F10-F15 Readiness Audit and Completion Plan

This audit reconciles the functional specification with its assessment, implementation owners, and remaining acceptance gates. Use the work IDs below to resume implementation without confusing local tests with operational completion.

Boundary: F10 is included. All 21 subfunctions through F15 were reviewed at source revision 8819a28951fd5c21858adc0a8e7a6d5278ba4745 on 2026-10-18. This is a source, ledger, and exact-issue-state audit, not a new runtime test campaign or live deployment validation. F01-F09 retain their previous assessments and are not recertified by this audit.

Audit results at a glance

The previous F10-F15 result was Ready 13 and Partial 8. The audit moved ten Ready decisions to Partial; the 2026-10-23 exact-source acceptance then restored F14.2 to Ready. The current result is Ready 4 and Partial 17. Across all 52 subfunctions, the inherited and corrected results total Ready 29 and Partial 23.

The plan contains 26 unchecked work packages and seven deferred or conditional scope records. Shared work is linked once rather than counted again for every affected function.

Three causes explain the misleading labels:

• Implementation was mistaken for completion: a passing local cohort did not close a required composed, authenticated, or independently observed acceptance gate.
• Real source work was called configuration: requester-policy activation, Release maximums, source-deployment cleanup, and source-specific health exporters still need implementation or wiring.
• Task markers and issue states disagreed: a closed issue can leave deferred work; an open checkbox can also describe code already delivered. Neither is completion evidence by itself.

Ordinary installation values do not make a generic read-model implementation incomplete. However, an owner’s explicit required operational gate cannot be excluded merely by calling it “installation qualification.” Optional expansion remains separate from the accepted support scope.

Concrete source cross-checks include the protected-download denial, independent activation readback, whole-tuple retained-Pattern qualification contract, and the Release maximum and CLI paths linked below. These support bounded conclusions, not a claim that every source file was reviewed.

Complete subfunction inventory

Local below means the owner records implementation and focused checks, not that those tests were rerun in this audit. Work IDs refer to the unchecked exit criteria in the following sections. The final column is the exact machine-readable mandatory package list. Use comma-separated IDs, not ranges; - means no mandatory package was identified for that bounded function.

Function	Previous -> current	What exists	Remaining mandatory work	Required packages
F10.1	Ready -> Partial	Local ingestion, versioning, extraction/retrieval/deletion and purpose-bound protected source export	Exact-topology lifecycle and approved protected-source provider qualification	K01, K02
F10.2	Ready -> Partial	Local manual and explicit prior-policy v2/v3 API admission, collection, dated retrieval, and containment	Actual prior-policy qualification/promotion, source-quality holds, and authorized source/answer receipts	K03, K04, K05
F10.3	Partial -> Partial	Local current-schema Rules and all nine revision phases	Retained legacy migration and governed rollout; do not reimplement completed phases	R01
F10.4	Ready -> Partial	Local inert semantic compilation, independent deterministic verification, and retirement	Governed ownership handover/withdrawal/hold receipts and actual bilingual assistive output	K06, K07
F11.1	Partial -> Partial	Attributed immutable cases, feedback, prediction outcomes, local retention controls, audited default-off collection preference and per-source Settings detail	Complete source histories, authorized context lifecycle, and retained-Pattern qualification	L01, L02, L03
F11.2	Partial -> Partial	Local replay, scenario evaluation, threshold-bound O7 measurement, and isolated LOCAL comparison	Coherent external sources, preregistered paired evaluation, and complete live cohorts	L03, L04
F11.3	Ready -> Partial	Durable promotion/recall, override attribution, demotion, and read projections	Production Release maximums and governed activation/recall evidence	A01, A02
F12.1	Partial -> Partial	Exact approvals, profiles, revocation, and inert standing-authority/lease contracts	Live authority behavior and provider-commit fencing; no shipped A3-E adapter is eligible	A02, A03, L02
F12.2	Partial -> Partial	Local risk/policy checks, seven safeguards, signed kill switch, and state readback	Release bounds, cross-path effects, and measured stop/recovery receipts	A01, A02, A03, A04
F12.3	Partial -> Partial	Hash-chained audit, minimization, retention/deletion, legal holds, and redacted export	Approved privacy gate and pinned operational governance receipts	D01, D02
F12.4	Partial -> Partial	Fifteen-signal policy, six actual signal and three field bindings, health/readiness, indexed consumed-canary age and actual deterministic-verifier rate	Nine signal/five field sources, latest-attempt result/state, measured delivery/recovery, stabilization, and unresolved operating decisions	H01, H02, H03
F13.1	Ready -> Ready	Authorized, source-qualified read views with explicit partial/unavailable states	No new mandatory read-view implementation gap found; dataset freshness and production source enrollment are not certified	-
F13.2	Ready -> Partial	Local report engine/exports and historical authenticated built-in RCA/PDF evidence	Production-datasource and authenticated error/isolation matrix beyond that built-in receipt	V01
F13.3	Partial -> Partial	Baseline/treatment formulas, guard metrics, independent settlement, and censored outcomes	Qualifying episodes, comparable windows, denominators, and uncertainty	L04
F14.1	Ready -> Partial	Source deployment, Trial, health, update/rollback, and one historical new-Trial receipt	Unchanged Trial rerun and both full-entitlement acceptance cases	I01
F14.2	Ready -> Ready	Accepted one-command source distribution and signed offline-package mechanics with retained isolated install evidence	No remaining mandatory distribution gap; Hub is not an available path	I02, I03
F14.3	Ready -> Partial	Availability/mode/profile, Trial watermark, model-health, and recall presentation	Real Release ceiling input and entitlement-transition readback	A01, I01
F14.4	Ready -> Partial	Typed bundle registration, publisher trust, disabled-first lifecycle, and revocation	Linked exact-artifact install/enable/disable/revoke/restart/disclosure receipts	X01
F15.1	Ready -> Partial	Exact-revision acquisition, independent local acceptance, workers, and historical isolation probes	Terminal installed scan, scheduled publication, and same-correlation business acceptance	S01, S02
F15.2	Ready -> Ready	Deterministic canonicalization, severity/confidence/priority separation, and adjudication	No new mandatory detected-issue processing gap found; detector calibration/promotion is separate S03 work	-
F15.3	Ready -> Ready	Signed scoped packs, guarded developer helper, and coverage-equivalent rescan verification	No new mandatory pack/rescan gap found; neither customer-repository editing nor arbitrary scanner accuracy is claimed	-

F10: knowledge and rule work

The owning records are ingestion, cloud knowledge, structured retrieval, distillation, and Rule governance.

• K01, composed ingestion acceptance: Retain local and deployed five-service receipts for upload, malware/protection inspection, extraction, index/citation, deletion, restart, and failure recovery, bound to exact service identities, versions, and contracts.
• K02, protected sources: Add and test the purpose-specific delegated provider authorization receipt before rights-managed download. Bind approved protection/SharePoint identities and policy; retain inspection, revocation, deletion propagation, cursor recovery, and backpressure readback. Local receipt, admission, bounded streaming, library availability and production binding are now implemented under #2214, with 132 behavior and 29 service-suite ownership checks. The approved external bindings and operational receipts keep this package open.
• K03, requester-policy implementation: Implement the accepted v2/v3 requester path with exact requester/policy/content/collection binding, independent prior authority, request-time and final-visibility checks, denial of self-grants and missing/expired/revoked authority, all required safeguards, and independent persisted readback. Until then, retain the manual lane. The explicit authenticated API path is now implemented under #2217 with a final 672 focused passes and 24 actual loopback PostgreSQL passes. Manual remains the default; actual independently approved policy, deployment read grants and operational promotion keep this parent package open.
• K04, extraction and answer quality: Resolve or explicitly hold hidden UI, lost link/heading context, cross-section table caveats, and every pilot source hold. Retain regression cases and independently reviewed EN/KO query/claim evidence before declaring the corpus retrieval-ready. #2232 adds opt-in normalizer 2.1.1 and reader 3.1.1 for duplicate scope/chrome, ancestor and authentication-control validation, with retained-byte upgrade and 329 focused passes. Old normalizer results stay immutable. Actual pilot-source holds and independently reviewed paired-language claims keep this parent package open.
• K05, source qualification: Retain rights, purpose-bound trust/revocation, independent reviewers, current scanner data, protected policy mounts, and bounded source enrollment. Exercise connected collection, due/restart behavior, changed/unchanged and partial outcomes, plus an internally reviewed OS-egress-denied package-to-answer drill with zero external calls. Bind an eligible immutable runtime release; older application-kit receipts cannot qualify it.
• K06, distillation conformance: Under #458, retain deployed ACL/purpose, withdrawal, legal-hold release/deletion, provider and cohort evidence before separate promotion. Inert compilation and local SQL tests are not these receipts.
• K07, human accessibility evidence: Retain actual EN/KO status/error/disclosure speech with the declared browser and screen reader. Keep UX-39 as U and needs-human until reviewed.
• R01, Rule upgrade and rollout: Close the existing F06.2 operational list: legacy migration, complete provider collection, trusted verifier, blocked-then-cleared review, scheduled expiry/notifier, override source, replica convergence, direct/offline activation, and frozen drift-quality measurement. Current code already implements phases 7-9.

F11-F13: learning, authority, health, and outcomes

The learning ledger delegates source/context/Pattern details to the prediction ledger. The governance, standing-authority, security, privacy, health, and reporting owners retain authority.

• L01, complete prediction inputs: Prove four independent source slices: external-change start anchor and journal continuity; target-scoped action/event-time watermark and contiguous audit; revisioned excluded-window history with positive initial/empty state; and post-window lifecycle reconciliation. Missing history remains unknown. Apply the closure-observation migration; wire audited EN/KO Settings preference and per-source availability to the actual collectors. #2224 removes false positive completeness from the current action-audit adapter: it emits complete=False with action_history_proof_unavailable and no normalized intervention records. Source/producer tests (58), Core consumer tests (41), Operator settings tests (4) and Console consumer tests (21) pass. This is a safety correction, not delivery of the missing positive event-time source; the package remains open. #2242 now records immutable first Thor dispatch entry before executor I/O, without a second same-state ActionRun save. Its 200 focused checks and actual isolated PostgreSQL check pass; observation mode stays non-writing and await-time authority loss blocks dispatch. This is not provider mutation, verified effect or the still-missing independent history proof. Recomputed canonical JSON identities reject cached boolean/number aliases during retry and observation awaits; the six reproduced cases are included in the final focused set. #2246 separates pre-dispatch rejection from executed failure after three actual typed-bus cases invoked rollback without executor I/O. The 219 focused and nine actual PostgreSQL checks preserve genuine-failure recovery and deny replay/restart without executing; parent L01 remains open. The current change implements the separate audited Settings preference forecast.history_collection.enabled, default off without any environment override. Actual collector/producer work is gated before reading, and a preference revision change during the read discards the result, including disable-and-reenable. Console exposes the four source bindings, checkpoint limitations and prerequisites. These local controls do not supply the four missing positive-history proofs, and the mandatory package remains open. #2248 repairs the original shared ActionRun cache, with 26 reproduced JSON value/key-alias regressions, actual preflight rejection, unchanged warm hash/copy counts and 203 focused passes. It preserves canonical identity bytes, but does not supply the independent history proof.
• L02, context safety: Complete authenticated exact-target proposal/review, bounded activation, expiry/revocation, Forseti decisions, Saga audit, and dispatch revalidation. Retain live queued and in-flight revocation/expiry, hold, stop/recovery, deadline, and subscriber-isolation receipts.
• L03, retained Patterns: Qualify current case/revision selection, copied-data lifecycle, frozen/access-scoped cohorts, restart-safe publication, and candidate promotion/demotion. Supply a trusted coherent external issuer covering the whole read/reuse tuple through acceptance; the isolated LOCAL lease implementation is not a production issuer. Close #2194 and the prediction owner’s P4-P6 gates, including its full-scope critique/hardening exit, without reopening already tested cleanup fixes.
• L04, measured outcomes: Freeze a source-versioned SLO/safety-budget evaluation plan before observing results. Retain independent grouped baseline/treatment episodes, mechanism evidence before upgrading raw response outcomes beyond O2, O7 action-specific days/sample sizes, complete recurrence windows, Wilson bounds, zero-escape evidence, missing/censored results, and uncertainty. Reuse the F07 measurement work for cost settlement rather than counting it twice.
• A01, Release ceilings: Replace the interim observation mode-only source with production Release capabilities maximums under #1822. Prove trusted exact-release admission, missing/mismatched input denial, and rejection above the ceiling without weakening recall. The current change distinguishes an absent authenticated Release source (release_maximum_unavailable) from a verified restrictive maximum. The official lifecycle artifact verifier does not establish Core’s current installed Release plus current revocation state, so wiring that prerequisite remains blocked under #1822. No unsigned input is accepted.
• A02, governed authority receipts: Retain fresh policy-administrator assignment/consent readback and one exact-revision single-operator approve/execute/independent-effect receipt with reduced quorum visible in audit. Exercise promotion, override, recall/demotion, restart, and rollback without interpreting display state as authority.
• A03, standing authority boundary: Resolve the provider constraint with a durable boundary that atomically checks the current lease/fencing generation at effect commit before admitting any A3-E ActionType. Separately bind the independent approval verifier and authorize effect-failure reversion; retain urgency and pinned live cohort/promotion evidence. ARM acceptance alone cannot satisfy the fence. This is deferred work, not an enabled automatic-execution path.
• A04, safety drills: Retain pinned cross-path safeguard/independent-effect, kill-switch engage/no-execute/disengage, break-glass, rollback, identity-recertification, audit-anchor, and no-execute approval receipts. Existing signed command and state-read implementations remain done.
• D01, privacy gate: Approve the deployment data inventory, accountable owners, retention schedule, model-provider terms, privacy assessment, regions, and compliance crosswalk.
• D02, operational data governance: Retain pinned deletion, legal-hold, access-review, audit-anchoring/WORM as required, and incident-response receipts. Verify deployed case-history retention/change-feed convergence and expiry rather than inferring it from defaults.
• H01, source exporters: Compose every one of the 15 catalog signals with a real runtime exporter; retain source identity/freshness, primary/fallback delivery, and recovery-drill receipts. Complete the linked Azure guest-event, throttling, and cross-service parity evidence. #2218 / PR #2220 now binds three actual local signals and three fields with 92 focused passes. #2228 adds a process-scoped duration from runtime entry to first protected readiness with a frozen observation timestamp and 80 focused passes in an overlapping test slice. #2250 adds indexed consumed-canary age with 99 focused checks, including four actual isolated PostgreSQL cases and exact byte/time boundaries. Five signals and three fields were bound at that checkpoint. The remaining signal sources and five field sources stay explicitly unavailable; canary result/state, operational index application and delivery/drill receipts remain open, and this package is not complete. #2240 observes actual normal embedding calls in an optional typed model sub-scope, independently bound to the full startup role/binding registration and checked by the Operator reader. The parent aggregate remains unavailable when other configured roles are unobserved. The final SDK/collector/Operator/integration, wheel and document selection passed 169 cases. This does not add a catalog signal/field or a Console model-scope view. This increment adopts #2253’s protected indexed age source rather than publishing a duplicate. The earlier six-signal/four-field local claim is withdrawn: a last-success receipt is not the latest scheduled-attempt result or complete canary state. The current increment adds the actual default T2 deterministic-verifier rate in complete policy-bounded process windows, with 20 source/composed/persisted-readback tests and 100% branch-inclusive source coverage. Unknown current results cannot re-expose a healthy old window. At that checkpoint coverage was six signals/three fields, with nine signals/five fields unavailable. This is neither verifier accuracy nor operational effect, and H01 stays incomplete. #2276 separately wires actual cross-check SDK/pool API observations; it adds no catalog signal. The current change supplies actual Core logical-round disagreement observations: 303 source/Core and 529 runtime/safety/SDK checks pass, with 100% three-module source branch coverage. Seven signal slots/three fields now have concrete implementations and wiring, but the new source requires a qualified native group and full known window. Eight other signals/five fields and installed qualification remain open; fixture checks are not operational improvement. #2286 adds the normal Core process’s actual ingest-to-first-tier-decision lag, not route timing. Pending siblings, incomplete pairs and stale/overflow/clock failures remain explicit. This eighth source does not close the seven remaining signals, five fields or installed H01 gates.
• H02, canary and stabilization: Retain the exact consumed canary audit within the implemented 30-second budget, frozen observation mode replay, stop/no-execute receipts, daily jobs, and recurrent-query signal. Prove stabilization opens, lowers authority on breach, and exits only when its configured conditions hold. #2254 hardens the existing scheduler prerequisite against non-finite intervals with 110 focused passes and actual creation/materialization/due-cycle controls. It does not register those daily jobs or the recurrent-query signal, choose the operating policy, or close this package. The current change now wires actual typed query-shape receipts through Operator and Bragi to Norns’ durable distinct-principal recurrence input and Mimir’s inert candidate queue. The 571 focused checks, including four actual isolated PostgreSQL scenarios, cover local mechanics. Installed transport, actual cohort qualification, daily jobs, policy and stabilization remain open; this increment does not complete H02 or add a self-health signal. The current creation-replay repair also preserves run state across repeated registration, distinguishes atomic duplicate races from other errors, and checks canonical JSON intent. Its 108 focused passes include four new isolated PostgreSQL cases and 100% creation coverage. This is another registration prerequisite, not registration or operational acceptance. The current change also preserves an approved configuration-review campaign’s exact baseline, scope, run IDs, cron and timezone from inert blueprint storage through materialization. Its 26 focused unit checks and two isolated PostgreSQL checks close an intent-loss defect without registering any task or inventing a principal, baseline or scope. Daily health and deployment-baseline jobs, approved installation inputs, transport, policy and stabilization evidence remain open, so H02 stays incomplete.
• H03, operating decisions: Resolve and version all seven open owner decisions: canary audit-freshness/round-trip alert budget; smoke fixtures/budgets/promotion wiring; alert ownership and fallback; audit investigation retention/query model; cold-start deadline; stabilization length/exit signals; and pre-launch load generator/per-tier budgets. The implemented five-minute cadence and 30-second consumed-audit check do not settle every decision.
• V01, report acceptance: For each declared production datasource retain identity, cutoff, freshness, unavailable/timeout, partial-widget isolation, and no synthetic-to-live substitution. Retain authenticated inventory/selection, rejected variables/formats, render-error isolation, and GET-only enforcement across Operator and Console. Historical built-in RCA/PDF evidence is reusable only for that unchanged scope.

F14-F15: distribution, extensions, and scanning

Use the source-deployment ledger, distribution ledger, productization ledger, bundle ledger, and scan ledger. The detected-issue processing ledger supports the bounded Ready decisions for F15.2/F15.3.

• I01, source acceptance: WP6 retains unchanged keyless rerun, new full-entitlement installation, and existing Trial-to-full transition evidence with exact digests, migration/health/readback, and second-plan results.
• I02, superseded paths: Public kit acquisition and source signed-kit adoption redirection are retired. azd-up.sh remains a Core-only diagnostic, while CLI help and tests advertise only source and signed offline installation.
• I03, distribution validation: The exact-source v18 invocation retained connected deployment, bounded recovery, image, inventory, health, and zero-change receipts on the selected revision. Proven shipped-wheel/offline construction remains separate and source installation requires no public release.
• X01, extension lifecycle: Retain one complete linked exact-artifact lineage covering install, enable, disable, revoke, restart, disclosure reload, typed action routing, and audit. Match artifact/release/environment identities and prove no bundle bypasses the governed path.
• S01, installed scan terminal: Resume #2128 only after its pause is lifted and a target is authorized. Retain terminal Console request, exact revision, coverage, isolated independent acceptance, publication, and readback. Local acquisition/worker/acceptance code exists; a ready worker or processing request is not a completed scan.
• S02, scheduled and business acceptance: Retain a scheduled scan/publication and the same correlation’s authoritative Incident, scan result (including an explicit result with no detected issues), external notification, and ownership handover. Supply real Owner access, protected source/channel bindings, primary/backup ownership, and duplicate replay checks. Do not restart paused work or fabricate provider identities.
• S03, detector quality: Before expanding promoted coverage, retain precision >=0.90 and at least one true positive in both dev and holdout, without tuning on holdout, for the remaining JavaScript path, C# command/path, Java SQL/path, and C# SQL verifiers. Broader owned analysis calibration and production promotion remain separate; deterministic issue processing and pack verification do not certify these detectors. #2237 fixes the qualification predicate to compare exact observed counts rather than rounded receipt precision, with 32 focused core/receipt-path passes and unchanged recorded outcomes. This arithmetic correction neither evaluates nor promotes the remaining detectors.

Deferred or conditional scope, not hidden completion work

Item	Treatment and observable condition for resumption
Ingestion capacity and pagination	Record reviewed-corpus p50/p95 stage/queue latency, throughput, growth, and failure rate before capacity claims. Add server pagination before raising the bounded collection window. Neither is proof of unlimited scale.
Automatic prior-version restoration	Keep containment separate. Resume only with Vidar-owned authority, higher-sequence tested recovery, no revoked-content resurrection, restart/duplicate convergence, two-phase audit, and independent effect.
Policy-authoring UI and physical policy tables	Later/conditional work in the governance owner; preserve current Operator API and StateStore CAS semantics. Review policy-key/crypto-role provisioning before enabling that add-on in a new topology.
Hub-managed lifecycle	Hub ledger has partial software but no complete install/update lifecycle. It is not a third available installation path or a prerequisite for source/offline use.
Public CLI publication	Protected release-batch workflow, first signed publication, and pinned public-install receipt remain productization work. Local source and signed offline paths do not depend on it.
EvaluationHost adapters	Dormant. If reactivated, admit results only as canonical operational cases; golden-answer success never grants promotion.
Broader knowledge and scan support	Non-Azure/delta/mirror/embedding expansion, arbitrary model fidelity, native coverage-guided fuzzing, and unmeasured detector classes require separate accepted scope and evidence.

Dependency-ordered execution plan

These are proposed work packages, not deployment authorization or a promise of calendar dates. Run no more than two independent active outcomes per FDAI maintainer. Each batch closes its owner ledger, assessment row, counts, and bilingual text together.

Wave	Work and accountable lane	Prerequisite	Exit
0	Evidence reconciliation, FDAI maintainer	This audit	Each work ID has an owner, current scope/revision, receipt requirement, and valid issue before tracked delivery; stale done/deferred markers are distinguished
1	K02-K04, knowledge owners; A01, Mimir/release owner; L01-L03, learning/evidence owners; H01, runtime owner	Reviewed small designs where public contracts/authority change	Focused source/composition/persistence and negative tests pass; independently review source work; retain observation mode
2	I01 -> I02; I03; X01; R01; D01; K05 preparation, release/installation/security/privacy owners	Selected immutable release, authorized target, approved identities/trust; relevant wave-1 source	Exact source, policy, grants, retention, and rollback bindings plus independent readback; missing prerequisites stay explicit
3	K01/K05-K07; A02-A04; D02; H02-H03; V01; L02/P6; S01 -> S02, installation and capability owners	Corresponding source/configuration ready, fresh explicit live authorization, and #2128 pause lifted for scans	Terminal correlated receipts cover positive, denied, unavailable, restart, expiry/revocation, duplicate, and recovery behavior
4	L04 and any elected S03 expansion, measurement/review owners	Preregistered cohort and thresholds; admitted source histories and complete windows	Independent comparable outcomes, denominators, uncertainty, guard metrics, and separate promotion decision

Reassessment rule: no mandatory checkbox is erased by narrowing the scope after testing. Mark Ready only when its accepted scope has implementation, focused verification, and every required owner gate. Otherwise keep Partial with exact remaining IDs; use Not assessed when evidence is unavailable rather than claiming a new defect. Routine per-installation configuration and explicitly deferred expansion remain visible without pretending they were validated.

2026-10-20 three-session reconciliation

This reconciliation separates Complete, Code complete / live unverified, Actual failure, External blocker, and Elapsed-time blocker. A result applies to the mandatory package as a whole, not only to one delivered source slice. No package below is Complete, so all 29 mandatory packages remain open.

Mandatory package outcomes

Package	Result	Current evidence	Next observation condition
K01	Code complete / live unverified	G02 now connects the standard worker and Core document-intent path without publishing private bytes.	Retain one approved production-profile corpus through normal upload, worker, current ACL, revocation, deletion, and authorized readback.
K02	External blocker	Protected-source export and binding contracts exist.	Supply approved source identities, credentials, immutable revisions, and independent readback without committing private content.
K03	Code complete / live unverified	#2296 closes the case- and whitespace-variant self-review defect; 93 focused tests and one loopback PostgreSQL test pass.	Bind an independently approved policy, deployed read grants, and operational promotion.
K04	External blocker	Extraction and answer mechanics have bounded tests.	Supply approved PDF, Office, OCR, Korean, table, image, and revocation corpora with independent answer-quality review.
K05	External blocker	Source-bound qualification tooling exists.	Select and approve the exact source, model, corpus, metrics, and holdout review before qualification.
K06	Code complete / live unverified	Local distillation and conformance gates preserve provenance and separate activation.	Retain the same checks on an approved production-profile corpus and installed path.
K07	External blocker	Accessibility requirements remain recorded.	Complete independent English and Korean keyboard, speech, disclosure, and screen-reader review.
R01	External blocker	The current rule lifecycle implements later governed rollout phases, but the complete operational list is not closed.	Supply approved provider sources, verifier identity, scheduled expiry/notifier, override source, replica convergence, activation, and frozen drift measurement.
L01	External blocker	Session 1 reran 220 unchanged learning/authority checks; four positive history inputs are still absent.	Retain external-change continuity, action/event-time watermark, exclusion history, and post-window reconciliation from authoritative sources.
L02	Code complete / live unverified	Local context-safety, revocation, restart, and negative paths pass.	Retain live queued and in-flight expiry/revocation, hold, recovery, deadline, and subscriber-isolation receipts.
L03	External blocker	Local retained-Pattern mechanics exist.	Supply one coherent trusted external issuer through frozen cohort, publication, restart, promotion, demotion, and acceptance.
L04	Elapsed-time blocker	Measurement contracts reject activity counts as improvement.	Complete preregistered independent baseline/treatment episodes, recurrence windows, denominators, uncertainty, and safety guards.
A01	External blocker	Release-ceiling code preserves unavailable provenance.	Supply authenticated installed-Release and current revocation input for #1822.
A02	Code complete / live unverified	Approval and evidence admission mechanics pass local negative and restart checks.	Retain current independent approval, application, audit, and effect receipts on an exact installation target.
A03	Code complete / live unverified	Standing-authority scope, expiry, and revocation guards exist.	Bind approved provider commit fences and retain live expiry, revocation, duplicate, and restart evidence.
A04	External blocker	Drill mechanics exist without granting authority.	Run separately authorized stop, rollback, recovery, quorum, and effect-verification drills.
D01	External blocker	Privacy gates and source boundaries exist.	Approve the real corpus, privacy schedule, provider terms, regions, and compliance crosswalk.
D02	External blocker	Retention and audit mechanisms have local coverage.	Retain deployed deletion, legal hold, access review, anchoring, incident response, and change-feed convergence.
H01	Actual failure	The current local snapshot exposes all 15 signals and eight fields without hiding gaps, but only seven signals have concrete exporter identities in this installation. Readiness is healthy, cold-start duration is stale, the other concrete signals are unavailable, and only version/configuration fields are available.	Bind the missing liveness installation input, seven remaining signal populations, five remaining field sources, primary/fallback delivery, and recovery drills, then retain a fresh complete snapshot.
H02	External blocker	Canary, recurrence, scheduler, and stabilization prerequisites have local tests.	Supply approved installation baseline/scope/job inputs and retain scheduled attempts, no-execute replay, authority lowering, and configured stabilization exit.
H03	External blocker	The five-minute cadence and 30-second consumed-audit check do not settle the seven owner decisions.	Version the canary, smoke, alert, audit, cold-start, stabilization, and load/budget operating policy.
V01	External blocker	Report rendering and source-isolation mechanics exist.	Retain authenticated per-datasource identity, cutoff, freshness, unavailable, timeout, partial-widget, format-rejection, and GET-only evidence.
I01	Complete	Keyless Trial, unchanged rerun, new full installation, and Trial-to-full transition evidence are retained.	Preserve the receipts and regression coverage.
I02	Complete	Source and signed offline paths are the only available distribution paths.	Keep azd-up.sh diagnostic-only and prevent it from becoming another installer.
I03	Complete	Exact-revision connected deployment, bounded recovery, managed-host continuation, inventory, health, and zero-change evidence are retained.	Preserve the v18 receipt chain and focused regressions.
X01	Code complete / live unverified	Bundle registration, disabled-first lifecycle, revocation, and typed routing exist. #355 closed as not planned without the mandatory lineage, and #2315 now owns the observable successor criteria.	Under #2315, retain one exact signed-artifact install, distinct reviewed enable, disable, revoke, restart, disclosure, routing, independent readback, and Saga audit lineage.
S01	External blocker	Installed acquisition, worker, and independent local acceptance code exists.	Keep #2128 paused until explicitly resumed, then retain one terminal installed request, exact revision, coverage, publication, and independent readback.
S02	External blocker	Scheduled processing, Incident, notification, and ownership handover mechanics exist.	After the pause lifts, bind approved identities and channels and retain one same-correlation scan or explicit no-issues result, Incident, notification, ownership handover, and duplicate-replay chain.
S03	Actual failure	Exact promotion arithmetic, labels, thresholds, and corpus remain unchanged. Catalog 1.8.0 promotes Java path traversal at precision 1.0 with TP 5/1 and FP 0/0; catalog 1.9.0 promotes C# SQL at precision 1.0 with TP 2/1 and FP 0/0. Java SQL still has dev precision 0.8824, and C# command/path still have no held-out true positive.	Improve Java SQL from dev only and supply independently labeled held-out C# command/path evidence. Keep the unchanged holdout, precision >= 0.90, and at least one true positive in both splits before any further promotion.

G03-G07 exit criteria

For #2256 closure, the operator selected the local development profile as the acceptance venue on 2026-10-22. The results below validate deterministic behavior, authorization boundaries, persistence, replay, and presentation locally. They do not claim an Azure deployment, production corpus, or elapsed production observation. Those operational qualifications are non-blocking and require a separate explicitly authorized outcome if requested.

Criterion	Result	Current evidence and next condition
G03	Validated locally	#2034 completed the independently authored and reviewed governed v3 protocol on exact merged commit 8fec8d423ca4dfbcca2802516dd5454bb8cd0352: 256/256 decisions completed with no wrong or unavailable selections, English and Korean answer accuracy reached 1.0, and latency stayed within the registered bounds. Durable PostgreSQL admission, explicit Var Owner approval, Thor registry promotion, Saga audit, and independent effect observation completed in the development profile.
G04	Validated locally	The local ontology and semantic round-trip batch passed 169 tests, the Console ontology batch passed 118 tests, and 35 loopback PostgreSQL document-intent tests passed. Together they preserve positive, unknown, unavailable, authorization-denied, secured-read, and withdrawal outcomes through the local request, Core, projection, and presentation boundaries.
G05	Validated locally	The 169-test ontology batch exercises complete and incomplete relationship projections, pending reconciliation, topology-history capture, replay, and fail-closed answers without inventing live Resources or edges. Local fixtures are acceptance evidence for mechanics, not evidence about any Azure inventory.
G06	Validated locally	Governed-learning and operational-evidence replay tests cover verified and mismatched outcomes, rollback receipts, recurrence accounting, immutable replay, and denial when independent evidence is missing. The result validates the local causal-evidence contract without claiming a production intervention.
G07	Validated locally	The document-ontology batch passed 460 tests across supported structured, PDF, Office, OCR, Korean, table, image, lifecycle, and review paths. The 35 PostgreSQL tests additionally prove exact approval, ACL denial, secured read, withdrawal, rollback, and empty post-withdrawal readback. The former 30-day and 500-proposal production observation targets are optional operational qualification, not local issue-closure gates.

Evidence reconciliation notes

• Historical tests: The assessment’s A-P cohorts and F06/F07 908+8 results are retained historical evidence, not newly executed results and not additive evidence for this audit.
• Stale completed task: Exact issue readback on 2026-10-18 shows #1061 closed as completed. Its older structured-retrieval publication checkbox is reconciliation work, not proof that the implemented extension is missing.
• Still open: Exact readback shows #1822, #2194, and #2128 open, and #458 reopened. The latest #2128 evidence retains the operator pause and all seven original criteria unchecked.
• Closed but unfinished: The owners record #355, #371, #372, #424, #632, #633, and #995 closed as not planned, not operationally validated. Schedule a new issue or the documented successor before delivery; this audit does not reopen them or remove their residual work.
• No unsupported renewal: No Azure/model call, scanner run, database access, deployment, promotion, issue mutation, or external recovery was performed for this audit.

Source-delivery hardening checkpoint

The first follow-up delivery is tracked by #2211. It implemented a repository regression guard, not the 29 feature packages. At that checkpoint, protected source download was unavailable. All 29 packages still remain open, but #2214 has since implemented K02’s local export contract and production binding as recorded above. #2217 implements the explicit K03 prior-policy API lane, and #2218 binds three H01 signals and fields. #2228 adds the fourth H01 signal, a process-scoped guarded-startup duration rather than fleet health. A01, learning integration, the remaining health sources and other source work are still incomplete; live, human, provider and cohort gates remain separate. A observation mode-only maximum cannot be replaced by an unsigned file or a test approval.

The owning test is already selected by CI and Pages. It now joins the specification, assessment, inventory and exact required-package column; rejects unfinished mandatory work behind Ready; and compares bilingual work IDs and completion. It does not authenticate receipts or prove that an author identified every obligation. Removing an obligation or changing scope still needs owner review.

Four direct critique rounds covered these twelve concerns each. Round 1 added the backlog guard. Round 2 corrected multiple-badge ambiguity, contradictory display labels, localized Not assessed handling, and incomplete aggregate checks. Round 3 corrected identity substitution and made mutation tests require the intended assertion rather than any thrown error. Round 4 applied the mandatory pre-publication gates, corrected seven display-terminology violations, and repeated the bounded review without changing readiness or closing any feature package.

Concern	Final bounded result
Accepted scope and authority	No runtime, approval, protection or execution behavior changed
Stable function identity	Exact F10-F15 ID set is checked, not merely 21 arbitrary rows
Complete inventory	Missing, extra and duplicate function rows are rejected
Mandatory completion	Ready with an unchecked required package is rejected even if document states agree
Valid completion transition	Completed packages allow Ready without deleting its obligation history
Work identity	Missing/duplicate package IDs and unknown references are rejected
Exact mapping	Duplicate required IDs, ranges and missing required columns are rejected
Bilingual parity	Current states, prior states, package references and completion match
Badge cardinality	Missing or multiple badges cannot hide a second visual state
State and display meaning	Unknown states and contradictory labels fail; localized Not assessed remains valid
Aggregation	Group and specification/assessment totals cover every present state
Regression and execution venue	Malformed transitions, mismatches and untracked Partial rows fail with intended assertions; existing CI runner selects the test

Final focused command: node --test site/test/functional-readiness.test.mjs, 25 passed, no skips. No confirmed above-Low defect remains in this bounded guard after round 4. This is not a Low-only assessment of the 29 unfinished packages, the entire product, or a live installation.

Related docs

To learn about	Read
Product requirements and corrected badges	Functional specification
Per-function decisions and historical test evidence	Readiness assessment
Normative owners rather than this cross-cutting index	Roadmap index