Skip to content

[CT-706] Prove authority and refusal enforcement readiness #94

Description

@flyingrobots

ADR-THINK-001 · Delivery plan · Complete issue catalog

Milestone: ADR-THINK-001 P7 — Authority and refusal readiness

Feature: F7.2 — Sufficiency enforcement readiness and answer witnesses

Outcome

Query classes prove the path from shadow to enforced planning only after fixtures, remediation, refusal policy, telemetry, rollback, and historical replay are independently verified; activation waits for P8 cutover.

User stories

  • As a Think user, I want a predictable staged-enforcement rehearsal with useful bounded results, so that semantic safety does not arrive as a sudden wave of unexplained refusals.
  • As a operator, I want per-class enablement, hold, rollback, and incident signals, so that a bad planner or policy can be contained without disabling all Mind reads.

Deliverables

  • Enforcement readiness checklist and signed decision per query class.
  • Non-authoritative canary, cohort, percentage, hold, and rollback controls that preserve capability law.
  • SLOs for answer, bounded, remediation, adjudication, refusal, error, and latency rates.
  • Historical replay, incident, and rollback runbooks with evidence retention and payload erasure.

Acceptance criteria

  • No class becomes production-authoritative before P8, and no class becomes eligible without passing governing fixtures, shadow evidence, remediation or final-refusal policy, and rollback rehearsal.
  • Rollback changes routing behavior but never fabricates authority or projection capability.
  • Historical answers remain reproducible across rollout and rollback.
  • High refusal, latency, error, authority conflict, or privacy signals automatically hold further expansion.

Test plan

Contract and unit

  • Rollout-state tests cover disabled, shadow, canary, cohort, expanding, held, rolled-back, and enforced states.

Integration and acceptance

  • A representative query class advances through the full non-authoritative staged rollout with witnessed results and historical replay.

Failure and recovery

  • Bad manifest, policy conflict, missing remediation, elevated refusal, latency regression, privacy event, and rollback failure trigger holds.

Resource and performance

  • Canary evaluation, telemetry, canonical escalation, and rollback remain bounded under production-shaped load.

Security, privacy, and erasure

  • Rollback and incident tooling preserve erasure and access-view constraints.

Build-time resources

Resource Exclusivity mode Scope
query-enforcement-state exclusive One authoritative rollout state per query class.
query-class-cohorts partitioned Disjoint class and user-cohort deployments.
enforcement-slo-policy exclusive Hold and rollback thresholds.
production-query-telemetry shared Read-only structural metrics and sampled witnesses.
enforcement-incident-lane exclusive One coordinated rollback or incident response.
  • exclusive: Only one active slice may mutate or lease the named resource.
  • partitioned: Concurrent writes are lawful only in disjoint partitions named by each slice.
  • shared: Concurrent read-only use is lawful; this slice does not mutate the resource.

Dependencies

ADR traceability

  • Implementation gates: None directly
  • Constitutional invariants: I9, I10, I11, I17
  • ADR acceptance criteria: AC13, AC14, AC15, AC19, AC20

Non-goals

  • Do not enable every query class in one cutover.
  • Do not weaken semantic checks as a rollback mechanism.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions