Factory docs, home
Page navigation

Pre-release documentation. Proposed designs are labelled in the page; acceptance evidence is maintained in the build review. Local paths appear as placeholders. Current state: Build review.

W31 conditional design reviewed; no source commission. Independently adoptable Verification remains unimplemented. The reviewed direction is to extract the existing D4 verification routine behind a caller binding, preserving one owner for dispatch, allowance and recovery. W30 implementation is not a prerequisite. Design evidence pins the reviewed proposal and its limits; the W31 packet records the next gates.

What exists

D4/R7 at 42d2700 remains a draft with outstanding composition, isolation, ownership and acceptance checks. Its verification routine already combines ProductGate, DeliveryLedger, Repository, native fencing, bounded collection, branded Evidence and Assurance. Reuse that responsibility; do not add a competing job ledger or dispatcher. Existing collector adoption proves a lower-level entry, whose repeated calls can launch repeated collection. It is not durable Workflow replay.

The proposed neutral data sketch records the original Product/Slice/Candidate, derived check/run identity, original allowance and historical proof separately from current validity. Its public type binding uses the actual RepositoryVerdict rather than another success schema. Neither declares an executable admission or settlement API.

Two missing contracts

External Candidate admission. Current ProductGate admits a check only when a matching sealed production admission exists; otherwise it refuses PRODUCTION_UNKNOWN before changing the gate. A Repository seal alone is insufficient. A separately reviewed rule in the existing gate must authenticate external origin, actual Repository/store identity, Product and frozen Slice/base/scope, independently owned acceptance/recipe/environment/input snapshots, current registry/grant and collector resource closure. Candidate output cannot appoint its evaluator or fabricate production provenance. Preserve old histories and the same original check/run, deadline and allowance; alternate origins or stores cannot reset them.

Original-dispatch proof discovery. D4 reopens a Verdict only after an observed dispatch names its digest. DeliveryLedger stores Verdicts by digest; it has no dispatch-to-proof lookup. Verdict retention and gate outcome are separate writes, so an orphan proof remains inert. Lost-ack discovery/reconciliation requires an unimplemented bounded rule and ABI in the existing owner, binding the original Product/Slice/Candidate/check/run/dispatch/epoch to one proof and its pins. Require current read authority and actual predecessor exclusion; distinguish missing, conflicting, corrupt and unavailable witnesses. A scan or caller assertion cannot establish the link. Missing acknowledgement grants no success, fresh job, collector retry or budget reset.

Evidence and completion gates

Verification retains Verified, Failed and Inconclusive: known failure remains Failed even with missing evidence; other insufficient evidence is Inconclusive. Historical Verdict meaning is separate from current validity. A failed current read cannot erase history or claim corruption. Verified is a local repository judgement, never integration, model, release or live-delivery authority.

Design evidence covers strict neutral/public type compatibility and 15 installed leaf probes using real EvidenceStore, Journal and native fence. Candidate identities and observations were synthetic; no Git seal, collector process, D4 Workflow or external admission was qualified. Resource declarations retain their explicit Node dependencies and source/compiled recipe pairing.

Before implementation: accept D4/R7, independently review both missing contracts and the extraction ABI, then commission an actual external tarball consumer with two qualified caller storage/authority arrangements. Prove real external sealing/collection, independent acceptance, current revocation, stable allowance through races and crashes, original-ID recovery, resource replacement, Evidence corruption and cross-Product isolation. Independent adoption, Knowledge, Fable/Opus, CI and live gates remain unchanged.

UI evidence

Design only; nothing implemented or installed. Browser and UI checks would reach Verification through a planned UI collector (roadmap X4) that stores one sf-ui-observation/1 Evidence record per acceptance check at its pinned stage. UI observation defines the record, its closed codes and judging rules U1 to U14. In short: the record holds runner facts only, and the judge decides each check against its pinned stage: a before check must fail on the base, and after and guardrail checks must pass on the Candidate. Only fixed locator, expect and pinned axe checks can pass or fail. Model-judged checks map to the reserved review check and stay Inconclusive. A stale or missing pinned replay, a pass only on retry, and an exceeded step budget or timeout are each Inconclusive with their own code. Artefacts are hashed into Evidence under the run's pins. The runner needs a reviewed confined boundary, a site allowlist and telemetry off. A run whose inputs, runner or browser differ from the pins does not start, and such a record never qualifies. Exploratory findings stay upstream and untrusted until a repro fails on the base and passes on the Candidate. The current Workflow does not accept local-ui Verdicts: a versioned collector/judge binding requires its own independent review. Not established: that binding, any runner, the install (owner decision PD-25) or a confined browser.

Source: docs/verification-workflow.md