# Core Architecture v2 S3 Tactical WorkGraph and Delegation Assessment

**Slice:** S3 — Authority Kernel & Human Control
**Campaign base:** `core-architecture-v2` at `92abcbeca51f134d5635ec87abde93b99ad0cb2c`
**Normative source:** `docs/CORE_ARCHITECTURE_V2.md` §§3.1–3.3, 10.2–10.3 and the canonical workflow/action-gate contracts. The target is frozen.
**Status dependency:** S1 execution identity and existing candidate truth; S1 and S2 are accepted at the base.

## Tactical WorkGraph

Mechanism labels classify the enforcing mechanism. Authority and identity gates are deterministic; no model judgment may authorize a decision, mutation, cancellation, or effect.

| WorkUnit | Objective and contract | Mechanism | Classification | Dependencies | Scope / file boundaries | Acceptance |
| --- | --- | --- | --- | --- | --- | --- |
| S3-U1 | Consume only a current, typed, attributable, scoped, unexpired, unreplayed HumanDecision. Keep product choice separate from exact action/effect authorization. Bind decisions to current operation, candidate, frozen policy identity, and controller epoch where applicable; model and semantic assessments never grant authority. | DETERMINISTIC | LUNA_OWNED | S1 candidate, execution revision, frozen policy, and epoch identities. | `src/security/humanDecision.ts`, `src/control-center/decision.ts`, operation decision-consumption/continuation call sites, and focused HumanDecision/control-center tests. No Control Center redesign. | Production has a consumer. Stale operation/candidate/policy/epoch, wrong decision kind/scope/actor, expired and replayed decisions fail closed. Product-choice data cannot satisfy action authorization; an action authorization is bound to exact action and effect digest. |
| S3-U2 | Require current controller ownership and epoch for controller-owned mutations and terminal transitions; bind the explicit Lead acknowledgement path to its current Lead actor, epoch, and operation revision. Cancellation establishes its fence before cleanup and rechecks ownership before terminalizing. Preserve candidate and frozen-policy lifecycle guards. | DETERMINISTIC | LUNA_OWNED | S1 controller epoch, operation execution revision, candidate truth. | `src/operations/state.ts`, `src/operations/controller.ts`, focused operation/controller-fencing tests. | Missing/stale controller epoch or token cannot perform controller-owned mutations or terminalize; stale Lead actor/epoch/revision cannot acknowledge; stale cancellation cannot kill writers or cancel after takeover; current owner can cancel only after writers are fenced and uncertain effects reconciled. |
| S3-U3 | Make OPA trust-boundary policy predicates match the exact canonical AEH role values used by execution identity. No aliases, case folding, or role-class expansion. | DETERMINISTIC | FLASH_DELEGATABLE | Frozen canonical role names in `src/participants/roles.ts`; no dependency on S3 integration. | Frozen assignment: `policies/core/trust-boundary.rego`, and the smallest focused OPA input/policy regression test. Worker may not edit other files. | Canonical Implementer and Reviewer inputs activate their intended deny rules; other canonical roles do not alias to those rules. Invalid/unknown role handling stays fail-closed where the adapter owns it. |
| S3-U4 | Align mutation and sensitive-effect authorization with frozen operation policy, role ceiling, exact scope, candidate, actor, lease, epoch, and action intent. Consume a matching HumanDecision only when frozen policy requires that exceptional authorization. | DETERMINISTIC | SEQUENTIAL_AFTER_CONTRACT_FREEZE | S3-U1 HumanDecision contract; S3-U2 controller fence; S1 execution identity/candidate truth. | `src/security/toolActionGate.ts`, `src/security/gatedAction.ts`, relevant security policy/lease helpers and action-gate/reconciliation tests. | Each sensitive effect is authorized from the current frozen policy and deterministic role/controller authority; external effects must appear in policy's allowed effects; required human authorization matches operation + candidate + action + effect and is consumed once. Stale or missing policy, candidate, actor, scope, lease, epoch, or required decision fails closed. |
| S3-U5 | Reconcile the affected S3 status/conformance rows and deterministic evidence lane after all implementation evidence is available. | DETERMINISTIC | LUNA_OWNED | S3-U1–U4 implementation and validation evidence. | `docs/CORE_ARCHITECTURE_V2_STATUS.md`, `docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md`; `docs/ENGINEERING_LEDGER.md` only for newly discovered evidence-backed defects or a finding materially closed/refined. Never edit the frozen target. | Current rows state CURRENT, STATUS, GAP, NEXT GATE, EVIDENCE, EVIDENCE LANE, LAST VERIFIED; lane claims distinguish contract/unit/build evidence from adversarial, provider, browser, and packed evidence. |

Dependency graph: `S1 identity/candidate truth → S3-U1 → S3-U4`; `S1 identity/candidate truth → S3-U2 → S3-U4`; `S3-U3` is independently mergeable; `S3-U1–U4 → S3-U5`.

## Delegation Assessment

`S3-U3` is safely separable because the defect is a finite, deterministic mismatch between OPA's lowercase comparisons and frozen title-cased canonical role values. Its inputs and files are isolated from the HumanDecision and operation-state contract work. It is meaningful because the policy predicates currently do not match production canonical role values, leaving the affected policy checks ineffective. The Lead freezes its acceptance contract and owns integration and review.

Before delegation, `list_profiles` returned the configured **DeepSeek V4.1 Flash Implementer** profile exactly as required: provider `opencode`, model `opencode-go/deepseek-v4.1-flash`, mode `build`, thinking `max`, features `{ "auto_accept": true }`. The worker must use its own Paseo worktree workspace and must not create agents or delegate.

## Inspection findings at the campaign base

- Initial Git status was clean on `core-architecture-v2`; HEAD exactly matched the expected base.
- `HumanDecisionLedgerV1` records operation and candidate plus kind/actor/reason/time, but has no consumer or policy/epoch/scope binding; its `active()` read is not replay consumption. Control Center currently records the posted decision and does not resume or authorize an operation.
- `transitionOperationToTerminal` and generic `mutateOperation` compare an environment epoch only when one is present; they do not require the current controller token. `cancelOperation` performs process/agent cleanup before `terminalizeOperation` reaches that optional epoch check.
- `ToolActionGate` checks operation and candidate plus selected role/lease/epoch, but does not validate the operation's frozen policy. The role-based external path uses Lead/Director status as approval and does not enforce `allowedExternalEffects` or consume an exact human action authorization. The controller path does require the current token.
- The OPA trust-boundary Rego compares `implementer` and `reviewer`, while canonical AEH roles are `Implementer` and `Reviewer`.
- S10 rootless/network/validator isolation and SAST are explicitly excluded. S4 CandidateImpact recompilation, S5 context/session runtime, and S6 acceptance-oracle work are explicitly excluded.

## Frozen Flash assignment

**OBJECTIVE:** Correct OPA canonical-role matching so frozen canonical AEH roles trigger the intended existing policy rules.

**EXACT SCOPE:** Update the trust-boundary OPA role predicates and the smallest focused regression test. Preserve every rule's meaning and decision behavior except that the predicates must match exact frozen canonical role semantics.

**FILES/AREAS:** `policies/core/trust-boundary.rego`; a focused test limited to the OPA input/policy contract (prefer `tests/architectureIntegrationContracts.test.ts` if sufficient). Do not edit other production files.

**DEPENDENCIES:** Canonical role spellings in the frozen target and `src/participants/roles.ts`; no S3 integration dependency.

**INVARIANTS:** No aliases, case folding, role classes, or capability expansion. Model output cannot select a role. Do not touch S1 identity/candidate/policy contracts. Do not weaken existing gates.

**ACCEPTANCE CRITERIA:** Exact `Implementer` and `Reviewer` identities activate their corresponding policy conditions. Other canonical roles do not match those conditions. Invalid/unknown identity handling is not broadened. Test evidence is deterministic and does not claim OPA provider certification.

**TESTS:** Run the smallest relevant focused OPA adapter/policy test, then report the exact command and result. Do not run broad suites.

**EXPLICIT EXCLUSIONS:** HumanDecision, state mutation, cancellation, tool-action authority, control-center behavior, S4/S5/S6/S10, docs except this assignment report, and changes outside the named file boundaries.

**NO NESTED DELEGATION:** Do not create, spawn, or ask another agent to work. Return changed files, diff summary, exact tests/results, and unresolved issues.

## S3 authority contract freeze (lead-owned)

Before implementing S3-U1/U2/U4, the Lead freezes these source-level contracts as an equivalent implementation of target §§3.1–3.3, 10.2–10.3, and the ActionIntent/ToolActionGate requirements:

- A persisted HumanDecision binds the current operation, complete CandidateRevision, operation execution revision, ResolvedOperationPolicy digest, controller epoch, external actor, typed purpose/scope, and expiry. Product choice, action authorization, and operation control are disjoint purpose variants. An action authorization additionally binds the exact registered action and canonical effect digest. Only a single matching, unexpired, unconsumed decision can satisfy a policy requirement; reject/approval are not interchangeable; consumed decisions cannot replay. Legacy decision storage must fail with an explicit unsupported-version/migration error rather than be dual-read.
- The paired Control Center session is the source of the human actor identity; a request body cannot self-assert or replace it. The authoritative producer reloads operation state and derives all operation, candidate, policy, execution-revision, and epoch fields from that state.
- Every ToolActionGate request validates the current operation/candidate and frozen policy identity. External effects are controller-only and must be named in the frozen policy's `allowedExternalEffects`; a canonical participant role, including Lead/Director, is never itself action approval. Local participant mutations continue to require the existing compiled role ceiling and current lease. A policy-declared HumanDecision requirement consumes only the exact action/effect authorization. Action intent/receipt identity records the frozen policy digest.
- Controller-owned operation mutations and terminal transitions require both the current durable epoch and its current token. The explicit Lead acknowledgement is separately authorized by the current Lead actor, epoch, and operation revision, without disclosing the controller token. Cancellation must claim/hold ownership before cleanup and recheck before every cleanup/terminal operation; no stale cancellation may fence another owner. Existing action reconciliation remains authoritative; cancellation cannot report terminal CANCELLED while unresolved action intents remain.
- S1 candidate, execution revision, frozen policy lifecycle and epoch invariants remain unchanged. No new role, capability, action kind, policy bypass, compatibility branch, or lifecycle state is introduced by this contract.

### Remediation Tactical WorkGraph

This WorkGraph supersedes the initial S3 WorkGraph above only for the still-open PRODUCT_CHOICE continuation gap. The frozen normative target remains unchanged. All implementation units are bounded to S3; S10 isolation/SAST, later slice work, provider certification, browser certification, and live OPA evaluation remain excluded.

| WorkUnit | Objective and frozen contract | Mechanism | Classification | Dependencies | Scope / file boundaries | Acceptance |
| --- | --- | --- | --- | --- | --- | --- |
| S3-R1 | Add typed `DecisionRequestV1`, `ContinuationRecordV1`, and stored resolution contracts. A request binds operation, complete candidate, `operationExecutionRevision`, frozen policy digest, current controller epoch, request ID, bounded options/consequences, authoritative evidence references, tried work, reason the system cannot decide, parallel work, resume target, and expiry. Continuation binds the saved target/reason plus the same identity and required revalidation. Product choice, action authorization, and operation control stay disjoint. | DETERMINISTIC | LUNA_OWNED | Existing HumanDecision V2, S1 candidate/execution identity, controller fencing. | `src/security/humanDecision.ts`, `src/operations/state.ts`, focused contract tests. | Malformed, expired, stale, cross-operation/candidate/policy/epoch, wrong actor/scope, and replayed decisions fail closed. Stored old versions fail explicitly. Product choice cannot satisfy action authorization or cancel/retry/acknowledge. |
| S3-R2 | Connect the real formal-SDD blocked boundary to the typed request lifecycle. Spec Manager may propose bounded options as semantic data; controller derives artifact references/digests and current identity, persists the request and continuation, and suspends at `SPEC_AUTHORING`. The controller alone consumes a submitted choice, revalidates current candidate/policy/revision/epoch and checkpoint, then resumes that saved stage. | HYBRID | LUNA_OWNED | S3-R1, current Spec Manager `BLOCKED` output at `src/operations/change.ts`, current operation controller. | `src/agents/outputContracts.ts`, `src/operations/change.ts`, `src/operations/controller.ts`, authority-owned state helpers, focused operation tests. | Missing/invalid choice options do not invent choices and stay blocked. No user/client input can create an operation, participant, ChangeSet, policy grant, tool grant, or mutation authority. A chosen option is applied only as bounded Spec Manager context; resulting authored requirements are compiled/sealed by existing controller-owned steps. Dependent execution bindings/policy are invalidated/recompiled before work continues. |
| S3-R3 | Expose current product-choice requests and submit the selected typed option from the authenticated Control Center session. The body includes only operation/request/choice IDs and an optional human reason; actor and all bindings are reloaded/derived by the server. | DETERMINISTIC | FLASH_DELEGATABLE (UI only, after this contract freeze) | S3-R1 request/projection/submission types and endpoint path/shape frozen below. | Flash may edit only `ui/control-center/src/App.tsx`, `ui/control-center/src/api.ts`, and `ui/control-center/src/styles.css` in a separate Paseo worktree. No shared-checkout work. | The UI renders the bounded issue/evidence/tried/why/options/consequences/parallel-work/resume-target details, submits only an offered `choiceId` over the CSRF-protected route, shows stale/expired/replay failures, and refreshes after acceptance. It sends no actor, candidate, policy, epoch, lease, or authority fields. |
| S3-R4 | Integrate the server projection and submit route with durable operation continuation and the one-time HumanDecision ledger. Use the paired-session actor and never a caller-supplied actor. Persist the submitted choice idempotently, consume it once under the controller's current binding, and keep product choice separate from exact action authorization and operation control. | HYBRID | SEQUENTIAL_AFTER_CONTRACT_FREEZE | S3-R1 and S3-R2; frozen Control Center contract below. | `src/control-center/contracts.ts`, `src/control-center/decision.ts`, `src/control-center/server.ts`, `src/entry.ts`, `src/security/humanDecision.ts`, focused Control Center/HumanDecision tests. | A post is accepted only for the currently persisted open request and one of its offered choices. Stale/replayed/mismatched submissions cannot resume. The controller revalidates the continuation before returning to its recorded stage. |
| S3-R7 | Close Campaign Director’s stale-binding finding: retain the original full decision binding; compare it against the current operation and continuation; validate checkpoint envelope identity against both; permit only the authorized same-candidate/same-epoch requirement revision and current-policy rebind. An unanswered request may be reissued only after checkpoint integrity and unchanged candidate/revision are proven. | DETERMINISTIC | LUNA_OWNED | S3-R1–R2, S1 candidate/execution identity, current controller epoch. | `src/security/humanDecision.ts`, `src/operations/state.ts`, `src/operations/change.ts`, `tests/operations.test.ts`. | Changed decision/current execution revision, policy digest, controller epoch, checkpoint binding, or recomputed-digest checkpoint tampering fails closed. A stale consumed choice cannot be rebound; after verifying its exact old receipt and unchanged candidate/revision, recovery opens a fresh current-binding HUMAN_REQUIRED request. An intact unanswered choice may be safely reissued after policy/epoch drift. |
| S3-R8 | Close the crash window between product-choice execution-semantics persistence and policy binding. Verify controller bootstrap binds/reuses the current deterministic policy before continuation loading; same-epoch recovery must validate the full checkpoint and saved choice binding, while takeover must require a fresh current-epoch choice. | DETERMINISTIC | LUNA_OWNED | S3-R7, controller ownership/epoch lifecycle, frozen policy compiler. | `src/operations/controller.ts`, `src/operations/state.ts`, `src/operations/change.ts`, `tests/operations.test.ts`. | An interruption with missing current policy is recovered through the bootstrap policy binder before checkpoint/resume; policy and continuation identity are current before resumption. Takeover never rebinds the old choice and durably returns to HUMAN_REQUIRED. No silent RUNNING continuation remains. |
| S3-R5 | Add deterministic regression coverage for request creation, durable continuation, exactly-once selection/consumption, restart/takeover fencing, stale and expiry rejection, successful spec-authoring resumption, and product/action/control purpose separation. | DETERMINISTIC | LUNA_OWNED | S3-R1–R4, S3-R7, and S3-R8. | Focused HumanDecision, operations, change-lifecycle, and Control Center tests. | Tests exercise actual state and submission consumers, use deterministic fixtures, do not weaken existing gates, and make no browser/provider/OPA certification claim. |
| S3-R6 | Update the S3 STATUS, CONFORMANCE, and applicable ENGINEERING_LEDGER record from observed implementation and validation results. | DETERMINISTIC | LUNA_OWNED | S3-R1–R5, S3-R7–R8, and final verification. | `docs/CORE_ARCHITECTURE_V2_STATUS.md`, `docs/CORE_ARCHITECTURE_V2_CONFORMANCE.md`, `docs/ENGINEERING_LEDGER.md`; never edit the frozen target. | Rows state CURRENT, STATUS, GAP, NEXT GATE, EVIDENCE, EVIDENCE LANE, and LAST VERIFIED accurately. Only full deterministic S3 Gate evidence can set Authority kernel to IMPLEMENTED; provider, browser, live OPA, and packed adversarial lanes remain separately pending. |

Dependency graph: `S1 identity + accepted S2 runtime → S3-R1 → S3-R2 → S3-R4 → S3-R7 → S3-R8`; `S3-R1 typed UI contract → S3-R3 (isolated Flash worktree)`; `S3-R2–R4 + S3-R7–R8 → S3-R5 → S3-R6`.

### Product-choice contract freeze for this remediation

- `DecisionRequestV1` is controller-persisted operation data, not an authorization token. It contains `version`, `requestId`, `operationId`, one complete `CandidateRevisionV1`, positive `operationExecutionRevision`, current `policyDigest`, current `controllerEpoch`, `issue`, nonempty controller-verified `authoritativeEvidence[]` artifact references and SHA-256 digests, nonempty `whatTried[]`, `whyUnresolvable`, one or more bounded `choices[]` (`choiceId`, `label`, `description`, `consequences[]`), `workThatCanContinue[]`, literal `resumeTarget: "SPEC_AUTHORING"`, `createdAt`, and `expiresAt`. Duplicate IDs, empty required text, unsafe artifact references, invalid digests, invalid timestamps, or an expiry not after creation are rejected.
- The current production producer is `runChangeOperation` after the durable Spec Manager output. Extend that output with typed `decisionRequests[]`; it remains model-proposed data. The deterministic controller accepts a request only when the Spec Manager result is current and provenance-accepted, the result is `BLOCKED`, the request is structurally complete, and current operation/candidate/policy/revision/epoch checks pass. It derives the artifact paths/digests and policy/identity bindings. An absent typed request stays explicitly blocked; it never synthesizes options from prose.
- `ContinuationRecordV1` stores `version`, `resumeTarget: "SPEC_AUTHORING"`, `reason: "PRODUCT_CHOICE"`, `requestId`, the same current candidate/execution-revision/policy/epoch binding, `checkpointArtifact` and `checkpointDigest`, `requiredRevalidation` containing candidate-current, operation-revision-current, policy-current, controller-epoch-current, and checkpoint-current checks, and suspension/resume timestamps/state. It is persisted before the operation exposes HUMAN_REQUIRED. The operation's semantic phase is `HUMAN_REQUIRED`; the public queued/running/terminal transport status is not repurposed.
- Once consumed, the continuation also retains `selectedDecisionBinding`, the exact original full HumanDecision binding used for the ledger receipt. That binding cannot change. Revised requirements may advance execution revision exactly once; after the resulting policy is rebound under the same candidate and controller epoch, both the continuation and checkpoint envelope are rewritten to the current full binding. Any stale consumed candidate/revision/policy/epoch fails closed. Only an intact WAITING checkpoint bound to the saved request and unchanged candidate/revision may be reissued under a new policy/epoch; no consumed choice is reissued.
- The Control Center GET projection contains only the current open `DecisionRequestV1` and continuation summary. POST `/api/v1/decisions` accepts `{ operationId, requestId, choiceId, reason? }` for `PRODUCT_CHOICE`; the session actor is authoritative. The server resolves the offered option and current binding itself. `ACTION_AUTHORIZATION` and `OPERATION_CONTROL` continue through their existing purpose-specific producers/consumers and cannot target a product-choice request.
- The sole accepted product submission is stored as an exclusive, durable request-scoped submission claim referencing the recorded HumanDecision. The controller consumes that exact decision once, stores the consumed decision ID and choice on the continuation before resumption, revalidates the full continuation, then returns to `SPEC_AUTHORING`. Recovery may complete a prior consume only when the durable ledger receipt exactly matches the request, binding, actor, purpose, and decision ID; a new controller epoch invalidates the previous request/submission and requires a new paired-session choice under the refreshed current binding.
- On resumption the selected bounded option is supplied as explicit user product input to the Spec Manager. The human reason remains audit data and cannot add options, scope, tools, participants, or capabilities. Existing OpenSpec compilation and `sealTask` remain the only producers of revised requirements/seals. After a revised contract is sealed, controller-owned execution semantics are rebound so old plans/bindings cannot be reused. The continuation is cleared only after the saved target succeeds under its revalidated binding.
- `DecisionRequestV1` and `ContinuationRecordV1` are the local schema chosen to implement target §§3.3 and 10.2; this does not add a semantic state or alter the target. Product choice creates no unrelated action, tool, participant, operation, lease, ChangeSet, or controller authority.

### Flash launch record and integration boundary

`list_profiles` returned the configured **DeepSeek V4.1 Flash Implementer** profile: provider `opencode`, model `opencode-go/deepseek-v4.1-flash`, `modeId: "build"`, `thinkingOptionId: "max"`, `featureValues: { "auto_accept": true }`. `create_agent` must materialize exactly `provider: "opencode/opencode-go/deepseek-v4.1-flash"` and `settings: { modeId: "build", thinkingOptionId: "max", features: { "auto_accept": true } }`. The UI work is safe to delegate only after the typed contract above is frozen, must run in a separate Paseo worktree, may touch only the three UI files listed for S3-R3, and must not create child agents. LUNA owns all request/continuation lifecycle contracts, server and operation integration, conflict resolution, review, and verification.

### Reinspection before this contract freeze

- The checkout is `core-architecture-v2` at `92abcbeca51f134d5635ec87abde93b99ad0cb2c`; it has exactly 56 modified tracked paths and 4 untracked paths, matching the supplied preserved S3 state. No checkout mutation, reset, discard, commit, or push occurred.
- S1 identity/candidate and S2 semantic-runtime rows are accepted in STATUS. The S3 dependency on current candidate, operation execution revision, policy, and controller epoch is represented in current source.
- `src/operations/change.ts` already receives provenance-accepted structured Spec Manager output. Its actual `SPEC_MANAGER_BLOCKED` throw is the narrow product-choice producer boundary; no other production `DecisionRequest` producer or operation resumption callsite currently exists.
- `src/control-center/server.ts` already has a paired-session CSRF-protected POST `/api/v1/decisions`, but its decision input has no product-choice lifecycle integration; `ui/control-center/src/App.tsx` has no decision view/callsite. The existing ledger only provides exact action/cancel consumption.
- Current operation status is QUEUED/RUNNING/terminal and `phase` is the operation stage projection. The continuation will be represented in the durable V2 operation record and exposed as semantic `HUMAN_REQUIRED`, without misusing the public transport status.
- S10 isolation, network/sandbox, validator isolation, and SAST remain excluded. Existing `opa` unavailability and pending provider/browser/adversarial lanes remain evidence-boundary constraints, not blockers to this deterministic remediation.


## S3 implementation evidence and remaining boundary

**Verified 2026-09-24.** The frozen WorkGraph was implemented with deterministic ownership/controller, HumanDecision, and action-policy checks; the OPA correction was integrated from the assigned Flash worktree and reviewed against the exact canonical role semantics. The explicit Lead acknowledgement mutation is bound to the current Lead actor, epoch, and revision; controller-owned mutations retain token-plus-epoch fencing. After this acknowledgement change, the focused S3 authority/action command passed 10 files / 146 tests; `npx tsc -p tsconfig.json --noEmit` and `git diff --check` exited cleanly; `npm run build` succeeded as `release-1790277810343-2032089-c2f3666b`. The full non-system Vitest run completed with 318/318 suites successful, 810 tests passed, 7 pending, and zero failed. The S3-relevant deterministic system suites passed 45 tests across adversarial, concurrency, and scenario-matrix files.

The final complete `tests/system` run after the acknowledgement change passed 63 tests and skipped none, with one failure among 64. The natural-language `intent` CLI black-box assertion at `tests/system/aehHumanBlackBox.test.ts:84` received empty stdout because the command exited before output with `SEMANTIC_ASSESSMENT_UNAVAILABLE`: configured model `opencode/opencode-go/MiMo-V2.6-Flash` is absent from the local Paseo provider catalog. Direct reproduction confirmed this provider-preflight cause; it is outside S3. The `opa` binary is also unavailable; OPA verification is source/adapter-level and does not claim live policy-provider execution.

Product-choice HumanDecision consumption and operation resumption remain open because the current operation lifecycle has no wired HUMAN_REQUIRED/PAUSED continuation. The Control Center UI API method now matches the typed request contract without a caller-supplied actor, but the app still has no decision callsite. No lifecycle schema was invented and the frozen target was not changed. Deterministic results do not certify packed adversarial cancellation/takeover, real providers, or browser behavior.


## S3 product-choice continuation remediation result

The earlier “implementation evidence and remaining boundary” section is a pre-remediation checkpoint; its statement that product-choice consumption/resumption and the App decision callsite are open is superseded by this completed same-slice work. The frozen target and the authority contract above were not changed.

- The production Spec Manager BLOCKED result now requires one typed bounded product decision. Controller-derived current candidate, execution revision, frozen policy, epoch, provenance-accepted evidence, and expiry are persisted in a DecisionRequest and checkpointed continuation before the operation projects HUMAN_REQUIRED.
- The paired Control Center App presents the request and submits only operation/request/choice IDs and an optional reason. The server derives the actor from the paired session and all operation bindings from current durable state. Request-scoped durable reservation and the HumanDecision ledger reject replay.
- The controller consumes the exact choice once, validates the checkpoint/current bindings, resumes SPEC_AUTHORING with the selected option as product input only, compiles and validates revised requirements, invalidates/rebinds dependent execution semantics and policy, and completes the continuation only after the implementation target finishes. Choice cannot grant action, tool, participant, lease, mutation, or operation authority.
- Campaign Director source review found that the consumed-decision receipt was checked against the decision’s own binding and that checkpoint loading omitted full binding comparisons. The Luna-owned fix now compares the original consumed decision binding against the current operation and saved continuation, checks the exact ledger receipt on that saved binding, and validates the checkpoint envelope against the continuation and current operation before returning or resuming. It prevents a consumed choice from being rebound after controller, policy, or execution-identity drift and allows only the single same-epoch requirement revision. The controller bootstrap rebinds deterministic policy before resume; if takeover invalidates a consumed choice, recovery verifies the old receipt and unchanged candidate/revision, then durably creates a fresh current-binding HUMAN_REQUIRED request instead of leaving the operation running without a route. Deterministic regressions cover both crash windows, revision/policy/epoch changes and digest-recomputed checkpoint tampering; safe unanswered-request reissue after takeover is also covered.
- Flash delivered only the three UI files in the isolated Paseo worktree already recorded above; it created no child agents, made no shared-checkout edit, and created no commit. LUNA performed contract freeze, lifecycle/server integration, review, conflict resolution, and verification.
- Final deterministic evidence after stale-binding and crash-window review: focused S3 authority/action tests passed 12 files / 167 tests; `npx tsc -p tsconfig.json --noEmit` and `git diff --check` passed; `npm run build` passed as `release-1790285875947-2230303-2736c95d`; full non-system Vitest passed 318/318 suites, 819 tests, 7 pending, zero failed; the three S3-relevant deterministic system files passed 45 tests. The shared-checkout UI API mock smoke passed pairing, strict DecisionRequest parsing, stale 409 handling, CSRF/submission-field checks, accepted response, and refresh. Checkpoint tests cover digest-recomputed binding tampering and valid unanswered takeover reissue. A deterministic interruption test starts after semantics binding and before policy binding, proves the checkpoint is unavailable without current policy, runs the same bootstrap policy binder used before controller resume, then proves the current checkpoint and continuation can resume; takeover coverage proves stale consumed choices become a fresh HUMAN_REQUIRED request.
- Deferred evidence remains accurately separate: no browser DOM/session run; no live OPA executable evaluation; no real-provider certification; and no packed adversarial/provider cancellation/takeover certification. The frozen S3 deterministic gate is complete for independent acceptance.
