chore(decisions): propose ADR 0355 — skill doctor evaluation evidence ships separately from its gated production import

[creosB]

Proposes ADR 0355 — "Skill Doctor evaluation evidence ships separately from its gated production import" (status: proposed).

Tracks the decision context for:

• skill-doctor fabrika import: five blockers before production #8035 — the Skill Doctor epic, ruled 2026-09-05 (comment 5555033755): pitch approved, children land skill-doctor: byte-exact upstream vendor, fabrika-shaped, tests in CI #8048–rule skill-doctor's report share posture and record it as an ADR #8054 sequenced; the production import proceeds through that approved sequence, each child its own reviewed change
• chore(benchmarks): skill-doctor evaluation PoC evidence #8037 — the draft evidence PR (sanitized docs + the OpenCode adapter), which this ADR separates from that import

What the ADR decides (if accepted)

The sanitized Skill Doctor evaluation and its OpenCode adapter are reviewable as their own change (#8037); the production import ships separately, as the founder-approved epic sequence on #8035 — packaged import first (#8048, open as draft PR #8063), collector next (#8049). Support for importing on discussion #7319 approves the direction only — no implementation choice. Evidence claims keep their scopes attached (25/25 tests under PYTHONUTF8=1 ≠ native-Windows-encoding fixed; the B+ over 2 sessions = pipeline executes, not grading validated; the ≥20-session rerun minimum is owned by child #8053 — a pass means corpus measured, not grading validity). Binding provenance ledger + pinned upstream commit; byte-exactness is today's posture, subordinate to the adaptation-policy call: the two-target test adaptation is proposed and awaits maintainer acceptance (the "Proposed adaptation" section of #8063). Open rulings remaining: the report share posture (#8054, ready-for:human) and the content-ingestion trust posture (#4859) — approved by neither this ADR, #8037, nor #7319. The report's branding default is already ruled on #8035 (fabrika-branded, internal).

Landing status (as of 2026-09-06)

#8063: tests/lint passed; leak scan ran and failed on path-policy matches — ~/.claude literals in authored docs (sanitized in e715358e) and in the byte-exact upstream references/supported-harnesses.md, held unchanged pending an upstream text change. The leak-guard infrastructure defect behind one vacuous green is fix-proposed in draft PR #8108.

Diff

One file: .decisions/0355-skill-doctor-evidence-before-import.md (new, 138 lines). No other surface touched.

Gates

• adr mint allocated 0355 against fetched origin (341 merged records, 0 in-flight ADR PR claims at fe039d1b); adr resolve 0355 rechecked absent immediately before push.
• adr resolve 0300 → live (citation filename verified); adr sweep --new 0355 on the final text → shortlist judged, no contradictions, no supersede/amend.
• .decisions/ is a governed root: per CODEOWNERS + ADR 0330 this PR takes control-plane team approval (humans-only; mechanically inert until the main ruleset enables require_code_owner_review, but the convention is the gate).

Awaiting required control-plane review; not approved for merge. Acceptance is the control-plane call (ADR 0330).

[github-actions]

No preview deploy

• No preview deploy for this PR — its diff touches no deploy-relevant path, so no preview stack was minted and e2e is not applicable. _(da4a666)

• web — Stage pr-8055 torn down.

[github-actions]

heal-ci: ROUTED — PR #8055 @ da4a666 → author

Scheduled stall sweep: this pull request classifies as linkage-refused, stranded 120 minute(s) at head da4a666e73f8d17a91117a3b9720428a509bca98.

Detection only — this run merged nothing, re-ran nothing and spawned nothing. What to do about the flag is a driver decision; /fabrika:heal-ci 8055 reads the full diagnosis.

Posted by the heal-ci-sweep workflow (#6146). The suppression key <pr>:<class>:<head> rides in the marker below, matched over the whole comment history: no further note lands here until this pull request changes class or gains a new head commit.

[usirin]

governance: PASS @ da4a666 content:1704a65930a2 — separate evidence and import; no guard weakened

Questions reviewed: may sanitized Skill Doctor evidence land separately from its production import; what does the recorded approval authorize; and does a two-session grade validate grading quality?

Sweep outcome: shortlist. Read the relevant scope and decision passages in all eight candidates. ADRs 0252, 0253 and 0243 govern the distinct review-eval measurement chain; this proposal neither replaces that chain nor turns Skill Doctor scores into merge verdicts. ADRs 0250, 0313, 0173, 0348 and 0388 address hook failures, queue waits, release reachability, Tuval commands and worktree isolation; none is changed by this record. Live status was confirmed with adr resolve.

The independent domain read covered live ADR 0300 and its amendment 0400. The cited founder comment #8035 (comment) rules internal-by-default branding and defers sharing. The approved plan and plan-gate comments on that issue establish the packaged-import dependency before the collector. The record leaves sharing and content-ingestion trust undecided, and bounds the two-session run to execution evidence rather than grading validity. No standing decision is contradicted.

Guard scan: no-anchors-in-reach, zero anchored invariants. This additive decision file edits no gate, workflow or exemption list; no gate invariant is in this diff's reach. Its future implementation obligations do not authorize skipping review or changing protected upstream files silently.

Self false; the self-review fence does not apply.

[usirin]

review-doc: PASS @ da4a666 content:1704a65930a2 — scoped evidence claims and cited decision record

This is a conversation-authored decision record with no issue acceptance criteria to bind. Reviewed the entire added ADR against the doc rubric.

Check	Result	Evidence
Placement and Diataxis	PASS	Single-mode explanation. The record explains why evaluation evidence and production import proceed separately, with alternatives and consequences. The child list records the approved dependency order; it is not an operational recipe.
Status and claims	PASS	Proposed status is explicit in the consequences. The founder ruling and approved plan on #8035 support internal-by-default branding, deferred sharing, and import-before-collector dependencies. The test claim retains PYTHONUTF8=1, and the two-session B+ claim proves execution only.
Writing-for-agents	PASS	The opening names the decision, binding constraints group the implementation limits, and unresolved sharing and trust questions have separate pointers. No competing instructions or unsupported implementation approval found.
CI-owned hygiene	PASS	review ci settled green at the reviewed head: 44 checks, 36 success and 8 skipped; 27 repository workflows inspected the head. No local reproduction of CI checks.
deviation-disclosure	N/A	The section is absent. This additive, conversation-authored ADR is not a build implementation PR, so it does not owe a build-authored deviation disclosure. No issue-link omission is being used to exempt code or skill changes.

Governance was independently judged and posted at this head. No code, runtime state, or tests change, so the silent-failure, type-design and test-gap sweep found no applicable code findings.

Verdict-written: 2026-09-20T20:09:06Z