts, tests/buildHygiene.test.ts, tests/deliveryFinalize.test.ts, tests/system/aehScenarioMatrix.test.ts, tests/system/aehAdversarialE2E.test.ts | Strict version-2 evidence binds the exact CandidateRevision, configured packed-artifact path/digest, running BuildIdentity, supply-chain policy digest, expected builder, CycloneDX SBOM digest, in-toto subject, SLSA predicate, and policy-required Cosign verification. S8 handoff/finalization consumes the existing policy-required S7 gate after current S6 acceptance and before public effects; no S7 semantics changed. Deterministic fake signing is contract evidence only, not REAL_PROVIDER certification. S7 was accepted by the Campaign Director on 2026-09-24; the S8 integration evidence is complete and its deterministic gate was accepted by the Campaign Director on 2026-09-25. |
| Certification contracts | src/certification/; tests/certification.test.ts, tests/certificationMatrix.test.ts | Required-evidence enforcement and capability matrix contracts have deterministic references. They do not establish real-provider certification. |
| Observability identity, metrics, and advisory evals (S12) | src/telemetry/identity.ts, src/telemetry/events.ts, src/telemetry/metrics.ts, src/evals/runner.ts, src/evals/scoring.ts, src/evals/advisoryInvariant.ts; tests/telemetry/, tests/evals/corpus.test.ts, tests/evals/advisoryInvariant.test.ts, tests/helpers/otlpReceiver.ts; evals/corpus/, evals/fixtures/, evals/scenarios/; scripts/s12ObservabilityCampaign.ts; docs/evidence/s12/s12-observability-evals-campaign.json | `TelemetryCorrelationV1` binds operation/candidate/execution/policy/controller/participant/runtime/session identity with one correlation digest across NDJSON events, OTel spans, and metric points; stale supplied identity is marked, never silently re-attributed. Real OTel counters/histograms cover operation/participant/runtime/validation signals and export canonical local snapshots. The versioned corpus records corpus/case digests and scores reproducibly. The advisory-invariant guard proves no authority decision module imports observation modules and fabricated eval/telemetry artifacts cannot change validation, durable operation truth, acceptance, or routing. The OTLP lane is verified against an in-process loopback receiver; it is not a hosted collector or REAL_PROVIDER certification. |

The Campaign Director independently accepted S4 on 2026-09-24 after reviewing the source diff and affected evidence records, then rerunning the focused candidate-assurance set (7 files / 41 tests), `npm run typecheck`, `npm run build` (`release-1790299600802-17-85b48e9b`), and the full non-system Vitest lane (153 files passed, 1 skipped; 848 tests passed, 7 skipped). These deterministic results do not certify REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E behavior.

## REAL_PROVIDER lane

**Status from the audit:** not established for the complete claimed capability matrix.

A real-provider claim requires a freshly packed candidate, provider startup and execution receipts, current candidate/build identity, deterministic oracle checks, and required independent review. Deterministic fixtures and contract tests cannot certify real providers.

The audit identifies incomplete or unproven real-provider journeys for semantic assessment, Librarian research, Paseo, Podman/OCI isolation, integration environments, browser validation, delivery/reconciliation, and the full capability matrix. S2 deterministic Paseo fakes do not change that status. The Serena smoke recorded in the historical ledger proves only the tested Serena MCP broker path; it does not certify unused provider lease/release lifecycle or model execution. S10 adds real local-provider evidence — actual rootless `bwrap 0.13.0` namespace execution and a real Trivy `v0.70.0` scan with checksum-verified provisioning — recorded separately from hosted-provider certification. Rootless Podman/OCI execution remains unprovisioned in this environment (`podman`, `buildah`, `crun`, `slirp4netns`, `fuse-overlayfs` absent; no `sudo`); that lane remains open for S13 and the product fails closed when the provider is missing.

**S13 real-provider campaign — 2026-09-26 (round 2; implementing lead report; `SLICE_BLOCKED`; independent review pending):** `tests/packed/s13RealProviderCertification.mjs` packs the current checkout through the repaired `runExternalSelfDogfood` bootstrap, installs the artifact into a disposable `/tmp` fixture, and certifies capability lanes against real providers with the new `createCapabilityJourneyOracle` (`src/certification/journeys.ts`). All four executed lanes are ACCEPTED: `startup`, `project-home`, and `multi-project` through real `codex-cli 0.156.1`/`gpt-6-luna` actor sessions (for example startup session `01a0e000-abad-7811-8b44-9c078f249b63`, `init`/`doctor` exit 0, packed artifact `c4048c90bd6b3d97…`), and `informational` through the packed product's real semantic runtime: a direct packed `intent` execution launched a real Paseo `opencode/opencode-go/mimo-v2.6-flash` assessor session with `thinkingOptionId: "low"` at the default topology (workhorse alias migrated to `mimo-v2.6-flash`), returned a schema-conformant INTENT assessment (`intent: informational`, `mechanism: MODEL`), and persisted it in the durable semantic cache. Round-2 product repairs on the real-Paseo path: `semanticModelDeadlineMsV1 = 300_000` with the request schema ceiling raised; `boundSemanticThinkingOptionV1` caps the Paseo thinking option by the request reasoning class; `waitForPaseoAgentHandle` polls the canonical snapshot every 2 s as a bounded fallback; Paseo timeline fetches use the supported `direction: "tail"` (the SDK rejects the retired `"backward"`); `parseSemanticAssessmentOutputV1` extracts exactly one typed JSON object from a prose/fenced reply; the launch prompt carries `outputJsonSchema`. The campaign's before/after tracked-file digest is identical (`9aecda35170ad80f…`, 668 files) and candidate execution is confined to `/tmp`; machine evidence `docs/evidence/s13/s13-real-provider-certification.json` SHA-256 `e30db968c9e34b98d342e9e676a6a9dc90a574280e78436953248b6bd8c38621`. These are REAL_PROVIDER + PACKED_E2E + SYSTEM_DETERMINISTIC results for four rows only; the remaining 18 rows are `BLOCKED_UNAVAILABLE_PROVIDER` (17; exact actions in the S13 WorkGraph) or `NOT_APPLICABLE` (`control-center`, BROWSER lane). No hosted-provider, OPA, OCI/Podman, signing, full packed-operation, or live-Paseo-lifecycle certification is claimed.

**S13 real-provider campaign — 2026-09-26 (round 3; implementing lead report; `SLICE_BLOCKED`; independent review pending):** round 3 built the missing harnesses and repaired six blocking defects. `tests/packed/s13GovernedOperationCampaign.mjs` installs a freshly packed artifact into a disposable `/tmp` fixture and runs the canonical `aeh operation start` control path with real Paseo participants (no deterministic stubs) plus a deterministic per-row oracle over the durable operation record, candidate revision, participant receipts, validation report and acceptance artifacts; the AUDIT lane advanced through real `gpt-6-luna` supervisor initialization, real reviewer sessions, configured validation and into consolidation, and the CHANGE/DIRECT lane through real semantic ROUTE/STACK assessments into implementation staging. Both rows remain non-accepted and are reported blocked with exact defects: AEH-V2-0074 (`EXECUTION_BINDING_STALE: ContextManifest changed after session materialization` when a persistent supervisor resumes with the consolidation context) and AEH-V2-0075 (a transient workspace digest mismatch at the DIRECT baseline assert whose retained worktree later recomputed to the exact bound r2 digest, plus a schema-valid-but-semantically-invalid STACK payload with no bounded retry). `tests/packed/s13RealPaseoLifecycleCampaign.mjs` certifies the `cancel` and `recovery` rows with real Paseo provider sessions owned by real durable operations: the lease owner is SIGKILLed, a strictly newer controller epoch is claimed, and the production cancellation path (real `paseo stop` + `inspectManagedPaseoAgent`) or the production provider-lease takeover (real inspect/stop of the exact prior session) is driven to a drained lease set; sessions `11e38c87-4a6e-401a-85e1-da0d75f32a7d` (`CANCELLED`) and `fb71ebe3-97af-4ed5-b57f-b251e66519c6` (takeover + resumed work). `tests/packed/s13GithubEffectCampaign.mjs` used the authorized disposable PRIVATE repository `JamesMorales04/aeh-s13-round3-20260927035743-6bb5da`: real issue #1 snapshot frozen with `contentSha256` `bd777eed917f9e30463af3c1790e3d5baf00017c17225ca32b1aa99a43bdfcc8` and the deterministic `ISSUE_DRIFT` gate proven against a post-freeze edit; a real controller-authority gated effect was denied `HUMAN_REQUIRED` without a human decision and `EXECUTED/SUCCEEDED` with an `ACTION_AUTHORIZATION` `APPROVE` decision in the durable `HumanDecisionLedgerV2`, pushing the real branch `aeh-s13-authority-20260927035935` observed through the GitHub API. `issue import` is PARTIAL/blocked by AEH-V2-0076 (`V2_AUTHORITY_REQUIRED` because the canonical Planner launch requires a managed operation), `delivery` is blocked pending an accepted packed change, and repository deletion is blocked by the missing `delete_repo` scope (AEH-V2-0077, exact 403 captured, private repository remains). Product repairs AEH-V2-0068…0073 (audit bootstrap route/assurance; policy rebind after workspace candidate binding; controller-side semantic assessments excluded from the managed participant lease lifecycle; repair continuation execution labels; supervisor event skill seed completeness; remaining semantic deadlines at the provider-appropriate ceiling) have focused regressions (`tests/governedOperationBootstrap.test.ts`, `tests/managedForegroundAssessmentLease.test.ts`, `tests/reviewerSerializationRecovery.test.ts`, `tests/supervisorPromptPolicy.test.ts`) and are shipped in the round-3 packed candidates. Round-3 focused validation: 21 files / 144 tests passed; `npm run typecheck` exit 0; `npm run build` exit 0; `git diff --check` clean; tracked checkout digest unchanged across all campaigns; candidates ran only in disposable `/tmp` fixtures; no credentials persisted. Machine evidence: `docs/evidence/s13/s13-governed-operations.json` SHA-256 `d91cc27370f9b9a0…`, `docs/evidence/s13/s13-real-paseo-lifecycle.json` SHA-256 `6ec75306455c3962…`, `docs/evidence/s13/s13-github-effects.json` SHA-256 `228e3e963f48fabe…`. The `cancel`, `recovery` and `authority` lanes are accepted by the implementing campaign; `issue-driven` is PARTIAL; all FULL_OPERATION rows remain blocked; no hosted-provider, OPA, OCI/Podman, signing, delivery, or complete packed-operation certification is claimed.

**S13 round-4 independent review — 2026-09-27 (fresh reviewer; `SLICE_BLOCKED`):** the reviewer verified the dirty candidate against the normative target and evidence (HEAD `e3dad34` at S12 acceptance with S13 fully uncommitted; `git diff --check` clean; TARGET byte-untouched; all thirteen documented evidence digests recomputed and matching; 12 focused files / 84 tests and `npm run typecheck` green independently). Accepted rows were re-derived, not inherited: `startup`, `project-home`, `multi-project`, `cancel`, `recovery` and `authority` are `UNCHANGED_EVIDENCE_VALID` on their execution paths; `informational` is `STALE_REQUIRES_RERUN`; `issue-driven` is `BLOCKED` (PARTIAL); `control-center` is `NOT_APPLICABLE`; the remaining FULL_OPERATION rows are `BLOCKED`. The AEH-V2-0075/0081 untracked-identity exclusion was independently reviewed and is not a product regression: it is explicit, path-scoped (`.serena`, `graphify-out`), untracked-only, deterministic, and an independent reproduction proved a newly created legitimate untracked source file still enters candidate identity and the DIRECT ChangeSet, while tracked provider configuration still contributes (R4-F1 requires the missing negative regressions and an intent-to-add-aware or config-driven filter). Blocking findings owned by round 5: R4-F2, the round-4 report's change-lane narrative (implementer receipt / isolated mutation / ChangeSet assembly; ledger rows 0079/0080) is not supported by `governed-lanes/change-direct.json` (`FAILED`, 0 receipts, unmutated candidate digest, no validation/acceptance) and the report mis-numbers defects and mis-attributes the residual failure, and a claimed watchdog trace was not preserved; R4-F3, `finalizeOperationSession` still targets operation state through the isolated execution root so isolated DIRECT participant receipts cannot persist (AEH-V2-0080 incomplete); R4-F4, the AEH-V2-0082 audit stall is an unverified hypothesis without traces and must be reproduced and classified before repair; R4-F5, AEH-V2-0076 is confirmed as the planner launch requiring a managed operation, with a repair contract that preserves controller→scoped authority→participant→effect semantics; R4-F6/F7, structured-output completion extraction, timeout classification and the fail-open workspace guard need hardening. No S13 row is independently accepted; machine record `docs/evidence/s13/review-round4.json`.

**S13 repair/re-run campaign — 2026-09-27 (round 4; implementing lead report; `SLICE_BLOCKED`; budget stop; independent review subsequently also `SLICE_BLOCKED` — see above):** round 4 repaired the round-3 FULL_OPERATION blockers and four further real defects in the same path, each with a focused regression. AEH-V2-0078 adds an explicit `continueBoundSession` continuation mode so the persistent Operation Supervisor can execute a consolidation event on its frozen generation binding (candidate, execution revision, blueprint, policy, controller epoch, participant generation and runtime session are all still verified; only turn-scoped prompt/context manifests are delivered anew, with the delivered turn recorded in the consolidation artifact). AEH-V2-0079 stops an isolated DIRECT provider session from claiming the operation workspace, so the implementer now edits the disposable isolated tree instead of the bound candidate worktree. AEH-V2-0080 resolves operation artifacts (transcripts, accepted structured results, receipts) against the operation control root for isolated runs. AEH-V2-0081 excludes untracked provider-generated scratch (`.serena`, `graphify-out`) from candidate/repository identity and from DIRECT ChangeSet pathspecs, and adds a bounded digest-input inventory to mismatch errors. AEH-V2-0083 makes controller-side assessor turns capture the completed OpenCode `assistant_message` via the canonical timeline fallback, and gives the semantic service one bounded retry for schema-invalid/unparseable output (with a repair note naming the expected judgment type and allowed evidence refs) and one bounded retry for a hung provider turn; other provider/authority failures still fail closed. The durable round-4 artifacts prove only this much: the packed change lane reached real ROUTE, STACK and CANDIDATE_IMPACT assessments and candidate r2 binding, then FAILED at implementation with 0 terminal and 0 candidate-bound participant receipts, an unmutated candidate source digest and no validation/acceptance artifact; the packed audit lane reached real supervisor initialization, reviewer sessions, validation and consolidation, persisted its audit report, but never reached a terminal operation state. The round-4 narrative claiming "a real implementer receipt, isolated mutation and ChangeSet assembly" (and the corresponding ledger rows 0079/0080) was not supported by `governed-lanes/change-direct.json` and is corrected in round 5 (R4-F2). Neither row reached an accepted terminal state before the budget stop: the audit silent non-terminal exit is AEH-V2-0082 (round-5 reproduction: the `return terminalizeOperation(...)` without `await` bypassed the controller catch, so the successful-terminalization gate rejection killed the detached controller and left the operation RUNNING; the round-4 "watchdog wake during consolidation" mechanism and the claimed `operation.watchdog.supervisor`/`no-lead` trace were unverified and are not preserved) and the change surface remains exposed to real model-output flakiness (AEH-V2-0083). `aeh issue import` remains blocked by AEH-V2-0076 (`V2_AUTHORITY_REQUIRED` outside a managed operation; no round-4 repair). Round-4 evidence: `docs/evidence/s13/governed-lanes/audit.json` SHA-256 `108b3cd5095a9422…`, `docs/evidence/s13/governed-lanes/change-direct.json` SHA-256 `7a51ad2d158bcd32…`, current single-lane summary `docs/evidence/s13/s13-governed-operations.json` SHA-256 `5d848848c16f5072…`. Round-4 focused validation: 25 files / 187 tests passed; `npm run typecheck` exit 0; `npm run build` exit 0; `git diff --check` clean; every campaign run records `checkoutUntouched: true` for its own before/after tracked digest; candidates ran only in disposable `/tmp` fixtures; disposable Paseo agents/workspaces and `/tmp` staging were cleaned; no credential persisted. No row is claimed without a fresh candidate/build-bound real-provider receipt, and the implementing lead does not accept its own slice.

**S13 round-5 repair/re-run campaign — 2026-09-27 (implementing lead report; `SLICE_BLOCKED`; fresh independent review owns acceptance):** round 5 corrected the round-4 narrative to match the durable round-4 artifacts (R4-F2), then repaired and re-ran the governed lanes. It reproduced the audit stall with `S13_GOV_KEEP=1`, per-stage timestamps and `operation.watchdog.*` trace capture and established the owning defect: the audit report was persisted and consolidation completed, but `return terminalizeOperation(...)` without `await` bypassed the controller catch, so the successful-terminalization gate rejection (`V2_TERMINAL_GATE_REJECTED: participant participant:<digest> has no terminal receipt`) killed the detached controller and left the operation RUNNING; the round-4 "watchdog wake during consolidation" mechanism and its claimed trace were not supported. Product repairs: `await` terminalization on every controller branch (gate failures become durable FAILED — AEH-V2-0082/0084); one durable bounded-work identity per participant with the Paseo session id as receipt provenance, supervision/assessor identity exemptions, and no second work-participant registration for authority-bound or semantic-assessment sessions (AEH-V2-0085); all finalization operation-state access routed through the authoritative operation state root (R4-F3); `aeh issue import` routed through a controller-owned managed operation with controller-issued Planner authority, deterministic intake terminal evidence and explicit candidate-lineage advancement (AEH-V2-0076); typed timeout classification, canonical last-assistant extraction, traced timeline-refetch failures and a verified `direction: "tail"` enum with retries still bounded at 2 (AEH-V2-0088); a positive-proof workspace guard (R4-F7); and an intent-to-add-aware provider-scratch filter with the mandated regressions (R4-F1). The packed `audit` lane is ACCEPTED against a fresh candidate with real Paseo sessions and `checkoutUntouched: true` (`governed-lanes/audit.json` SHA-256 `84517c3dd6445f9ad1baee8b76221fc67babd75bd7841f615f7862d1c3740a6e`; packed artifact `8e67326c784c8ff7…`, build `release-1790503184568-945008-1133190d`, gitSha `e3dad34`; last-invocation summary SHA-256 `2f25ebd7df6bf05a671e7c018c571dc6cfd8720c0f030d0897661299c5cc948d`). The packed `change-direct` lane reached a truthful terminal `FAILED`: the final build-bound run stopped at the STACK semantic assessment (unparseable real assessor output after the bounded retry, `SEMANTIC_ASSESSMENT_INVALID`, captured in `operation.error`) before any participant launched, leaving candidate r2 bound to the baseline digest with no ChangeSet/validation/acceptance artifact; an earlier round-5 run on the same source reached the DELEGATED path with 4 candidate-bound receipts and a failing configured validation command (`governed-lanes/change-direct.json` SHA-256 `beb0f5d606fdf1ee…`), so the row remains blocked on real model implementation quality (AEH-V2-0083-class). Round-5 scope is limited to the lanes actually re-run: `informational` is STALE_REQUIRES_RERUN, `cancel`/`recovery`/`authority` remain accepted by path-equivalence but still bind pre-round-5 builds (a build-bound sweep remains mandatory before any checkpoint), `issue-driven`'s managed import path is REAL-PROVIDER VERIFIED in a stable clone of the authorized disposable repository (terminal `SUCCEEDED`, bounded Planner receipt, sealed `GH-1` contract digest, frozen snapshot digest, route `FORMAL_SDD`; `docs/evidence/s13/issue-import-round5.json` SHA-256 `41cfbc5ce0328f72…`) while the full row remains blocked on change/delivery, and `delivery` remains blocked. Focused regressions green; `npm run typecheck` exit 0; `npm run build` exit 0; `git diff --check` clean; TARGET untouched; candidates executed only in disposable `/tmp` fixtures; workspaces archived and staging cleaned; no credential persisted. The full suite also caught one real round-5 regression (AEH-V2-0089: resumable launch identities had their context execution closed by receipt recording; fixed by keeping launch identities non-terminal and mapping receipts by participant or session identity) and left environment/pre-existing failures only (189/196 files passed; `tests/system/` daemon and run-kind objective-gate fixtures, `tests/paseoSdkResolve` local mise shadowing, and parallel-only flakes), recorded as unresolved and independent of the certification lanes.

**S13 round-5 independent review — 2026-09-27 (fresh reviewer; `SLICE_BLOCKED`):** the reviewer verified the dirty candidate against the normative target and evidence (HEAD `e3dad34`; 46 tracked modifications / 11 untracked paths; `git diff --check` clean; TARGET byte-untouched; all 13 documented digests recomputed — the 10 matching ones confirmed, and `s13-github-effects.json`, `github-lanes/authority.json` and `github-lanes/issue-driven.json` found rewritten in round 5 while the docs still cite the round-3 digests), re-ran 23 focused files / 197 tests and `npm run typecheck` green, and independently reproduced the full suite (196 files: 192 passed / 4 failed; 1,307 tests: 1,275 passed / 25 failed / 7 pending). `audit` is `CURRENT_ACCEPTED` (packed artifact `8e67326c…`, real supervisor/reviewer sessions, report + validation + oracle PASS, `checkoutUntouched: true`); `change`/`direct-change` remain `NOT_ACCEPTED` with the primary root cause classified `AEH_PROVIDER_ADAPTER_DEFECT` (operation-scoped assessor turns lose provider-enforced structured output while the runtime capability requires it; STACK retries compose to up to four real model turns; rejected output not captured) and a precise Round-6 repair contract recorded. The full-suite groups are `PRE_EXISTING_REPRODUCIBLE` at `e3dad34` for `aehHumanBlackBox`/`aehAdversarialE2E`/`paseoSdkResolve` and `TEST_DEFECT` for the racy `tests/operations` descendant-cancel fixture; a pre-existing clean-build asset-resolution defect (`init` ENOENT without stale shared `dist/releases` assets, which `npm pack` still ships) was reproduced. Only `audit` is current; `startup`, `informational`, `project-home`, `multi-project`, `cancel`, `recovery`, `authority` and the managed issue import are `STALE_REQUIRES_RERUN`, `issue-driven` is not satisfied by the import path alone, and no canonical 22-row `requiredEvidence` requires trivy/opengrep/podman/buildah/opa/cosign. Machine record: `docs/evidence/s13/review-round5.json`.

**S13 round-6 repair/re-run campaign — 2026-09-27 (implementing lead; `SLICE_BLOCKED`; fresh independent review owns acceptance):** round 6 repaired the round-5 provider-adapter root cause (AEH-V2-0092: assessor turns now request the provider-enforced structured-output channel on the session-creating turn with the unchanged schema and no AEH result sink/lease/participant registration; total real-model attempts bounded at two under one owner; bounded sanitized rejected-reply fingerprint and failed assessor session id persisted in failure details and traces), the two real product defects the lane then exposed (AEH-V2-0090: `CANDIDATE_IMPACT` dimensions constrained to the canonical typed vocabulary so the compiler can no longer fabricate unresolvable `unit-test` requirements; AEH-V2-0091: deterministic DIRECT contracts bind configured command/validator ids as requirement→validator traceability), the pre-existing clean-build asset-resolution defect (AEH-V2-0093: `init` resolves the running release's own assets; stale shared `dist/releases` leftovers removed; fresh-build/packed regressions) and the racy descendant-cancel fixture (AEH-V2-0094), plus the R5-F1/F4/F6/F7/F10 record/harness corrections. Build-bound results: `audit` ACCEPTED (`governed-lanes/audit.json` `7c893bf0ea8bea20…`, packed artifact `04ef5ed70f7bbaf7…`, real reviewer/supervisor sessions); `startup`/`informational`/`project-home`/`multi-project` 4/4 ACCEPTED fresh packed candidates (`s13-real-provider-certification.json` `61532afa9695cc19…`); `cancel`/`recovery` 2/2 ACCEPTED (`s13-real-paseo-lifecycle.json` `1dbf617e2b3b1658…`); `authority` ACCEPTED (`s13-github-effects.json` `e1c89b57bb76dfc7…`, branch `aeh-s13-authority-20260927191920`); managed issue import terminal `SUCCEEDED` (`issue-import-round6.json` `de71fac2982856ad…`) while full `issue-driven` remains NOT ACCEPTED. `change`/`direct-change` is NOT ACCEPTED after four build-bound real runs: the provider-enforced STACK assessment on `opencode-go/mimo-v2.6-flash` returned syntactically malformed JSON (stray `}` at offset 2107; 5,330 bytes; sha256 `c1d23da8…`) in both bounded attempts before any participant launch, recorded as `REAL_PROVIDER_CONTRACT_LIMITATION` without weakening the schema, synthesizing fields or raising retries (`governed-lanes/change-direct.json` `d05a8fc650be78f1…`); all dependent FULL_OPERATION rows remain blocked. Focused validation 23 S13 files / 200 tests plus `buildHygiene` 6/6, `npm run typecheck` exit 0, `npm run build` exit 0, `git diff --check` clean, TARGET byte-untouched, every campaign `checkoutUntouched: true` (`ca509c20808e2102…`, 668 tracked files), no credential persisted, all disposable workspaces/staging cleaned.  **Round-6 structured-output amendment (2026-09-27; `SLICE_BLOCKED`):** a bounded real-channel probe (`docs/evidence/s13/structured-output-probe.json`, `a72260fb2b1fe9fb…`) replaced the provisional `REAL_PROVIDER_CONTRACT_LIMITATION` with a precise decomposition. Paseo passes `format: json_schema` to OpenCode; OpenCode exposes the canonical structured object through an injected `StructuredOutput` tool (`info.structured`, `finish: "tool-calls"`, `StructuredOutputError.retries = 0`, no platform input validation); AEH consumes the canonical object whenever OpenCode produces it. Owning AEH defects: the Semantic Assessor permission projection's `"*": "deny"` suppressed that tool (repaired with an explicit `StructuredOutput: "allow"`), and runtime-level `structuredOutput=true` was treated as model proof while the assessor shipped on `opencode-go/mimo-v2.6-flash`, which is not certified for schema-bound output (repaired with a deterministic certified capability registry, required level `SCHEMA_BOUND_TOOL_CALL`, the `@structured-assessor` default `opencode-go/gpt-6-luna`, and fail-closed selection with no silent fallback; `mimo-v2.6-flash` stays documented as uncertified). The canonical output discipline is retained as error reduction only. After the repair the AEH-mediated assessor probe passed in 11.3 s with a typed STACK judgment (digest `5afdd632…`, real session `51014e17…`), and one bounded change/direct-change rerun reached real implementation, validation PASS, a real independent Reviewer with `candidate.assurance.reviewer.0.reviewer` PASS, `evidence.requirement-coverage` PASS and review `ACCEPTED/CONVERGED`, then terminal `FAILED` at deterministic AcceptanceOracle wiring (`VERIFICATION_VALIDATION_PATH_MISSING` for base contract assertions; review evidence matcher requiring binding+artifact on one participant record — AEH-V2-0099/0100). The run exceeded the 900 s lane bound and terminalized later (`governed-lanes/change-direct-late-terminal.json`, `ffb42c0e838a7a74…`). Schema, retry bounds, and acceptance/evidence gates are unchanged.

**S13 round-6 amendment independent review — 2026-09-27 (fresh reviewer; `SLICE_BLOCKED`):** the reviewer verified the uncommitted candidate at HEAD `e3dad34` from source and evidence (55 tracked modifications / 13 untracked paths; `git diff --check` clean; TARGET byte-untouched; all 17 documented round-6 digests recomputed and MATCH), re-ran the mandated focused files (`semanticStructuredOutputCapability`, `semanticAssessment`, `semanticAssessmentRuntime`, `semanticAssessorTopology`: 4 files / 34 tests green) and `npm run typecheck` exit 0 (`docs/evidence/s13/review-round6.json`). A: the structured-output repairs are valid and should be frozen — assessor projection exactly `{"*": "deny", StructuredOutput: "allow"}` (`src/agents/permissions.ts:188-195`), canonical levels and required `SCHEMA_BOUND_TOOL_CALL` with the probe evidence digest equal to the file sha256 (`src/semantic/structuredOutput.ts:16-46`), per-model fail-closed selection (`src/semantic/assessment.ts:474`), `@structured-assessor` = `opencode-go/gpt-6-luna` with `workhorse` still `mimo-v2.6-flash` (uncertified), prompt-level discipline only, interior-stray-brace rejected, single-owner retries, and no change to `acceptanceOracle.ts`/`run.ts`; caveat R6-F1: the 11.3 s AEH-mediated probe digest/session is not durably recorded. B: the amendment rerun is durable for implementation/candidate r3/reviewer receipt+session artifact/review stage completion and the `REJECTED` AcceptanceOracle disposition (digest `53436e86…` recomputed canonically, 9 blockers), but "validation PASS", "requirement-coverage PASS", "reviewer.0.reviewer PASS" and "ACCEPTED/CONVERGED" are source-gated inferences without durable artifacts; the late-terminal record has no preserved generator (R6-F2/R6-F4). C: AEH-V2-0099 (DIRECT base validation requirements empty, `src/core/run.ts:274` + `src/architecture/candidateAssurance.ts:206-224,353-354`; oracle gate `acceptanceOracle.ts:133-137,329`; independent `requirement-coverage` matcher `src/evidence/graph.ts:31,59-60`) and AEH-V2-0100 (binding on the controller-issued authority participant `src/operations/state.ts:1253-1258`; artifact on the runtime session `src/workers/agentPrompt.ts:1238-1268`; single-record predicate `acceptanceOracle.ts:282,285-292,301`) are confirmed AEH-owned deterministic defects with the repair owned by producer/compiler/resolution/provenance wiring, not the gate. D: all strict gates must remain and TARGET does not contradict the repair boundary. Blocking for any checkpoint: every accepted round-6 lane artifact (audit, certification 4/4, cancel/recovery, authority, issue import) binds a pre-amendment build (the amendment edited `permissions.ts`/`assessment.ts`/`runtime.ts`/`structuredOutput.ts`/`presets/agents/default.jsonc` after those runs; `informational` durably records the now-refused `@workhorse`/`mimo-v2.6-flash` assessor), so a post-amendment build-bound sweep is required (R6-F3). Round 7 owns AEH-V2-0099/0100, a clean terminal change run with persisted validation/review/acceptance artifacts, and the sweep; S14 must not start.

**S13 round-12 repair record — 2026-09-28 (implementing lead; `SLICE_BLOCKED`; independent review pending):** the P0-A/P0-B repairs and their focused regressions are recorded in [S13 WorkGraph](CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md) under "Execution record — round 12". `AEH-V2-0115`: `src/spec/openspec.ts` now validates a READY Spec Manager result against the canonical OpenSpec change layout before any controller-owned write and fails with `SPEC_MANAGER_CONTENT_NOT_CANONICAL` naming the exact artifact; `artifacts.specs` is one canonical `{ capability, content }` delta per capability and persists to `specs/<capability>/spec.md` (`tests/specManagerCanonicality.test.ts`). `AEH-V2-0110`: escalation requires the Planner's typed `formalizationNeed: REQUIRED` and records the deterministic constraint plus the exact contributing model judgment (`tests/changeRouting.test.ts`); the delegated pipeline ordering, provisional-policy supersede, skill selection, permission projection, validation resolution and approval-stop handling each have focused regressions. The live `change-delegated` lane reached planning COMPLETED, participant compilation, a real Implementer session and validation PASS, then blocked on `AEH-V2-0116` (Paseo/OpenCode provider approval stop despite `auto_accept`); no accepted DELEGATED completion and no live FORMAL_SDD rerun exist yet. These are deterministic source/regression results; they do not establish REAL_PROVIDER completion for those rows.

**S13 round-13 conformance record — 2026-09-28 (implementing lead; `SLICE_BLOCKED`; independent review owns acceptance):** AEH-V2-0116 is diagnosed at the Paseo daemon boundary with a machine record at [docs/evidence/s13/round13-0116-diagnosis.json](evidence/s13/round13-0116-diagnosis.json): the AEH-owned causes are repaired (scoped `external_directory` projection for the frozen launch root and git metadata roots; bounded pending-approval capture from the Paseo wait result; Codex sandbox projection `approval_policy: never` + `sandbox_mode` from the compiled ceiling; Paseo parent handle suppressed for isolated launches, AEH-V2-0117) and two provider behaviors are documented rather than waived (`auto_accept` is disabled whenever a `toolPolicy` is present; Paseo children inherit the parent workspace/cwd). The repaired real lane reached Implementer COMPLETED, validated ChangeSet and r2->r3 assembly, then failed closed at remediation on the new AEH-V2-0118 (repairers returned no-change after a PASS validation report; terminal FAILED with no durable error). The five Round-12 stale rows were marked `STALE_REQUIRES_RERUN` and their build-bound reruns are recorded in the Round-13 WorkGraph record. No row or slice is accepted by this record; S14 must not start.

## BROWSER lane

**Current status:** the actual Playwright Chromium runner is available and the maintained `tests/browser/` journey executes against the current candidate; the controller-owned paired decision journey and the required pause/resume controls remain pending (see below). Earlier paired-home and unpaired-readiness screens are infrastructure/projection observations only and do not satisfy the S9 gate.

No durable-operation browser journey is established yet. The external validator adapter and requirement mapping are deterministic/contract evidence only. Acceptance requires the actual pairing flow and Control Center UI to expose operation, candidate, policy, operationExecutionRevision, and controller epoch on the current decision surface, followed by the real choice/consume/resume/replay/cancel/refreshed-projection chain. Durable artifacts may contain only screenshots after fragment removal, assertions/observed states, and network method/path/status; pairing fragments, auth material, cookies, CSRF values, headers, and nonce are excluded.

The persisted `docs/evidence/s9/lead-gateway-failure.json` and `.png` record the actual fresh-start Lead Conversation configuration failure. `lead-gateway-passing-journey.json` records paired startup/home and a later message timeout only; it is not a complete journey. A safe scan of the current S9 JSON evidence found no credential-like string patterns.

**Reusable Playwright journey — 2026-09-26 (PASS):** the maintained `tests/browser/playwright.config.ts`, `tests/browser/s9-control-center-journey.e2e.ts`, and `tests/browser/fixture/` harness run the complete user-facing journey against the candidate release selected by `dist/current` (`release-1790386441389-339964-cb4c88d0`, gitSha `c03647107bc8b8e4464afd4d8a189ac80e7a01aa`, dirty, package 0.8.4). The fixture starts a real governed CHANGE operation through the public `npm run aeh -- operation start change` path with the file-scripted deterministic Paseo runtime (`AEH_DETERMINISTIC_PASEO_RUNTIME=1`), so the real controller owns the Spec Manager suspension, exactly-once consumption, continuation revalidation/resume, and a second suspension. Against the real pairing URL in candidate-pinned Playwright 1.62.1 Chromium 151.0.7922.34, the journey passed 46/46 assertions: paired navigation with fragment clearing, HttpOnly `SameSite=Strict` session establishment and one-use pairing replay rejection; authenticated operation and rendered current-scope projection (operation, candidate digest, policy digest, `operationExecutionRevision`, controller epoch, requestId); rendered-choice submission to a durable one-time `HumanDecision`; exactly-once controller consumption with replay/unknown-request/out-of-scope/foreign-operation/missing-CSRF rejection; continuation revalidation and resume; a second controller-owned suspension; pause through the UI to durable `PAUSED` with an empty drain receipt and stale-writer fencing; resume back to `HUMAN_REQUIRED`; scoped cancellation to durable `CANCELLED` with stale-writer and stale-decision rejection; and final API/rendered projection equality. Sanitized evidence: `docs/evidence/s9/s9-playwright-journey.json`, `s9-playwright-paired-home.png`, `s9-playwright-final-projection.png`.

**Browser lane boundary:** this is real-browser product evidence with a deterministic scripted provider boundary; it is not REAL_PROVIDER certification, and the scripted boundary is never enabled for production journeys. Independent slice review remains required before S9 acceptance.

**S11 candidate-bound browser/visual lane — 2026-09-26 (ROUND-2 REPAIRED; independently accepted by R3; checkpoint commit 78de9c5):** the first S11 submission was rejected and the round-1 repairs were verified by R2, which then rejected the slice on the declared BDD/gherkin route plus one LOW artifact-durability finding; both are repaired here. The first rejected submission's sanitized summary is preserved as `docs/evidence/s11/browser-visual-campaign.rejected-pre-repair.json`, and the round-1/round-2 history is retained in the S11 WorkGraph. The round-2 campaign ran `npm run test:browser-visual` (build `release-1790457526805-146084-3114256d`, package 0.8.4, gitSha `9b74c04167e79748312160b6476cc5050050b4e3`, dirty) against candidate `CAND-S11-BROWSER-VISUAL` r1, identity digest `7ad7d70b2ef1ac17c407b344d2c2811f3d58533699730380564bf1701d28cc44`, workspace source digest `7e791da17e1951ae96bf1c8c26438de4510c01cae1703370e2e4b3b0fdafcd42`. Each lane owns a candidate-scoped durable Playwright output directory (`.harness/evidence/playwright-s11-output/<lane>/<candidateId>-r<revision>-<identity12>`), so a later campaign run against a different candidate cannot delete a prior candidate's screenshots, and the campaign itself calls `verifyProviderLaneEvidenceV1` for both lanes after the full run and fails unless both are `ok`; the earlier `a20fd933b163` candidate evidence still verifies after this run. BROWSER evidence `.harness/evidence/browser/CAND-S11-BROWSER-VISUAL-r1-7ad7d70b2ef1/s11-browser-journey.json` digest `d98c455a87ddcff72214d4100fb0938628496e1407300ef3a783e3d800aef932` (provider `playwright` 1.62.1, command digest `69fd954a1be2a2b6ea3810f3ee388af0fc8310868c530bf6913a64eb5b9a361c`) and VISUAL evidence `.harness/evidence/visual/CAND-S11-BROWSER-VISUAL-r1-7ad7d70b2ef1/s11-visual-journey.json` digest `409aca8d6deb4758f55b437ae4a92c18a6943b4b9fca66d7d164099c2b9b6515` (provider `playwright-visual` 1.62.1, command digest `452a32a8b149e1797cabe39a1d57d394acd84d28766e3309275d2499bbf8f9f3`) both verified `ok=true` with a MATCH workspace binding and all artifacts present. VISUAL evidence binds the committed reference baseline artifact `tests/browser/s11-control-center-browser-visual.e2e.ts-snapshots/s11-control-center-heading-chromium-linux.png` (digest `21ec4e27bdb2aca29a3ed672c310e35533289f4636ddfe26f93badd8772565ee`) and the comparison configuration `playwright-toHaveScreenshot` `{ animations: disabled, maxDiffPixelRatio: 0.05 }`; a missing baseline fails closed with `VISUAL_REFERENCE_BASELINE_REQUIRED` and is never generated from the candidate. The machine summary is written inside the ignored evidence directory (`.harness/evidence/s11/browser-visual-campaign.json`); its pre/post-summary digest checks are stable and a fresh worktree-digest recomputation after the campaign equals the recorded candidate source digest. The R2 blocking BDD finding is repaired: `src/validators/registry.ts` routes `gherkin`/`bdd` through the capability provider, so configured BDD provider commands run, a missing runner returns typed `CONTRACT_PROVIDER_UNAVAILABLE` instead of the bare skip message, and a PASS persists candidate-bound CONTRACT evidence; `src/validators/gherkin.ts` is deleted and the maturity inventory cites `src/validators/registry.ts`. Focused S11 set 9 files / 57 tests; full non-system 177 files passed / 1 skipped (1161 passed / 7 skipped) with only the known `tests/paseoSdkResolve.test.ts` mise drift failing (3 tests, also failing at S10/S11 baseline); four deterministic system suites 4 files / 23 tests; `npm run typecheck` exit 0. Boundary unchanged: real built candidate, real Control Center UI/server/controller, real Chromium, scripted deterministic Paseo provider conversation; not REAL_PROVIDER, not a freshly packed product journey.

## ADVERSARIAL lane

**Status from the audit:** complete self-hosting adversarial/fault-injection evidence is not established.

Deterministic negative tests and fail-closed contract checks remain in the deterministic evidence map. They are not substituted for adversarial or fault-injection campaigns against the packed candidate, authority boundary, cancellation/takeover behavior, isolation, and external-effect reconciliation. S10 adds real isolation adversarial probes (network egress, out-of-scope writes, host PID/credential visibility, environment leakage, required-provider absence) executed against the actual sandbox rather than asserted by fixtures; packed adversarial/fault-injection campaigns remain open.

## PACKED_E2E lane

**Status from the audit:** partial campaign/bootstrap evidence exists; complete packed-candidate product journeys are not established.

Some historical campaign evidence exercised packed consumer behavior, but the all-packed campaign description was overbroad: some campaign tests import src/** directly. A packed bootstrap or a direct-source campaign is not proof that every product journey ran from a freshly packed candidate. Full packed journeys remain required for claimed capabilities, including human interaction, candidate mutation/repair, acceptance, delivery, provider lifecycle, and disposable self-modification.

## Reconciled claims

- **Authority and HumanDecision:** deterministic production consumers now cover exact action authorization, cancellation, and product-choice continuation. Requests and decisions bind current operation, candidate, execution revision, policy, actor, purpose/scope, and epoch; product choice cannot satisfy action authorization or operation control. Live OPA, browser, provider, and packed adversarial certification remain separate pending lanes.
- **Cancellation:** deterministic focused and system evidence now proves current controller ownership is claimed before cleanup, current operation/candidate/policy/revision/epoch fences are rechecked through cleanup, and unresolved action intents prevent terminal CANCELLED. Packed provider/fault-injection certification remains pending.
- **Validation, review, and acceptance:** these are separate gates. S4 runs or blocks impact-derived validation and creates actual independent Reviewer assignments with candidate-bound session evidence. S6 consumes current validated checks and exact Reviewer session provenance as evidence, requests bound managed Lead semantic evidence when frozen policy requires it, and has the controller persist the AcceptanceOracle disposition. Validation PASS, review agreement, participant completion, completion claims, and Lead opinion alone remain insufficient. Delivery and terminal success require the current oracle and objective gate.
- **CandidateImpact:** actual post-assembly impact is consumed by a deterministic assurance compiler against exact candidate and frozen policy identity. `requiresIndependentReview` causes an actual configured Reviewer assignment when required; it remains input evidence and grants no authority. S6 consumes the compiled candidate/impact/policy-bound assertions and evidence-strength floor without changing the S4 interface.
- **Repairer:** candidate-assembly repair support exists, so the public path is not accurately described as having no Repairer mutation support. S4 verifies the runTask validation-repair path and review quality-repair/rollback recompile impact-driven assurance against the resulting current candidate. S6 consumes only the resulting current candidate evidence; broader recovery/provider journeys remain pending.
- **StructuredResult provenance:** candidate/generation/epoch checks exist, and blueprint/policy checks are optional in some paths. Production launchers omit required full digests; do not describe provenance as wholly absent or complete.
- **ExecutionCatalog:** the catalog exists and is used. Its policy/tool authorization fields are not fully enforced; do not describe the catalog itself as absent.
- **Packed campaign:** some campaign tests import source directly. Do not label all campaign tests or all product journeys packed.
- **Serena:** the historical smoke covers the tested Serena broker path only, not provider-lease lifecycle or real-model certification.
- **Dated validation counts:** retain them only as dated checkpoint evidence in the ledger. They are not rerun results or current certification.
- **Delivery description:** the prior audit note about the PR description predates the S6 gate. Current delivery effects are accepted only after the persisted current AcceptanceOracle disposition; required managed Lead evidence precedes that disposition.

## S7 deterministic evidence — 2026-09-24

The focused provenance/build/delivery command (`npx vitest run tests/provenance.test.ts tests/provenanceS7.test.ts tests/architectureIntegrationContracts.test.ts tests/deliveryFinalize.test.ts --reporter=dot`) passed 4 files / 49 tests. `npx tsc -p tsconfig.json --noEmit` exited 0; `npm run build` exited 0 and produced `release-1790308464438-2905085-da2062aa`; JSON schema parsing and `git diff --check` passed. The final full non-system suite passed 158 files with 1 skipped and 965 tests with 7 skipped. The build-hygiene pack test emitted missing-source-map notices but passed.

The three relevant deterministic system suites ran 3 files: scenario and concurrency passed, while the adversarial suite had 2 failures (43 passed / 2 failed overall). Both failures are terminal-operation fixtures rejected with `OBJECTIVE_COMPLETION_REQUIRED` because they lack the objective-completion/current AcceptanceOracle evidence now required by S6. The S7-specific missing-manifest system case was isolated and passed (1 passed; 27 other tests skipped). No S6 lifecycle fix was made in S7. REAL_PROVIDER signing was not run because the `cosign` CLI is unavailable in this environment; fake Cosign/PATH-shim fixtures are deterministic contract evidence only. At the initial S7 close, Campaign Director acceptance remained pending; the schema remediation and subsequent independent acceptance are recorded below.

## S7 schema-guard remediation — 2026-09-24

The independent-review finding was reproduced before the schema edit in the Flash worker's separate Paseo worktree. The command npx vitest run tests/projectSchemaS7.test.ts --reporter=dot reported 1 failed file, 6 structural guard tests failed, and 15 tests passed. The failures identified absent parent required entries for provenance.required, sbom, signing, and verification plus absent nested required guards for their boolean flags. The 15 passing tests already confirmed both then constraints and the existing loadProjectConfig behavior. The worker also confirmed the 21-test suite passes against a throwaway schema copy with the full guards; the primary schema remained unchanged until the red run was reported.

The repair adds those guards to both allOf conditions and changes no runtime config contract. No AJV or other JSON Schema validator is installed or resolvable, so JSON Schema coverage is explicitly structural; the test suite uses the existing Zod loader to accept the shipped non-strict template, minimal/false-flag configurations, and valid strict configuration, and reject missing artifact/key cases. No validator dependency was added. Post-fix, npx vitest run tests/projectSchemaS7.test.ts --reporter=dot passed 1 file / 21 tests. The focused command npx vitest run tests/provenance.test.ts tests/provenanceS7.test.ts tests/projectSchemaS7.test.ts tests/architectureIntegrationContracts.test.ts tests/deliveryFinalize.test.ts --reporter=dot passed 5 files / 70 tests; npx tsc -p tsconfig.json --noEmit exited 0; npm run build exited 0 (release-1790309899170-2946073-c0075cc5); JSON schema parsing passed. The S7 system scenario matrix passed 8 tests and the adversarial missing-manifest case passed 1 test (27 skipped).

The previously recorded broader adversarial run's two terminal-operation fixture failures remain the known S6 objective-completion/current-AcceptanceOracle fixture issue. This remediation did not rerun or change those fixtures. The isolated S7 system checks passed. The Campaign Director independently reran the focused S7 set (5 files / 70 tests), TypeScript noEmit, production build (release-1790310426337-17-0e3b4825), full non-system Vitest (159 files passed, 1 skipped; 986 passed, 7 skipped), scenario matrix (8 passed), adversarial missing-manifest case (1 passed, 27 skipped), JSON schema parsing, and diff checks. REAL_PROVIDER signing remains unrun because cosign is unavailable; deterministic fake signing remains fixture evidence only. The Campaign Director accepted the S7 deterministic gate on 2026-09-24; broader external certification lanes remain pending.

## S8 deterministic evidence — 2026-09-25

The fresh S8 lead preserved the strict delivery boundary and did not change the normative target. The default issue implementation now prepares the existing issue and starts a managed controller run; controller-owned Paseo execution workspaces remain pre-acceptance local resources, bound to current S3 controller authority/fencing and durable action receipts. Public GitHub issue/branch creation, push, and pull-request creation remain behind the current AcceptanceOracle and policy-required S7 provenance gates.

The deterministic classification is explicit: paseo.workspace.create is LOCAL_RESOURCE_CREATION and is excluded from configuredExternalEffects(); it still passes ToolActionGate with the current controller epoch, CandidateRevision, and policy, persists intent/receipt, and retains conservative reconciliation. When provider identity cannot be confirmed, the action remains unresolved/HUMAN_REQUIRED. Executor exceptions for gated delivery actions trigger observation and only conclusive SUCCEEDED/FAILED observations persist reconciled receipts; UNKNOWN or failed inspection leaves the intent unresolved.

The default issue-flow regression omits autoHandoff (the public default), verifies internal workspace action before acceptance and zero GitHub POSTs until acceptance, then verifies accepted branch/push/PR operations have intents and receipts. GitHub was stubbed and the push target was a local bare repository. The focused set passed 6 files / 87 tests; npx tsc --noEmit exited 0; npm run build exited 0 (release-1790320485365-3155731-ff216f06); full non-system Vitest passed 161 files with 1 skipped and 1,012 tests with 7 skipped; affected deterministic scenario/concurrency suites passed 2 files / 17 tests. The first post-adjustment full non-system attempt had one unrelated operations cancellation fixture fail with CANDIDATE_WORKSPACE_MISMATCH; the isolated test and subsequent full-suite rerun passed. These results are deterministic contract/unit/integration/build evidence only. No REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E claim is made. The Campaign Director accepted the S8 deterministic gate on 2026-09-25. S9 had not started at acceptance.

### S8 direct CLI surface remediation — 2026-09-25

The remaining stale surface was `aeh sdd handoff <taskId>`, whose help promised GitHub issue/branch and Paseo worktree delivery while the handler called a managed-operation-only `handoffTask`. No supported accepted-delivery controller lifecycle was available to a normal CLI invocation. The command now advertises an unsupported/deferred state and returns `SDD_HANDOFF_UNSUPPORTED` before project configuration loading, with the current AcceptanceOracle and policy-gate rationale. It does not call `handoffTask` or create local/public resources. Managed finalization continues to own accepted delivery. The new `tests/cliS8Handoff.test.ts` passed 2/2 and checks help text, direct invocation diagnostics, config-load ordering, and an unchanged empty temporary directory. The existing direct handoff regression still checks that pre-acceptance invocation causes zero GitHub requests and creates no delivery intent.

Final lead verification: focused S8 action/delivery/issue/CLI set passed 14 files / 198 tests; `npx tsc --noEmit` exited 0; `npm run build` exited 0 (`release-1790322187860-3232428-0507f39b`); full non-system Vitest passed 162 files with 1 skipped (1,014 passed, 7 skipped); affected deterministic scenario/concurrency suites passed 2 files / 17 tests. The Campaign Director independently reran the focused S8 delivery/action/issue/CLI set (7 files / 89 tests), `npx tsc --noEmit`, `npm run build`, the full non-system suite (162 files passed, 1 skipped; 1,014 passed, 7 skipped), and affected scenario/concurrency suites (2 files / 17 tests); all passed. `git diff --check` and equality of `docs/CORE_ARCHITECTURE_V2.md` with HEAD passed. The Flash worker's baseline regression failed 2/2 against the stale command and passed 2/2 against the lead's CLI change. All evidence is deterministic fixture evidence; no REAL_PROVIDER, BROWSER, OPA_PROVIDER, ADVERSARIAL, or PACKED_E2E certification is claimed. **Campaign Director disposition: SLICE_ACCEPTED on 2026-09-25.**

## S9 R12 startup-failure remediation — 2026-09-25

The exact R11 packed runtime selected by the disposable Lead identity (`release-1790355907478-3853360-f5e934b6`) moved the operation to `RUNNING`, loaded config, resolved `baseRef=main`, and synchronized the portfolio before its catch boundary. One supported packed `operation execute` therefore exited with `No resolvable Git base ref found; configured baseRef=main.` after claiming epoch 2, leaving operation `CHANGE-20260925T171424Z-648c3147` at revision 8 `RUNNING/preparing` with no participants, DecisionRequest, or provider leases and a dead recorded PID. HEAD and that selected release share the pre-catch ordering; the current worktree repair moves the RUNNING transition and all bootstrap under the existing fenced failure-finalization catch. Unresolvable base refs remain explicit `FAILED` outcomes; no synthetic base or bypass was introduced.

The isolated Flash regression proved RED against baseline commit `c036471`: `npx vitest run tests/operationStartupFailure.test.ts` exited 1 with 1 failed at `resolveBaseRef` before the catch. Luna selectively integrated that single test file; current repaired-source run `./node_modules/.bin/vitest run tests/operationStartupFailure.test.ts --reporter=dot` passed 1 file / 1 test, and `npm run typecheck` exited 0. The Director independently ran the earlier test draft — 1 file / 1 test passed; typecheck exited 0; and build exited 0 (`release-1790359341925-17-d65e1b0f`). Fresh packed package `/tmp/aeh-s9-r12-pack/agentic-engineering-harness-0.8.4.tgz` has SHA-256 `77a0c39f580053b67abe2ee444e6b4c2995ecdc417a66613eb56d5dc20a1782d` and selects `release-1790359422074-34-41c17389`; read-only inspection confirms the catch encloses RUNNING patch, config loading, base-ref resolution, and portfolio synchronization.

The Director owns same-root Git initialization and fenced recovery of this operation. Stop at durable `HUMAN_REQUIRED` for independent UI choice control. No actual browser decision journey is claimed by this remediation record; the real Playwright assertions must verify current operation/candidate/policy/revision/epoch scope before the Director submits a choice. Pairing/auth secrets are excluded from durable evidence.

## S10 real isolation and candidate-bound SAST evidence — 2026-09-25

The S10 WorkGraph, frozen contracts, delegation assessment, and exact environment facts are in [S10 WorkGraph](CORE_ARCHITECTURE_V2_S10_WORKGRAPH.md). The pre-S10 gaps were independently reproduced by a bounded worker in an isolated Paseo worktree (`docs/evidence/s10/baseline-gaps.md`): no rootless execution provider was wired, isolation was only asserted through argument arrays, validator commands ran as host child processes, SAST artifacts were not candidate-scoped, and `opa`/SAST tools were absent.

The focused S10 suites passed 5 files / 30 tests: `tests/security/isolation.test.ts` (10), `tests/security/sastEvidence.test.ts` (8), `tests/security/validatorIsolationFailClosed.test.ts` (5), `tests/security/providerFailClosed.test.ts` (4), and `tests/system/aehIsolationCampaign.test.ts` (3). The system campaign executes the real `bwrap 0.13.0` provider and proves from inside the sandbox: host uid preserved without privilege, `CapEff=0`, `NoNewPrivs=1`, only `lo` in the new network namespace, a real connect to `1.1.1.1:80` failing with `ENETUNREACH`, a live host PID not visible, host `$HOME/.ssh` and sibling directories masked, `/etc` read-only, declared workspace writes reaching the host and out-of-scope writes not, an inherited host secret absent, and a required validator command unable to mutate the repository or reach the network. Deterministic fail-closed suites prove that an unavailable or unsupported isolation provider never runs a required validator and returns an explicit blocker.

`S10_TRIVY_BIN=/tmp/opencode/s10-tools/trivy/0.70.0/trivy node tests/packed/s10RealSastCampaign.mjs "$(pwd)"` ran against built release `release-1790389747122-486790-a8c37684` and passed all assertions. Candidate r1 with a planted GitHub PAT produced a real FAIL with one `github-pat` finding; candidate r2 with the secret removed produced a clean PASS. Both candidate-bound `SastEvidenceV1` artifacts verify by digest and bind the exact candidate/workspace source digest, tool version `0.70.0`, command digest, raw-output digest, findings, and isolation evidence; the r1 evidence is rejected as `SAST_EVIDENCE_STALE` for r2. Sanitized summary: `docs/evidence/s10/real-sast-campaign.json`. Trivy was provisioned user-locally from the official release with checksum verification (tarball SHA-256 `8b4376d5d6befe5c24d503f10ff136d9e0c49f9127a4279fd110b727929a5aa9`, binary SHA-256 `379d59f24a4a828c55de5f0b91b6805cc35d13580180b658820e648611256166`) and is not a committed dependency.

`npm run typecheck` and `npm run build` exited 0 (`release-1790389747122-486790-a8c37684`). The full non-system suite passed 175 files with 1 skipped (1,134 passed, 7 skipped) except `tests/paseoSdkResolve.test.ts` (3 failures), which was rerun 3/3 failing in a clean detached worktree at baseline `0afebcd` and is therefore environment drift from a mise-installed Paseo CLI, not an S10 regression. The scenario matrix, concurrency campaign, and isolation campaign system lanes passed 3 files / 20 tests; the adversarial suite passed 26 tests with its two long-recorded S6 objective-completion fixture failures unchanged. These are CONTRACT, UNIT, INTEGRATION, BUILD, SYSTEM_DETERMINISTIC, ADVERSARIAL, and local REAL_PROVIDER results. They do not establish hosted-provider certification, rootless Podman/OCI execution, live OPA, BROWSER, or complete PACKED_E2E certification.

## S12 observability and advisory eval evidence — 2026-09-26 (R1 repair applied)

The S12 WorkGraph, frozen contracts, delegation assessment, execution record, and the R1 repair record are in [S12 WorkGraph](CORE_ARCHITECTURE_V2_S12_WORKGRAPH.md). The implementing lead did not accept its own slice; a fresh independent reviewer rejected the first submission on gate B (F1 corpus/case identity did not bind the executable scenario harness; F2 the campaign's canonical metric digest was not reproducible across executions) and one LOW docs finding (F3 counts), and these repairs are recorded below. A second fresh independent reviewer (S12-REVIEW-R2) independently verified all three repairs and accepted S12 on 2026-09-26 (checkpoint commit ff1f420) across gate A (telemetry identity), gate B (local export reproducible), and gate C (evals advisory only).

`src/telemetry/identity.ts` defines `TelemetryCorrelationV1` (operation, candidate id/revision/source digest/identity digest, operation execution revision, policy digest, controller epoch, optional participant/generation/role, runtime name/session id, build release/digest), canonical `aeh.*` attribute keys, a domain-separated correlation digest, deterministic derive/resolve/verify helpers, and explicit `TELEMETRY_IDENTITY_MISMATCH`/`TELEMETRY_IDENTITY_UNRESOLVED` violation records. `recordEvent` resolves the current durable operation identity at emission time; a supplied identity that differs from current durable truth is recorded with a violation while the record keeps the current identity, so stale/wrong-candidate attribution is detectable and never silently accepted. `tests/telemetry/identity.test.ts` (5 tests) and `tests/telemetry/events.test.ts` (4 tests) cover derive, digest stability/separation, stale candidate/revision/policy/epoch/participant/session mismatch detection, unresolved identity, disabled telemetry, and trace-ID stability.

`src/telemetry/metrics.ts` creates real OpenTelemetry counters and histograms (`aeh.operation.count`, `aeh.operation.duration`, `aeh.participant.launch.count`, `aeh.participant.result.count`, `aeh.participant.turn.duration`, `aeh.runtime.session.count`, `aeh.validation.run.count`, `aeh.validation.duration`) and wires them into the production run finish, participant lifecycle, Paseo session binding, and validation finish. The local exporter serializes canonical sorted snapshots to `.harness/telemetry/metrics.ndjson` with no hosted credential. The R1 INFO hardening verifies a caller-supplied correlation against current durable truth before recording and records nothing on mismatch; a detached observation with no resolvable durable operation keeps the supplied identity and is documented as unverified. `tests/telemetry/metrics.test.ts` (4 tests) verifies instrument coverage, correlation attributes, digest reproducibility for a fixed scenario and candidate identity, no-op behavior without resolvable identity, and rejection of a stale correlation while the current correlation records. `tests/telemetry/export.test.ts` exercises the real OTLP/HTTP JSON exporter against `tests/helpers/otlpReceiver.ts`, an in-process loopback receiver, and verifies POST `/v1/traces`, `application/json`, resource build/service identity, stable trace IDs, parent span correlation, and equality between the OTLP span correlation digest and the NDJSON event identity digest.

The versioned corpus `evals/corpus/corpus.json` ships five cases (`validation-gated-change`, `validation-fails-closed`, `scope-governance`, `context-budget-projection`, `quality-convergence-thresholds`) with fixed fixtures and deterministic scenario harnesses over production deterministic subsystems. `src/evals/runner.ts` records `EvalCorpusIdentityV1` (`corpusId`, `corpusVersion`, corpus `digest`, per-case `caseDigest`) and the running `EvalBuildIdentityV1` on every result, and `src/evals/scoring.ts` adds a deterministic comparable projection. The R1 F1 repair binds each case's executable scenario harness closure — the declared scenario entry file plus every relative import reachable inside the project `evals/` tree (for example the shared `_result.ts` writer) — into the case and corpus digest; production `src/` code is bound separately by the recorded build identity; a missing declared scenario fails closed with `EVAL_CORPUS_SCENARIO_MISSING`. The digest rule is canonical: the unique file set is keyed by normalized repo-relative path, sorted, and hashed as `relativePath\0fileSha256` lines joined by newline, so an independent implementation reproduces it exactly. `tests/evals/corpus.test.ts` (5 tests) proves the manifest is shipped, the temp-copy corpus digest equals the repository corpus digest, a scenario-harness mutation changes the case/corpus digest while another case's digest is unchanged and restoring the file restores the digest, every case runs through the production runner with its expected status and a score ≥ 90 for PASS cases, build/corpus identity is recorded per result, and status/score/metrics/checks/corpus identity are identical across repeated runs.

`src/evals/advisoryInvariant.ts` and `tests/evals/advisoryInvariant.test.ts` (5 tests) prove the advisory invariant: no authority decision module imports eval/telemetry modules; observation read helpers are not referenced outside eval/telemetry/CLI surfaces; fabricated eval results, dashboards, telemetry events, and metric snapshots cannot change a `verifyTask` FAIL to PASS or alter its checks; durable operation bytes, candidate, execution revision, policy, and controller epoch are unchanged; forged eval acceptance claims are not loadable as an acceptance disposition and `requireAcceptedCurrentOracleV1` still throws `ACCEPTANCE_ORACLE_REQUIRED`; and deterministic route resolution is unchanged.

The deterministic campaign `npx tsx scripts/s12ObservabilityCampaign.ts` uses a fixed repo-relative ignored scenario root (`.harness/s12-campaign-scenario`, recreated per execution) so candidate/project/policy/correlation identity and the canonical metric snapshot digest are canonical across independent executions. Six independent executions in separate processes (two after the final artifact build) produced identical values: canonical metric snapshot digest `838305e6fba955d5a55d64f6252909130cf7b8f005eff8eb22388d16cd522878`, candidate identity `698e397a9d2f91bde5fa0a090a4c69062e1236bc33d1dc9c9f512d38d4128de3`, policy digest `b1c8f1dd4c74e952ded2a8537fce24ef1edefdb1b6a4f22bf828f9a20d0d2ecd`, participant correlation digest `20af5760d1f4b6c7a600a259e9cb0bdf9f3ab3bae84fe3be5d9f6ce6fed5d796`, base correlation digest `ad2d03c5d1d9427b70c21f686801de7d975fe842fb46129ab9e939ab654879f3`, corpus digest `11e9b972994e316c853893e9744109e146ffad96563ff3924dd8632b0e5485a6`. The regenerated artifact `docs/evidence/s12/s12-observability-evals-campaign.json` (SHA-256 `285e78ff74ca9bfe64bd3ece8d4018387e4508ba2597fbad94a9871f867398af`; built release `release-1790462333370-272879-9c0e6a52`, package 0.8.4, gitSha `9cdbbfe30b48f5cbcdd214724768e981cb368714`, dirty) binds the durable scenario identity, five identity-bound trace events with a single stable trace ID and a detected stale-identity violation, required metric instruments with two reproducible snapshot digests plus the canonical cross-execution digest, the verified local OTLP payload, all five eval case results with corpus/case/build digests and repeat reproducibility, and the advisory-invariant checks. Focused S12 set passed 9 files / 28 tests; affected wiring set 13 files / 126 tests; `npm run typecheck` exit 0; `npm run build` exit 0; the full non-system Vitest ran 186 files and 1,195 tests: 1,185 passed / 3 failed / 7 pending (file statuses 185 passed / 1 failed in this run; the 7 pending are the opt-in real-provider contracts in `tests/providerContracts.test.ts`, reported as skipped by some reporter views). The only failures are the 3 pre-existing `tests/paseoSdkResolve.test.ts` mise environment-drift tests unchanged from S10/S11. Scenario matrix, concurrency, and contract/integration lanes passed 3 files / 20 tests; the isolation campaign passed 1 file / 3 tests; the adversarial suite passed 26 tests with its two long-recorded S6 objective-completion fixture failures unchanged. These are CONTRACT, UNIT, INTEGRATION, BUILD, SYSTEM_DETERMINISTIC, and LOCAL_OTLP loopback results. No REAL_PROVIDER, hosted-collector, BROWSER, or PACKED_E2E certification is claimed.

## Certification rule

CertificationCore evaluates frozen policy against candidate/build-bound receipts and required evidence. It does not infer provider success from deterministic tests, browser success from adapter contracts, adversarial robustness from ordinary negative tests, or packed-candidate behavior from direct-source test imports. Each lane reports its own status and evidence.


The Campaign Director’s S3 source review identified two stale-binding recovery gaps in consumed-choice and checkpoint loading. On 2026-09-24, the Luna-owned remediation added a durable `selectedDecisionBinding`, exact full-binding comparisons in `loadConsumedProductChoice` and `loadOperationProductChoiceCheckpoint`, checkpoint identity rewrites only during the authorized same-epoch requirement/policy lifecycle, and fail-closed consumed-choice recovery. Regressions cover decision and current-operation execution revision, policy digest, and controller epoch drift; checkpoint envelope tampering with recomputed content digests; safe reissue of an intact unanswered request after takeover; and fresh HUMAN_REQUIRED reconfirmation after consumed-choice takeover. The bounded interruption after execution-semantics persistence is recovered by the controller’s existing bootstrap: it binds current policy before continuation loading, then verifies current continuation/checkpoint identity. If takeover changed the epoch, it retains the old decision binding only to verify the consumed receipt and opens a new current-binding HUMAN_REQUIRED request; it never rebinds the old choice. Final verification: focused S3 authority/action command passed 12 files / 167 tests; `npx tsc -p tsconfig.json --noEmit` exited 0; `npm run build` exited 0 (`release-1790285875947-2230303-2736c95d`); full non-system Vitest passed 318/318 suites, 819 tests passed, 7 pending, zero failed; the three S3 deterministic system suites passed 45 tests. `git diff --check` was clean. These are deterministic contract/unit/build/system results only; live OPA, browser, real-provider, and packed adversarial/provider takeover certification remain pending. The full-suite packed build-hygiene test emitted missing-sourcemap notices but passed. The Campaign Director independently accepted the S3 deterministic gate after reviewing the lifecycle implementation and rerunning the focused authority set (12 files / 157 tests), TypeScript check, and production build; OPA, browser, provider, and packed adversarial evidence remains unclaimed.

**S13 round-7 repair/re-run campaign — 2026-09-28 (implementing lead; `SLICE_BLOCKED`; fresh independent review owns acceptance):** round 7 repaired AEH-V2-0099 (`contractValidationRequirementsV1` compiles each frozen contract requirement's declared validators into exactly one candidate-bound ValidationRequirement whose id is the deterministic check id — configured `command.<id>` commands and configured validator ids — with `resolveConfigured` recognizing the convention, and the executed `candidate.assurance.validation.*` checks are merged into the report on every assurance recompilation; unknown ids skip and id collisions fail closed; no oracle gate changed) and AEH-V2-0100 (`recordParticipantReceipt` persists the receipt's accepted result artifact on the controller-issued `participant:<digest>` identity while keeping launch identities non-terminal and the session id as provenance, so the oracle's unchanged single-record predicate reads one coherent chain). Six bounded real packed `change-direct` executions against fresh packed candidates then newly reached and round 7 repaired four further deterministic defects with focused regressions: AEH-V2-0101 (`objectiveParticipantAccountingV1` derives completion of in-flight participants from the same candidate-bound settled receipt the successful-terminal gate consumes), AEH-V2-0102/0105 (one bounded supervisor-consolidation correction, total ≤2 turns, for schema-invalid or provenance-mismatched output with the schema/exact-set gates intact), AEH-V2-0103 (one current-candidate review evidence item per reviewer/dimension/assertion; multi-round collision fixed in the evidence producer) and AEH-V2-0104 (superseded review-round checks replaced by the current round so a remediated candidate's report status is not poisoned). The final run's AcceptanceOracle returned `ACCEPTED` (all three assertions covered, zero blockers, persisted acceptance artifact sha256 `c325323b156c131d6442aab407b6f0042e6ed1e262558c968723096311f487e9`; evidence bundle `7368ddfdc91597cea7dc6a80598df1c99f0f719dd119d5da65ff89e3cc710ded`; disposition `52b46f6710e5ff2715a3e6e698c0e49bfc0dcb872b341143ace99512de44dad9`), validation PASS, review `ACCEPTED/CONVERGED`, real `opencode-go/mimo-v2.6-flash` implementer/reviewer and `codex/gpt-6-luna` supervisor sessions, durable `ROUTE`/`STACK`/`CANDIDATE_IMPACT` assessments, packed artifact `8ed50f013d4984e4210145d7ca1cef6cbfbc1dd80558b2244dad7c903d84603d` / packed build `release-1790556382520-15303-fd958fbb` / source `e3dad34`, and `checkoutUntouched: true` (`bfee9f4bcf471394…` before == after, 668 tracked files). Objective completion then failed closed with one `PARTICIPANT_INCOMPLETE`: the implementer's receipt binds the pre-assembly candidate r2 while the current candidate is r3, the assembled result of that turn's ChangeSet; the same equality is enforced by `evaluateTerminalGate` (`CANDIDATE_MISMATCH`). This is **AEH-V2-0106**, newly discovered, reproduced by `tests/objectiveParticipantAccounting.test.ts`, and owned by round 8 (assembly-aware receipt binding or deterministic assembly-lineage verification; no gate weakened, no evidence re-bound). The lane bound was never raised: run 4 recorded `timeout: true` at 900 s with the cancel rejection and an explicit 600 s late-terminal wait, and the final run completed in 735 s. R6-F1/R6-F2/R6-F4/R6-F5 are addressed by the campaign's new terminal-artifact capture (terminal record, validation report, reviewer artifacts with sha256, acceptance artifact, semantic assessment cache metadata) and the per-campaign tracked-digest correction (`ca509c20808e2102…` pre-amendment; `da9e3497c38d76b7…` round-6 post-amendment; `bfee9f4bcf471394…` round 7); R6-F6 verified. R6-F3 (post-amendment build-bound sweep of the accepted rows) and every dependent FULL_OPERATION row remain open; `change`/`direct-change` remains NOT ACCEPTED and the full matrix is not exhausted. Round-7 focused validation: 42 files / 417 tests green; `npm run typecheck` exit 0; `npm run build` exit 0; `git diff --check` clean; TARGET byte-untouched; all disposable Paseo workspaces and `/tmp` staging archived; no credential persisted. S14 must not start. Machine record and per-run digests: `docs/CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md` (Execution record — round 7).

**S13 round-8 repair/re-run campaign — 2026-09-28 (implementing lead; `SLICE_BLOCKED`; fresh independent review owns acceptance):** round 8 closed the round-7 structural blocker and accepted the change trunk through a real packed lane. **AEH-V2-0106** — ASSEMBLING now records a deterministic `CandidateAssemblyReceiptV1` per successful candidate transition (base/assembled candidates, ChangeSet and patch digests, producer identity, and the settled source receipt's id and canonical digest when one exists; explicit rebases record the observed source base), and `resolveCandidateLineageReceiptV1` walks that durable chain from the current candidate to accept a work receipt only when the chain names the exact receipt and the receipt's observed candidate is on the walked ancestry. `objectiveParticipantAccountingV1` (factored into `src/operations/participantAccounting.ts`, shared with the terminal verifier) and the successful-terminal receipt decision consume the resolver; `evaluateTerminalGate` is unchanged and still validated per receipt binding; unsettled/unrelated/divergent/broken/tampered lineage fails closed; AEH-V2-0089 resumability and the one-coherent-chain invariant are preserved. The bounded packed `change-direct` lane then reached terminal **`SUCCEEDED`** (operation `CHANGE-20260928T013113Z-5de819ea`, 513 s inside the 900 s bound, packed artifact `c2312547b41b4266…`, build `release-1790559036086-103163-19047f5f`, source `e3dad34`) with AcceptanceOracle `ACCEPTED`, complete objective completion (zero blockers; implementer launch identity closed `COMPLETED` through the persisted r2→r3 assembly lineage), validation PASS (real `opencode-go/mimo-v2.6-flash` implementer/reviewer and `codex/gpt-6-luna` supervisor sessions), `checkoutProof.checkoutUntouched: true` (`948a0bdac7b84de8…`, 668 tracked files), retained acceptance-artifact bytes recomputing to `4263478a961036d2…`, and four workspaces archived with `remaining: []` (`governed-lanes/change-direct.json` `f1bddeff05744e02…`; persisted operation-record copy `change-direct-round8-operation.json` `8e19971981b536f2…`). **AEH-V2-0107** — the first `change-delegated` run escalated to FORMAL_SDD/ELEVATED while the frozen policy stayed DELEGATED/NONE and the Spec Manager failed `EXECUTION_POLICY_STALE`; `rebindEscalatedChangePolicy` now advances execution semantics (establishing a pre-escalation baseline when no digest exists), clears the superseded policy through the canonical API, and rebinds the escalated route/assurance before FORMAL_SDD participant launches, idempotently and without weakening any guard (`tests/specEscalationPolicyRebind.test.ts`, 3). **AEH-V2-0109 / R7-F2/F4** — the campaign scopes checks/rows to the observed route, retains per-lane checkout proof, workspace inventory plus archive/query results, recomputable acceptance-artifact bytes and `candidateAssemblyReceipts`; `tests/changeControllerOwnedDecision.test.ts` re-ran green in this environment (R7-F3). Newly discovered and OPEN: **AEH-V2-0108** — the build-bound `audit` sweep run hung 20+ minutes in consolidation on a real supervisor continuation after one bounded correction, with the watchdog only observing `supervisor-busy`; the operation exceeded the 900 s bound plus a 600 s late wait and was still `RUNNING` (`governed-lanes/audit.json` `004baabdb0e18e52…`), so a bounded supervisor-turn deadline with typed fail-closed terminalization is required. **AEH-V2-0110** — the `change-delegated` rerun failed run-internal `planning` in ~17 s with no Implementer participant and no ChangeSet; the fixture validation failed on the unchanged tree and two bounded Repairers produced zero ChangeSets (`governed-lanes/change-delegated.json` `0abdbae41a8b98d7…`), so the DELEGATED planning path requires classification and repair. Round-8 focused validation: 43 files / 345 tests green; `npm run typecheck` exit 0; `npm run build` exit 0 (`release-1790560487803-148242-437c07ef`); `git diff --check` clean; TARGET byte-untouched (`4954b62bfb3a820f…`); every round-8 lane workspace archived with query evidence and all `/tmp/aeh-s13-gov-*` staging removed; no credential persisted. The post-amendment build-bound sweep (R6-F3) is not complete (`audit` failed on AEH-V2-0108; `startup`, `informational`, `project-home`, `multi-project`, `cancel`, `recovery`, `authority` and managed issue import not re-run); dependent FULL_OPERATION rows, the full `issue-driven` contract and `delivery` remain blocked; `control-center` remains BROWSER-only; AEH-V2-0077 remains an external user action. S14 must not start.

**S13 round-9 provider/effect conformance (implementing lead report; subsequently independently reviewed `SLICE_BLOCKED`):** the final-build R6-F3 sweep passed for `startup`, `informational`, `project-home`, `multi-project`, `cancel`, `recovery`, and `audit`. Audit reached `SUCCEEDED`/PASS in 230.784 s with the 300 s supervisor continuation ceiling; its lane artifact binds packed build `release-1790567348887-372966-19891328` (build digest `8e949b0113ddd90ae0573826b8ecc348915ae2c65fd6a77c99f82ba9b6a9ab84`). Round-9 GitHub `authority` and `issue.operation-import` PASS claims are tainted by AEH-V2-0113(A), because that campaign created an unauthorized repository; delivery was blocked.

R8-F1/R8-F2 are corrected in the governed-campaign emitter: the route-specific `applicableChecks` determine the lane aggregate and `candidateAssemblyReceipts` are projected into each lane envelope. The final direct lane carries these projections but is **not accepted**: it timed out with repeated Reviewer `EMPTY_OUTPUT`, and the campaign cleanup archived workspaces while its controller remained alive. The later candidate workspace mismatch is cleanup-race evidence, not product acceptance. AEH-V2-0112 adds a guard to avoid archiving until terminal/exited and requires a clean rerun.

AEH-V2-0110 remains unreconciled. The retained clean-run Planner artifact is schema-valid but says `formalizationNeed: NONE`, while the operation route is FORMAL_SDD; the Explorer payload needed to explain that route was not retained. A prior transient inspection appeared to show REQUIRED, but that artifact was discarded and is not evidence. The formal lane remains blocked: the Spec Manager had no accepted structured result, terminal `SPEC_MANAGER_RESULT_ARTIFACT_MISSING`; its session reported a CandidateRevision workspace mismatch, preserved as an agent-reported explanation only. The controller-owned OpenSpec content persistence path preserves the Spec Manager’s read-only authority and has focused unit coverage, but live end-to-end acceptance is pending. AEH-V2-0111 tracks the candidate/result identity failure. Full issue-driven delivery, multi-worker, repair-family, context-handoff, permission-delegation, and distributed-execution rows remain unexecuted or blocked. S14 must not start. See [S13 WorkGraph](CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md), “Execution record — round 9”, and the machine evidence under `docs/evidence/s13/`.

**Round-9 governance incident AEH-V2-0113 disposition:** the campaign created private repo `JamesMorales04/aeh-s13-round3-20260928032308-4317d7` without a required pre-authorized target. Its HTTP 403 deletion failure is `USER_CLEANUP_NON_BLOCKING`; the user will delete it manually. No deletion retry occurred. Only the Round-9 authority and operation-bound issue-import claims are tainted; unrelated S13 evidence remains independently classified.

**S13 Round-11 repair/conformance (implementing lead; `SLICE_BLOCKED`; fresh independent review owns acceptance):** Round 11 repaired the AEH-V2-0114 policy violation without touching product source: `tests/packed/s13ExternalEffectAuthorization.mjs` is the only producer of `ACTION_AUTHORIZATION` approvals for the GitHub campaign and derives them exclusively from an externally supplied user authorization file (`S13_GH_AUTHORIZATION_FILE`), verifying owner, PRIVATE_ONLY scope, machine-readable effect coverage (`scope.allowedEffects`), resource namespace (`scope.resourceNamePrefix`) and the anti-self-authorization governance flag; missing/unreadable/invalid/out-of-scope/uncovered inputs fail closed with typed errors before any `gh` command, staging directory or approval, and a private `Symbol` brand makes the validated object unforgeable. The amended prospective artifact (`docs/evidence/s13/github-effect-authorization.json`, sha256 `272c5cb7…`) encodes the same user grant; repository create/delete and issue edit/delete remain uncovered, and no production ban is hardcoded. The current-build sweep is complete against packed build `0fc216b8…`: capability certification 4/4 PASS (`s13-real-provider-certification.json` `1ab94ea7…`, `checkoutUntouched: true`) and governed `audit` PASS (`f31a34f8…`, summary `370b53a2…`, terminal `SUCCEEDED`, oracle PASS, one workspace archived / zero remaining). The user-authorized GitHub rerun reuses existing issue #1: `issue-driven` PASS (`49ac06e1…`, snapshot `bd777eed…`, drift gate PASS, operation-bound import completed) and `authority` PASS (`aaa8ea76…`, external-authorization provenance check PASS, `HUMAN_REQUIRED` without approval, `EXECUTED/SUCCEEDED` with it, real remote branch observed); delivery remains BLOCKED for lack of an accepted candidate. Every lane envelope records `harnessRevisions` (R10-F6 closed). The FORMAL_SDD live diagnostic reached a READY Spec Manager structured result (`0f85afd2…`, payload `66ba2766…`) and then failed OpenSpec compilation because the persisted `specs` content lacked delta headers — recorded as AEH-V2-0115 with the exact canonicality repair; AEH-V2-0110/0111 stay open. See [S13 WorkGraph](CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md), “Execution record — round 11”.

**S13 Round-10 reviewer disposition and evidence (corrected in Round 11):** R9-F1A's repository-target cause is fixed with an explicit fail-closed repository input and a regression that proves no `gh` command or staging starts without it. **AEH-V2-0114 correction:** the attempt-4 authority and operation-bound issue-import claims are `NOT_ACCEPTED (tainted)` — the campaign fabricated the approving HumanDecision (`actorId: "human:s13-github"`) for both `github.issue.create` and `git.push`, issues #2/#3 were created after an explicit pre-effect user denial, and the attempt-4 remote-ref observation, though genuine, was approved by that fabricated decision. The attempt-4 artifacts remain historical bytes. Round 11 removes every campaign-generated approval and consumes an externally supplied typed user authorization artifact that must cover the exact effect and fails closed before any `gh` command or staging; the default lane reuses the user-authorized existing issue #1. See `docs/evidence/s13/review-round10.json` (R10-F1/F2/F3) and the Round-11 WorkGraph. Delivery remains unexercised/blocked for the issue-driven fixture.

Round-10 lifecycle emitters now record round/run-specific artifact paths and packed build identity. `s13-real-paseo-lifecycle-round-10-lifecycle-round10-1.json` (SHA-256 `7e285c57463827e7a92d5358f925ccd6bbc932d90bd2330699e354e6549ce3ff`) reports cancel/recovery PASS against build digest `0fc216b8ebd40b942c19556b2f1e7d6c44be802bb27142d41deeab2cc7ee71ac`; the prior lifecycle artifact remains unchanged.

The Round-10 packed `change-direct` lane terminalized `SUCCEEDED` with AcceptanceOracle `ACCEPTED`, validation PASS, and route-scoped `change`/`direct-change` rows PASS. Its candidate workspace digest matched before cleanup; four workspaces were archived after controller exit, with a persisted post-campaign operation read. The OpenSpec candidate-preparation fix is unit-regressed, while live FORMAL_SDD result acceptance and the Round-9 route trigger remain open. The frozen 22-row capability plan is provider-path classification, not current row status; see the Round-10 WorkGraph for the current evidence table. Overall S13 remains `SLICE_BLOCKED`; no self-acceptance and S14 must not start.

**S13 round-15 conformance record — 2026-09-29 (implementing lead; `SLICE_BLOCKED`; independent review owns acceptance):** round 15 reconciled the unreviewed Round-14 source edits (all retained/completed with focused tests; none reverted), completed the reviewer-failure ownership repair and reran the packed `change-delegated` lane. **AEH-V2-0119:** a Reviewer provider/runtime stop, unparseable contract, or candidate mutation attempt now terminalizes immediately as a typed `SYSTEM_FAILURE` (`agent.system-failure` check naming reviewer, kind, session id and provider detail) instead of entering the autonomous remediation loop as synthetic debt; the reviewer prompt states its authorized read boundary plus changed files/requirements; the control-root `external_directory` read projection for write-denied roles and the deterministic remediation budget are retained. **AEH-V2-0120:** supervisor generations rotate on candidate drift (positive current-policy proof; bounded consolidation correction with delivered-turn evidence; turn timeout), and every terminal review result carries a quality state so the run summary cannot crash. **AEH-V2-0121:** the governed fixture now provides a real `check:architecture` validation for the canonical architecture impact dimension, and the campaign binds a real Paseo Lead agent (`aeh.provider` label, deleted after the lane) through `PASEO_AGENT_ID` so headless DELEGATED operations can obtain the Lead semantic evidence the AcceptanceOracle requires; no oracle or validator gate was weakened. Build-bound lane results on the current source: run 1 FAILED on `UNRESOLVED_VALIDATION_REQUIREMENT: impact-review-architecture` (fixture gap, now fixed); run 2 FAILED on the reviewer `external_directory` stop whose terminal path then crashed on the missing quality state (now fixed); run 3 reached `review=ACCEPTED` and failed closed only at `ACCEPTANCE_LEAD_REQUIRED` (headless Lead binding, now wired). A second machine restart (~2026-09-29T02:46Z) interrupted lane attempt 4 and cleared its fixture; no terminal state is claimed for it. Evidence: `docs/evidence/s13/round15-environment-loss.json`, `governed-lanes/round-15/run-20260929-r15-delegated-{1,2,3}/change-delegated.json`, summaries `s13-governed-operations-round-15-run-*.json`. These are REAL_PROVIDER + PACKED_E2E results for a non-terminal lane; `delegated-change` remains NOT_ACCEPTED and S13 stays uncommitted. S14 must not start.

**S13 round-15 continuation (2026-09-29):** after three deterministic repairs the packed `change-delegated` lane reached a real terminal `SUCCEEDED`: **AEH-V2-0122** (the managed Lead continuation no longer requests an AEH structured-result turn for an interactive Lead session with no result channel), **AEH-V2-0123** (resolved plan-authored base validation requirements are executed into candidate-bound report checks when no matching check exists, and the Lead prompt projects the resolved `validationRequirement(kind)=status` evidence per assertion), and **AEH-V2-0124** (settled `SUCCEEDED` receipts of non-producing roles bound to an ancestor candidate of the verified assembly chain are completion evidence, and a retried launch identity's transient `FAILED` no longer shadows its settled receipt). Final lane: AcceptanceOracle `ACCEPTED` (0 blockers), Lead evidence PASS, objective completion complete, assembly receipt r2->r3, 6 real Paseo sessions, `checkoutUntouched: true`, 4 workspaces archived / 0 remaining; lane sha256 `1ea0e5f093ee9403cccf925e6e0a779e794f87557a38a4d1a422649a2561ac09`; summary sha256 `22efabf50786ab102ee704ede0efbbb582e2ad3355236cf01625321d61e8e282`. This is an implementing-lead result; independent acceptance is pending and `delegated-change` is not self-accepted.

**S13 round-15 final lanes (2026-09-29):** the packed `change-direct` lane reached a real terminal `SUCCEEDED` on the current source (AcceptanceOracle `ACCEPTED`, 0 blockers, rows `change`+`direct-change` PASS, 3 real Paseo sessions, `checkoutUntouched: true`, 4 workspaces archived / 0 remaining; lane sha256 `4865e915b344b4eab5b54e91c284960683d81316eb150886604bbbf8965948a6`), and the GitHub campaign under the external user authorization (`docs/evidence/s13/github-effect-authorization.json`) with the user-authorized existing issue #1 passed `issue-driven` (snapshot/`ISSUE_DRIFT` gate PASS, operation-bound import completed) and `authority` (unapproved `HUMAN_REQUIRED`, approved `EXECUTED/SUCCEEDED`, real branch `aeh-s13-authority-20260929043030`) with no issue/repo created, edited or deleted; overall `SLICE_BLOCKED` on `delivery`. The FORMAL_SDD live lane now passes spec-authoring, canonical persistence and OpenSpec compilation (with the new pre-persistence SHALL/MUST normative gate, AEH-V2-0111-residual, and the explicit `AEH_RESULT_JSON=` marker instruction, AEH-V2-0125) and is blocked only by **AEH-V2-0126** (OPEN): authoring persistence into the candidate workspace changes its digest without a deterministic candidate advancement, so the Planner result is rejected `CANDIDATE_WORKSPACE_MISMATCH` (run 3 sha256 `6f134a4e7346df06…`). Round-15 final validation: 18 focused files / 152 tests green, `npm run typecheck` and `npm run build` exit 0, `git diff --check` clean, TARGET byte-identical, all lane workspaces archived, no credential persisted.

**S13 round-17 conformance record — 2026-09-29 (implementing lead; `SLICE_BLOCKED`; independent review owns acceptance):** round 17 reconstructed the interrupted Round-16 record from durable artifacts, verified the campaign formal stage-shape oracle fix (`formal.spec-authoring`/`formal.spec-compilation` now read the durable stage object's `status`), and implemented the missing `delivery` capability. Product change **AEH-V2-0127**: `delivery.github.pullRequests` (default true) enumerates the requested effect set in frozen policy (TARGET 11); when false, `configuredExternalEffects` omits `github.pull-request.create` and `finalizeAcceptedIssue` returns a push-only `FINALIZED` after the gated push. This is required by the external authorization chain: the user authorization covers `git.push`, `github.branch.create` and `github.issue.create` only, and no self-amended authorization was minted. The real GitHub lane (`r17-gh-4`) drives an accepted candidate r2 with a persisted accepted AcceptanceOracle through: frozen policy `[git.push, github.branch.create]` → denied `HUMAN_REQUIRED` → external-authorization-derived `APPROVE` bound to the exact `git.push` effect digest → branch create + commit + push → `FINALIZED` → remote ref `aeh-s13-delivery-1-s13-delivery-certification-change-r17-gh-4@5adaddcf297b` observed, with 3 intents/3 receipts, 0 pull requests and issue #1 unchanged (`docs/evidence/s13/github-lanes/round-17/r17-gh-4/delivery.json` sha256 `2b8dc82d9cdae78a891f7cecb2b2f66cfe91ccc78d2cad314eaf03ac47e4535b`); the GitHub summary is `PASS` (issue-driven + authority + delivery). The re-freeze (`5d1b6680…`, 668 files, build `release-1790684324214-110138-2fd31d7e`) supersedes `2bee2ec0…`, and every current row was re-run on it: `change-formal` (`r17-formal-3` terminal SUCCEEDED, oracle PASS, AcceptanceOracle `ACCEPTED`), `change`/`direct-change` (`r17-direct-1` SUCCEEDED), `delegated-change` (`r17-delegated-1` DELEGATED with 0 escalations, terminal SUCCEEDED), `audit` (`r17-audit-2` SUCCEEDED), certification 4 (`r17-certification-2` 4/4 PASS), `cancel`/`recovery` (`r17-lifecycle-2` 2/2 PASS), all with `checkoutUntouched: true` and archived workspaces. `r17-formal-2` and `r17-audit-1` are preserved truthful intermediate failures (Lead `REQUIRES_PRODUCT_DECISION` on validator coverage; supervisor consolidation contract) with no gate weakened. Remaining conformance gaps: full issue-driven end-to-end (the authorized issue #1 is not implementation-grade and the authorized repository has no configured validation), `multi-worker`, repair family, `context-handoff`, `permission-delegation`, `distributed-execution` (`BLOCKED_NO_EVIDENCE`; no harness fabricated), and 3 pre-existing `tests/paseoSdkResolve` environment-drift failures in the non-system suite. See [S13 WorkGraph](CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md), “Execution record — round 17”. S13 stays uncommitted; S14 must not start.

**S13 round-18 conformance record — 2026-09-29 (fresh lead; `SLICE_BLOCKED`; independent review owns acceptance):** round 18 validated the current tree (typecheck/build/diff-check clean; tracked digest `9e36dcb5…`, 668 files; zero `src/`/`schemas/`/`presets/`/`templates/` edits after the Round-17 freeze, so freeze `5d1b6680…` stays current) and reproduced the canonical full non-system suite exactly (203 `*.test.ts` files: 201 passed / 1 skipped / 1 failed; 1,357 tests: 1,347 passed / 7 skipped / 3 failed; the only failures are the pre-existing `tests/paseoSdkResolve` mise/Paseo-CLI environment drift, classified with exact observed paths and unmodified fixture/source files). Missing capability rows: **`certification-repair` is CURRENT_ACCEPTED** in a real-provider lane (failure packet from a deterministic oracle failure, a real `codex/gpt-6-luna` actor receipt and one bounded real repair turn, oracle re-check PASS, candidate tree digest unchanged; `r18-certrepair-4`). **`multi-worker` is blocked by the newly discovered AEH-V2-0129** (`workflow.planning.worktreeIsolation` declared but never consumed; same-wave parallel implementers share one workspace and the second launch fails closed `PASEO_PROVIDER_LEASE_TAKEOVER_BLOCKED`), after the frozen Planner produced the qualifying two-unit task DAG (`workUnits=2, waves=1`) and the single-unit control run reached terminal `SUCCEEDED`. **AEH-V2-0128** records the semantic-vs-deterministic delegation-floor disagreement that fails concrete multi-file DIRECT-classified requests with `EXECUTION_POLICY_STALE`. `repair`/`product-repair` stayed unevidenced because the real implementer satisfied the seeded failing validator in one turn (forcing requirement recorded); `distributed-execution` is harness-built but blocked by AEH-V2-0129; `context-handoff`/`permission-delegation` have no harness and exact dependencies recorded. The `issue-driven` full-journey lane (P0-A option b) is built and exercised against the user-authorized private repository and existing issue #1; `r18-journey-4` reached implementation COMPLETED and validation/review RUNNING but was non-terminal at the 900 s bound, so no delivery effect occurred. P1-C maps the `[delivery-record, provenance]` tokens in the delivery lane; P1-D record corrections are applied; AEH-V2-0118 is closed. Real E2E verification lanes remain separate from the deterministic gates; no gate was weakened and no slice acceptance is claimed. See [S13 WorkGraph](CORE_ARCHITECTURE_V2_S13_WORKGRAPH.md), “Execution record — round 18”. S14 must not start.

**S13 rounds 19–20 conformance record — 2026-09-29/30 (fresh leads; `SLICE_BLOCKED`; independent review owns acceptance):** the product freeze **F1** (`trackedDigest 1183af86…`, 668 files, buildDigest `408b178d…`, HEAD `e3dad34`) supersedes the Round-17 freeze for every row whose executed path includes the AEH-V2-0129/0128-repaired modules; no `src/**` change has occurred since. Round 19 delivered on F1: `multi-worker` **PASS** (`r19-mw-1`: two work units, one wave, zero conflicts, wave-barrier PASS, per-unit assembly receipts, accounting 6/6/0), `repair`+`product-repair` **PASS** through the deterministic forcing geometry (`r19-gov-3`: repair packet with 6 failures, Repairer repair in scope, re-validation/re-review, oracle re-check ACCEPTED), `change-multifile` **PASS** on the AEH-V2-0128 geometry, `audit`/`change-delegated`/`change-formal`/`change-direct` **PASS**, certification 4/4 **PASS**, cancel/recovery 2/2 **PASS**, and GitHub `issue-driven`+`authority`+`delivery` **PASS** reusing existing issue #1 with the external-authorization loader (no issue/repo create, edit or delete). Round 20 delivered the **full `issue-driven` journey PASS** (`r20-journey-8`): real issue snapshot/drift → real governed change operation terminal SUCCEEDED with oracle ACCEPTED → human rejection boundary → external-authorization-derived approval bound to the exact effect digest → real branch push with receipts/reconciliation → issue untouched, workspaces 6/4/2 accounted, checkout untouched. The journey's delivery half runs through a fixture-scoped RUNNING delivery-operation shell that preserves the real terminal operation's identity and copies its persisted acceptance artifact byte-for-byte (disclosed in the lane evidence); the real operation stays terminal and no epoch is re-claimed. **P0-B `distributed-execution` remains open:** the repaired harness (`37b7615a`→`4fc7b11c`) starts the worker at the advanced candidate worktree with the binary from the fixture root; `r20-dist-4` claimed and ran a real job but returned a FAIL receipt with **AEH-V2-0131** (`transport: "inherit"` has no approved pre-prompt session materialization boundary; only `direct`/`podman` are supported), and the operation stayed RUNNING past the bound. **AEH-V2-0130** (campaign resource lifecycle) is open with source-level evidence (`terminalizeOperation` has no resource reconciliation; only CANCELLED reconciles; participant/lead agent records and operation workspaces have no product archive path; no recovery orphan sweep), a runnable bounded stability campaign, and the exact repair plan; two stalled non-terminal distributed operations are preserved and reported. P1-D confirmed the `[delivery-record, provenance]` mapping from `r19-gh-1`; P1-E reproduced the canonical non-system suite (203 files / 1,361 tests: 1,351 passed / 3 failed / 7 pending; the only failures are the pre-existing `tests/paseoSdkResolve` drift) and a focused 25-file / 207-test S13 set; typecheck/build exit 0; `git diff --check` clean; TARGET byte-identical; no gate weakened and no slice self-acceptance. S14 must not start and its disposable self-hosting composite must include resource cleanup/reconciliation.

**S13 round-21 conformance record — 2026-09-30 (fresh lead; `SLICE_BLOCKED`; independent review owns acceptance):** round 21 closed the three remaining P0 rows on the **F2** product freeze (`src/distributed/{types,worker}.ts`, `src/workers/agentPrompt.ts`, `src/workers/directProcess.ts`; trackedDigest `eca252d6…`, build `release-1790743087471-1508623-8ef2a92d`, buildDigest `599ed363…`, HEAD `e3dad34`, TARGET sha256 `4954b62b…`). **`distributed-execution` PASS** (`r21-dist-14`: terminal SUCCEEDED rev 137, oracle ACCEPTED, lease prepared=2/released=2/completed=2, real worker session `ses_f0faef79affe4CUexX9g65Hk4T`, controller-recorded candidate-bound participant receipt, `checkoutUntouched: true`); the 13 preceding attempts are preserved truthful failures of the same provider assessor-stream transient plus the final worker-clone `ENOENT` gap that the F2 repair closes. **`context-handoff` + `permission-delegation` PASS** (`r21-ctxperm-11`, harness `c6806c78…`): the round-20 blocker was four harness defects (MCP entrypoint, participant identity convention, continuation manifest reuse, evidence reads), all fixed without touching a product gate; the lanes produce a real provider session, a durable grant bound to the exact binding/session, nine addressable refs, a durable `ContextContinuationV1` (sequence 2, previous session = the launch session), the monotonic lease subset and a durable permission decision receipt. **`issue-driven`/delivery PASS with a terminal delivery operation** (`r21-journey-3`): governed operation SUCCEEDED → authorized issue #1 snapshot/drift → `HUMAN_REQUIRED` → external-authorization-bound approval → real push `aeh-s13-issue-journey-1-issue-journey-greet-contract-comment-r21-journey-3@c4b62e45cdea` with `git.commit`/`git.push`/`github.branch.create` intents and receipts → delivery shell terminalized through `transitionOperationToTerminal` (SUCCEEDED rev 82); the Director R20-F7 RUNNING-shell objection is resolved by the product path; the prior `r21-journey-1` (supervisor consolidation 300 s timeout) and `r21-journey-2` (non-fast-forward push against the round-20 branch; branch pattern is now run-scoped) are truthful failures. **AEH-V2-0130:** R20-F1 fixed (terminality-gated reconciliation; `reconciliation-terminal-only` PASS) and the bounded stability pass `r21-resource-1` produced truthful `RESOURCE_ACCUMULATION_OBSERVED` evidence (terminal SUCCEEDED iteration residue 4 workspaces + 4 idle agents + 28 provider/MCP processes, ceilings 0; second iteration grew; no-monotonic-growth FAIL); 17 orphaned Paseo-managed opencode servers (~5.3 GB) with proven AEH ownership were reclaimed and their agents archived; the durable ownership/reconciliation architecture remains required and is handed to S14. Deterministic validation: focused S13 set 27 files / 242 tests passed; canonical non-system suite 203 files / 1,363 tests (1,352 passed / 4 failed / 7 pending) — only the pre-existing `tests/paseoSdkResolve` environment drift (3) and one `buildHygiene` concurrency artifact that passes in isolation (6 tests); `npm run typecheck`/`npm run build` exit 0; `git diff --check` clean. Record fixes R20-F2/F3/F5/F6/F8 are applied as recorded in STATUS and the ledger. No gate was weakened and no slice self-acceptance is claimed; S13 stays uncommitted and S14 must not start.
**S13 rounds 22–23 conformance record — 2026-09-30 (fresh leads; `SLICE_BLOCKED` pending the fresh final independent review; independent review owns acceptance):** round 22 re-ran the final-review stale rows on **F2-final** (`trackedDigest 3f9b7fd8…`; buildDigest `599ed363223ff2dd85e17296c9f34c4134bf83e363916c4c57d7e7492ccbdf45`; HEAD `e3dad34`): certification `startup`/`informational`/`project-home`/`multi-project` 4/4, lifecycle `cancel`/`recovery` 2/2, GitHub `issue-driven` (operation-bound `issue import` on F2, FINAL-F5) / `authority` / `delivery` 3/3, full `issue-driven` journey with a product-terminalized delivery shell and re-read delivery-record digest (`r22-journey-1`, FINAL-F6), `context-handoff` + `permission-delegation` 2/2 (FINAL-F2), `certification-repair` PASS, governed `audit`/`change`/`direct-change`/`delegated-change` (`r22-gov-1`), `multi-worker` + `formal-sdd` (`r22-gov-2`), and `distributed-execution` F2-current (`r21-dist-14`). Round 23 closed the only remaining rows — **`repair` + `product-repair` terminal `SUCCEEDED` / AcceptanceOracle `ACCEPTED` on F2-final (`r23-gov-9`)**: first-pass implementation FAILED on the seeded `src/config.mjs` FAREWELL defect → canonical `RepairPacket` (`CHANGE-20260930T161148Z-cf61361e-1.json`) → canonical Repairer COMPLETED → re-validation/re-review `PASS` → oracle re-check `ACCEPTED`; packed build `release-1790784657836-398208-ea436d52` (buildDigest `599ed363…`), 8 real sessions, `checkoutUntouched: true`, 7 workspaces archived with complete accounting (lane `20621848…`, summary `6ce06bc0…`). FINAL-F3 is corrected (`3f9b7fd8…` recomputed), FINAL-F4 is closed by the preserved byte-identical release copy (`docs/evidence/s13/round-22/frozen-release/`; 892-file manifest all MATCH; the copy still recomputes to `599ed363…` after suite builds), FINAL-F5/F6 are cited from `r22-gh-1`/`r22-journey-1`. Diagnoses recorded: the `r22-gov-4`/`r23-gov-5` formalization escalation was a legitimate Planner `REQUIREMENT_CONTRADICTION` fixed by fixture-only request wording (no product escalation suppressed); the `r22-gov-3`/`r23-gov-6` review stall is an open bounded-participant liveness observation (`AEH-V2-0132`); `r23-gov-8` exposed superseded-attempt objective-completion accounting (`AEH-V2-0133`); the `r22-gov-1` `AEH_RESULT_PROVENANCE` intermittent is an open observation (`AEH-V2-0134`). The repair lane ran on a disclosed fixture-level brain-model override (`opencode`/`opencode-go`/`gpt-6-luna`) because the default codex ChatGPT account was usage-limit blocked; the override is recorded in `fixtureTopology` in the lane and summary. Deterministic validation: focused S13 set 27 files / 242 tests passed; canonical non-system suite 203 files / 1,363 tests (1,353 passed / 3 failed / 7 skipped; only the pre-existing `tests/paseoSdkResolve` drift, `buildHygiene` passed); `npm run typecheck` and `npm run build` exit 0 (fresh `599ed363…` release); `git diff --check` clean; TARGET byte-identical. Resource amendment: `docs/evidence/s13/round-23/resource-reconciliation.json` and `round23-cleanup-after.json`; **`AEH-V2-0130` remains OPEN and is carried to S14**, whose disposable self-hosting composite must include cleanup/reconciliation and the bounded repeated-operation stability campaign. No gate was weakened and no slice self-acceptance is claimed; S13 stays uncommitted and S14 must not start.

**S14 self-hosting composite gate — 2026-09-30 (implementing lead; `IMPLEMENTATION_COMPLETE`; independent review owns acceptance):** `src/certification/composite.ts` implements the TARGET §16 deterministic composite gate as a frozen-policy, versioned-manifest evidence aggregator. The policy requires 21 items — candidate truth, controller fencing, capability authority, policy freeze, execution identity/provenance, SkillManifest/context, Paseo session identity, impact-driven assurance, validation, independent review, acceptance/oracle, repair/replan, human exception handling, delivery reconciliation, runtime recovery, sandbox/security, supply chain, real-provider certification, adversarial/fault injection, resource-lifecycle reconciliation (amendment) and the disposable self-modification campaign — backed by 20 versioned S1–S13 artifacts pinned by sha256 plus 4 S14 runtime artifacts. Verification is deterministic and fails closed on a missing or unsafe manifest, unsupported version, policy-digest mismatch, frozen-pin drift, tampered bytes, and stale/replayed candidate bindings; readiness is computed only from those checks (`readinessAuthority: DETERMINISTIC_COMPOSITE`, `modelAuthority: NONE`), and supplied model readiness claims are recorded but never evaluated. Result at candidate buildDigest `4c05da6ee7fe249bcad4adb730f2c08bb34867910eb512171bccb40b63203284`: `ready: true`, 21/21 items, 46 verified artifact instances, 0 failures (`docs/evidence/s14/self-hosting-composite-report.json` sha256 `75b18b6e…`; `certification-evidence-manifest.json` sha256 `d8c4a69d…`; policy digest `458145cb…`), byte-reproducible with the recorded `generatedAt`. The disposable campaign (`s14-campaign-2`, packed candidate in a `/tmp` fixture) reached terminal `SUCCEEDED`/`ACCEPTED` with `change`+`direct-change` PASS; the checkout tracked digest is byte-identical before/after and all mutation roots are disposable (`docs/evidence/s14/disposable-self-modification-campaign.json` `3d224601…`, raw lane `98d62371…`). The bounded stability campaign measured truthful raw terminal-owned residue and verified terminality-gated reconciliation to a 0/0/0 baseline without reclaiming any non-terminal resource (`docs/evidence/s14/resource-stability-campaign.json` `090dace9…`); **`AEH-V2-0130` remains OPEN** (`OPEN_PRODUCT_DEFECT_EVIDENCED`) for the missing product-side durable ownership/terminal reconciliation. Adversarial checks: 11/11 gate-level negatives plus candidate-identity suites 18/18 (`docs/evidence/s14/adversarial-negative-checks.json` `d987ff6e…`). Boundaries: the versioned real-provider matrix is the independently accepted S13 F2-final binding (`599ed363…`); the S14 campaign's disclosed fixture-level `S13_GOV_BRAIN_*` override is recorded and acceptance-neutral as in S13 R23-F3; the stability campaign's reconciliation is campaign-side verification, not product-side reconciliation. Validation: focused 4 files / 43 tests; canonical non-system suite 204 files / 1,378 tests (1,368 passed / 3 failed / 7 skipped — only the pre-existing `tests/paseoSdkResolve` mise drift); typecheck and build exit 0 (`4c05da6e…`); `git diff --check` clean; TARGET byte-identical. No gate was weakened; the implementing lead does not accept its own slice.

**Final closure conformance record — AEH-V2-0130/0135 product repair and Core Architecture v2 completion (fresh final-closure slice lead, 2026-09-30; independently accepted by the fresh closure reviewer (`docs/evidence/closure/review-closure.json`, `SLICE_ACCEPTED`)):** after the independent S14 acceptance, the closure slice implemented and re-verified durable product-side operation→resource ownership and terminal/recovery reconciliation (see [Closure WorkGraph](CORE_ARCHITECTURE_V2_CLOSURE_WORKGRAPH.md)). Product surface: `src/runtime/operationResources.ts` (durable per-operation registry + deterministic reconciliation + durable receipt), `src/operations/controller.ts` (workspace disposition/registration, post-terminal reconciliation, restart sweep, terminal-entry recovery), `src/operations/monitorProcess.ts` (liveness monitor registered as a `managed-process` resource), `src/semantic/runtime.ts` (exact pre-operation triage session/workspace release by launch id), `src/runtime/providerLifecycle.ts` (generic policy-driven provider-session ceiling), `src/candidates/direct.ts` (DIRECT staging-root ownership), `src/core/errors.ts` (`RESOURCE_CEILING_EXHAUSTED`). Classification is exact: `LIVE_OWNED` preserved, `TERMINAL_ORPHAN` released/archived/stopped only with durable ownership proof, `UNKNOWN_OR_UNOWNED` preserved and reported; no PID-name or filesystem-glob inference is used. REAL_PROVIDER re-verification on closure candidate buildDigest `d266692df02c95bf432ac6c9578e127fb16584aa2cc18d469d2c844ed3b0a2ab`: resource-stability `closure-resource-6` both iterations terminal SUCCEEDED, product receipts `cleanupComplete: true`, census `0/0/0` both iterations, `campaignSideCleanup: "NONE"` (`docs/evidence/closure/resource-stability-product.json` `94ed891d…`); disposable self-modification `closure-disposable-1` terminal SUCCEEDED / oracle ACCEPTED / rows PASS with checkout byte-identical and 13/13 receipt dispositions (`docs/evidence/closure/disposable-self-modification-campaign.json` `17e9b926…`); lifecycle `cancel`/`recovery` 2/2 PASS (`docs/evidence/closure/paseo-lifecycle-cancel-recovery.json` `ead63493…`). Composite regeneration: `ready: true`, 21/21 items, 47 verified artifacts, 0 failed, policy digest `1be97df3f7736b76dd048648555874b43e7df5cffc5d1ba98b6a0b555085aef3`, manifest `ff1b5b11…`, report `9b9e2134…`, `readinessAuthority: DETERMINISTIC_COMPOSITE`, `modelAuthority: NONE`, byte-reproducible with its recorded `generatedAt`; the resource-lifecycle item is backed by PRODUCT-owned reconciliation. Adversarial negatives rerun 11/11 gate-level with 0 failed checks and 18/18 product-level tests (`docs/evidence/closure/adversarial-negative-checks.json` `c100197d…`). Exact source-impact staleness is recorded at `docs/evidence/closure/staleness-analysis.json`; focused closure regressions 43 suites / 201 tests passed (`docs/evidence/closure/focused-regressions.json` `0d670ab4…`). **AEH-V2-0130 and AEH-V2-0135 are FIXED/CLOSED. The closure review is `SLICE_ACCEPTED` for its documented scope, and S13/S14 have separate accepted records.** At the 2026-09-30 closure snapshot, the closure reviewer’s summary included an umbrella statement that S1–S12 are accepted, but the explicit S10 WorkGraph still required fresh independent re-acceptance and no S10-specific accepted review artifact had been found. STATUS therefore left S10 open at that snapshot. The later S10-specific acceptance is recorded below; the umbrella summary and S14/closure slice acceptance did not serve as S10 acceptance. AEH-V2-0132/0133/0134, the `tests/paseoSdkResolve` mise drift and S14-F4 remain explicitly open post-v2 observations; two closure provider flakes are preserved as truthful failed-run history.


## S10 final re-acceptance — 2026-10-01

The S10 acceptance-artifact gap is closed by a fresh independent reviewer’s **`SLICE_ACCEPTED`** verdict on reviewed head `87144f77a0c99088d385e848925643b5439422ea`. The S10-specific machine artifact is [review-s10-final.json](evidence/s10/review-s10-final.json), SHA-256 `f940aa6a72587bccf800ceade409de9b22c175af21fa83b2b99cd63e3362a9cc`. This is bounded final re-acceptance, not a repeat of the original S10 implementation campaign. The original rejection and its remediation record remain intact in [the S10 WorkGraph](CORE_ARCHITECTURE_V2_S10_WORKGRAPH.md).

The reviewer’s source-impact analysis compared checkpoint `9b74c04167e79748312160b6476cc5050050b4e3` through the reviewed head. It classified the frozen isolation/SAST primitives and core tests as `UNCHANGED_CORE_CONTRACT`; HOME/PATH and Trivy harness changes as `TEST_FIXTURE_CHANGE`; later candidate-assurance/provider-lane wiring as `INTEGRATION_EXTENSION_COMPATIBLE`; and provider provisioning/workflow updates as `CI_ENVIRONMENT_CHANGE`. It found no material S10 product-contract change and no unknown classification. Candidate-bound SAST remains required and fail-closed.

Fresh hosted isolation evidence is GitHub Actions [run 36881830507](https://github.com/JamesMorales04/agentic-engineering-harness/actions/runs/36881830507), exact head `87144f77a0c99088d385e848925643b5439422ea`, `rootless-isolation` job `110435254361`. The job provisioned rootless bwrap 0.6.1, passed the non-root namespace preflight (`CapEff=0`, `NoNewPrivs=1`), and ran `npm run test:rootless-isolation`: the real system campaign passed 3/3. It had no skip or `continue-on-error` path. The hosted merge tree SHA `13103c24b8bf0f7651baefc424d18dbf2bbf5af6` equals the reviewed candidate tree. The same exact-SHA workflow completed all 13/13 jobs.

Focused S10 deterministic verification passed locally: 5 files / 30 tests, including the unmocked real rootless system campaign (3/3); the executable Trivy resolver regression passed 2/2. In exact-SHA CI, `tests/security/isolation.test.ts` (10), `tests/security/sastEvidence.test.ts` (8), `tests/security/validatorIsolationFailClosed.test.ts` (5), `tests/security/providerFailClosed.test.ts` (4), and `tests/s10TrivyResolver.test.ts` (2) passed. Required provider/scanner absence, stale candidate evidence, tampered JSON/raw evidence, and no-fabricated-PASS behavior remain explicit failures. The dedicated `provider-contracts` lane also passed 17 tests across 4 files.

Fresh candidate-bound real SAST evidence is [real-sast-campaign-85877349558477e21386d6e076393d15e2a90e26-release-1790866583369-387000-cc8194fb.json](evidence/s10/real-sast-campaign-85877349558477e21386d6e076393d15e2a90e26-release-1790866583369-387000-cc8194fb.json), SHA-256 `36a871fd9c8fcafdbb4c116b1d2974fa385ab17e510ba74bdbbf6a430b8e3641`. Its build identity binds implementation SHA `85877349558477e21386d6e076393d15e2a90e26`, build digest `93fb2b413e748a3fdc9070a0cf6c0100facd37f15894dea7ce4904dacd258bbe`, and `dirty:false`. Trivy 0.70.0 binary SHA-256 is `379d59f24a4a828c55de5f0b91b6805cc35d13580180b658820e648611256166`; the actual bwrap 0.13.0 execution denied network and recorded namespace evidence. Candidate r1 (`CAND-S10-REAL`, identity `8648ef7b56d0ecab99eef2cbe33028c6696f4bb26fe11d86270f40de9a850e9b`, source/workspace digest `219debfd791e9a5c71a393f0934bdf57627dca92f583d2a7af59737c40b40a32`) produced one real `github-pat` finding and FAIL. Candidate r2 (identity `154b31f503fcf1da67b1eedbc336ff815a390c28d4beb89fc425231397651ff7`, source/workspace digest `621e7c4125bbe9fde0f3c97c4baee0ac14dd0298a2c5218a985dea0540b8ef2b`) produced a clean PASS. Stale revision and tampered evidence were rejected; all campaign assertions passed.

The S10-specific acceptance did not substitute for, regenerate, or alter S14/closure evidence. The review verified that the historical SAST input remains byte-identical (`docs/evidence/s10/real-sast-campaign.json`, SHA-256 `3a106006801c7b03e79795b8beedc4af240facd16ee78675d8f8194fd7848779`), S14 manifest bytes remain pinned (`d8c4a69da36f0b7c29788380aaa6802b3ff1a0bc1346e7f1bb2f3019c7207d3c`), closure manifest bytes remain pinned (`ff1b5b113adeae34c88abdd6aaa5c2a4770f98a8d3b47cb8dd521def4ea54b9a`), and the new S10 artifact is not a composite input. `s14RerunRequired` is `false`. The normative TARGET remains SHA-256 `4954b62bfb3a820fc6e5d0369b2d37bc35a94988d19e485f00d30675aafa555a`.
