# V9-003 evidence pack — invoked tutor and agent safety/authority

## Work-package charter

- **Task ID:** `V9-003`
- **Owner:** Metis tutor, agent-runtime, prompt-trust, and safety maintainers;
  implementation owner: the owners recorded in PROMPT-002, PROMPT-003,
  SAFETY-001, and AUTH-003
- **Accountable product:** V9 Metis
- **Priority:** P0
- **Target milestone:** 2026-07-15 SOTA gap-closure program
- **Risk tier:** Critical prompt injection, learner-data exposure,
  source-poisoning, unsafe advice, academic-integrity bypass, cross-tenant, and
  delegated-tool escalation risk
- **Dependencies:** PROMPT-002 invoked boundary migration, PROMPT-003 held-out
  adversarial portfolio, SAFETY-001 persisted safety admission, AUTH-003
  governed DAG execution, and V9-002 evidence-abstaining mastery semantics
- **Affected schemas and content slots:** Metis tutoring trust zones and
  evidence state, safety records, governed agent plans/executions, and
  common-run projections; no production lesson/content slot was released
- **Rollout strategy:** Require trust-zone rendering, source authority, current
  safety admission, explicit handoff/evidence transitions, and least-privilege
  runtime grants on the ordinary invoked paths
- **Rollback strategy:** Revert only with all four constituent packages and V9
  path bindings. Never retain provider or tool execution after removing its
  trust, safety, tenant, grant, or evidence gate.
- **Evidence-pack location:** This file plus the four constituent packs below

## Constituent implementation evidence

V9-003 is a reconciliation package, not a duplicate implementation:

- `evidence/v1-v9/PROMPT-002.md` proves invoked trust-zone fencing for tutor
  topic/mode, learner/mastery/orchestration state, integrity policy, source
  previews, conversation history, and prior model output.
- `evidence/v1-v9/PROMPT-003.md` proves the held-out indirect/source-poisoning
  attacks and provider/boundary-change invalidation.
- `evidence/v1-v9/SAFETY-001.md` proves current tenant/input/policy-bound safety
  screening before V9 planning, provider generation, or tools and a recorded
  fail-closed learner degradation.
- `evidence/v1-v9/AUTH-003.md` proves the executable DAG boundary with runtime
  registry binding, grants, exact permissions, tenant/data scopes, human gates,
  isolation, durable idempotency, and transactional publication.

The reconciliation source commit is `a05226b6d0e950456b9491048b5bfb8e06af0845`.
Constituent publication identities are retained in their owning packs;
AUTH-003's authoritative runtime source is
`b8107e5841c48b100ceca1f116608d8c39f97ecf`.

## Verification record

The original reconciliation passed 36 focused Python tests and 20 focused
TypeScript adversarial regressions across trust rendering, invoked tutoring,
mastery abstention, safety classification, governed DAG execution,
indirect/source poisoning, learner-data fencing, integrity/unsafe-advice,
cross-tenant denial, and undeclared authority escalation.

The refresh ran against source commit `649ae21ae3c44c9c8e259067674348e9ad51b205`
on `app-dev-server`, Linux `6.8.0-134-generic` x86_64, Node.js `v24.15.0`, pnpm
`10.25.0`, Git `2.43.0`, Python `3.12.3`, pytest `9.0.3`, and Vitest `4.1.7`.

| UTC interval                                      | Exit | Command and result                                                                                                                                                                                                                                                                                                                           |
| ------------------------------------------------- | ---: | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 2026-07-29T11:12:03.294Z–2026-07-29T11:12:44.207Z |    0 | `pnpm exec vitest run --config testing/prompt-injection/vitest.config.ts testing/prompt-injection/src/v8-v9-boundaries.spec.ts --maxWorkers=1` — 41/41 current V8/V9 boundary and typed-data integrity cases passed                                                                                                                          |
| 2026-07-29T11:12:50.240Z–2026-07-29T11:12:50.595Z |    0 | `PYTHONPATH=apps/metis/service/src python3 -m pytest -q apps/metis/service/tests/test_prompt_trust_helpers.py` — 3/3 trust-rendering cases passed                                                                                                                                                                                            |
| 2026-07-29T11:11:38.804Z diagnostic               |    2 | A five-file system-Python refresh stopped during collection because the host interpreter lacks the service-declared `fastapi` and `pydantic-settings` dependencies. It executed no affected cases and is not counted as a product verdict. The existing full service results in the four source packs remain the authoritative Python proof. |

Installing into the shared system interpreter was intentionally avoided. This
evidence-only publication does not change the Python service boundary, and the
missing local environment is recorded rather than replaced with a mock.

## Adversarial, artifact, and authority evidence

- Negative fixtures include indirect prompt injection, poisoned sources,
  learner-state instruction smuggling, academic-integrity circumvention, unsafe
  advice, stale/missing safety records, cross-tenant grants/effects,
  read-to-write escalation, undeclared tools, output-budget exhaustion,
  insufficient publication roles, tampered plan authority, and required human
  gates.
- Unit, contract, adversarial, invoked integration, durable recovery, and
  common-run adapter coverage are applicable and retained in the constituent
  packs. Browser/mobile/accessibility/Unreal tests are inapplicable because
  V9-003 changes backend invocation and authority, not rendered surfaces.
- AUTH-003 exercises API-driven human gates but does not invent a human
  reviewer. No real learner content approval is required or claimed for this
  control-plane implementation; V9-004/V9-005 own actual operator and study
  authority.
- No V9-003 implementation blocker remains. Live providers, approved production
  sources, real learners, efficacy studies, and release/canary evidence remain
  open under V9-001, V9-002, V9-004, V9-005, EVAL, and REL.

| Current artifact                                              | SHA-256                                                            |
| ------------------------------------------------------------- | ------------------------------------------------------------------ |
| `apps/metis/service/src/metis/services/tutoring_service.py`   | `a413391549a5e36f8457ab6deefa920704413135a31ef1978ec74c81238a310f` |
| `apps/metis/service/src/metis/services/agent_dag_executor.py` | `e85cc9236273d89efb77080c2a39080255c8ae9f785d3b3f4a5fcd663864633f` |
| `testing/prompt-injection/src/v8-v9-boundaries.spec.ts`       | `9a85eb61e2f965f6f22d791374d41573ed0627afcc3afdf58760cd648c343ab1` |
| `evidence/v1-v9/PROMPT-002.md`                                | `6d97a4e7b42a32f5e4990390934f9f73fa4a1d10e203811e201525a90bc2ab9c` |
| `evidence/v1-v9/PROMPT-003.md`                                | `3c6ca88dcf1e55f1728a5a42d1716b79a1d21faa446f2eeb6414a42f4c9c4fb0` |
| `evidence/v1-v9/SAFETY-001.md`                                | `65403d3d3c9d7c3a95d4c387ee95b024166c93d07f426f7d05c40dd0ecfc4e6e` |
| `evidence/v1-v9/AUTH-003.md`                                  | `8c2c7d4784ce0c607262bc4afe300a93f7944a430501d610e8e05caa167b2def` |

No production model/provider version, content artifact, release verdict,
placement, screenshot, or media receipt exists for this package. The generated
capability/status views remain source-driven by their canonical registries; the
new completed-work-package verifier separately derives and enforces V9-003's
task-specific evidence linkage.
