Recorded gaps

Twelve known, reasoned-about gaps in muster’s current behavioral/static checks, each with the schema id / title / evidence / what-was-tried / why-left / closes-when / status. An entry missing any of these seven fields is incomplete — that is exactly what scripts/check-register-schema.mjs mechanically enforces, so this register cannot silently decay into a bare-titles TODO list.

status is one of tracked-defect (has, or should have, an open issue; not yet resolved) or accepted-tradeoff (a deliberate, possibly-permanent design choice — closing the underlying GitHub issue does not mean the entry should be deleted).

---

RG-001 — Live control gate satisfiable by a dead endpoint

• id: RG-001
• evidence: verdict.passed (tests/cts/skills-suite.test.ts:422-425) — the live-model discrimination-control assertion only checks passed === false; garrison-hq/muster#76.
• what-was-tried: none yet — filed as an issue, not fixed.
• why-left: fixing it is a code change to a test file, out of this docs-only mission’s write scope.
• closes-when: the assertion additionally requires runsErrored === 0 — the field already exists (runsErrored, src/adapters/skills/types.ts:170) and is asserted elsewhere, just not in this file.
• status: tracked-defect

RG-002 — isControl exit-code semantics diverge between skills and a2a

• id: RG-002
• evidence: applyControlInversion (src/adapters/a2a/index.ts:356-371) flips a firing control to passed: true so a healthy a2a run exits 0; skills’ healthy run exits 1 instead (the control is designed to fail and is never inverted); garrison-hq/muster#77.
• what-was-tried: none — this is a by-design difference between the two adapters, not yet reconciled.
• why-left: reconciling requires a behavior decision (which adapter’s convention is canonical) that is out of scope here.
• closes-when: a muster FR picks one convention and the other adapter is migrated to match it.
• status: tracked-defect

RG-003 — Single-tool bias in the should-trigger axis

• id: RG-003
• evidence: gradeAxis (src/adapters/skills/trigger.ts:237) — the aggregate-rate calculation that produces the same 1.0 rate for a placeholder-description skill (10/10) as for a purpose-built weather skill (30/30); garrison-hq/muster#82. Not literally identical raw scores — the shared quantity is the normalized rate, not the count.
• what-was-tried: none — this is the grader’s current, shipped aggregate-rate design; no alternative discriminative axis has been prototyped.
• why-left: fixing requires a richer discriminative axis design sensitive to description quality, independent of tool-set cardinality.
• closes-when: a should-trigger axis exists that discriminates by description quality rather than only by aggregate trigger rate.
• status: tracked-defect

RG-004 — Crosslayer contradiction detector’s accepted false-negative surface

• id: RG-004
• evidence: stripHtmlComments (src/crosslayer/contradiction-lint.ts:374) — one of PR#85’s two fixes; docs/rubric/crosslayer-contradiction-gate.md’s own “Accepted false-negative surface” section documents the deliberately-accepted remaining gap; garrison-hq/muster#84 (closed), PR#85 (merged).
• what-was-tried: HTML-comment-body-leakage stripping and a subject-matter gate; a broader semantic-similarity gate was considered and rejected as out of scope.
• why-left: the tradeoff is accepted by design.
• closes-when: never, by design, unless a semantic (non-lexical) similarity check is added as new scope.
• status: accepted-tradeoff

RG-005 — Heartbeat behavioral cases time out once an endpoint is configured

• id: RG-005
• evidence: runManifest (src/adapters/heartbeat/index.ts:489) — with MUSTER_ENDPOINT set, 10 of 118 tests time out at vitest’s default because behavioral cases run whenever an endpoint is configured, including fixtures that look static; garrison-hq/muster#75.
• what-was-tried: none — the timeout is only observed once MUSTER_ENDPOINT is set, which ordinary CI does not do, so it has not yet forced a fix.
• why-left: needs either a per-test timeout override or a static/behavioral split in the fixture suite.
• closes-when: one of those two fixes ships.
• status: tracked-defect

RG-006 — No published rubric for four adapters

• id: RG-006
• evidence: MUSTER_RUBRIC_CITATION (src/crosslayer/contradiction-lint.ts:36) — crosslayer’s findings cite an in-code string constant rather than a published rubric document.
• what-was-tried: none — “every check cites a rubric” is currently aspirational for tools, memory, heartbeat, and crosslayer.
• why-left: publishing a rubric for these layers is not this docs-only mission’s job.
• closes-when: each adapter publishes its own rubric document and the rubric index is updated to list it.
• status: accepted-tradeoff

RG-007 — behave/a2a and skills/sop disagree on what a dead endpoint means for the exit code

• id: RG-007
• evidence: endpoint-fatal exit 2 in doBehaveRun (src/cli/index.ts:479-489) and doA2aBehavioralRun (src/cli/index.ts:1157-1162) vs. exit 1 on any failure in doSkillsRun (src/cli/index.ts:1580-1584) and doSopRun (src/cli/index.ts:1684-1686); garrison-hq/muster#78.
• what-was-tried: none — discovered during this mission’s own verification pass.
• why-left: this is not two valid per-adapter designs left as an intentional divergence. contracts/cli.md specifies one uniform rule: an endpoint unreachable for the entire run is an execution fault, exit 2. behave and a2a implement that contract; skills and sop do not. This is an open product question: either the contract should be amended to bless a documented per-adapter divergence, or skills/sop should be fixed to match the contract they already claim to implement.
• closes-when: a muster FR either amends contracts/cli.md to document a per-adapter exception, or migrates skills/sop to exit 2 on total-endpoint-failure.
• status: tracked-defect

RG-008 — Judge OR-of-two-positions leniency

• id: RG-008
• evidence: verdictA || verdictB (src/adapters/openclaw-sop/judge.ts:264-266) — a k-of-n run passes if either the “Answer A” or “Answer B” order-swap call voted PASS, not requiring both to agree.
• what-was-tried: none — this is the judge’s current, shipped design.
• why-left: requiring both swap positions to agree would materially change the judge’s leniency/strictness balance.
• closes-when: a muster FR either affirms OR-of-two as intentional or requires both positions to agree and migrates the implementation.
• status: tracked-defect

RG-009 — xfail mechanism is a fixture convention, not a manifest-level feature

• id: RG-009
• evidence: xfail_discrimination_control (examples/behave/manifest.yaml:34-45) — the shipped discrimination-control case is named with an xfail_ prefix and a comment, but nothing in the manifest schema or runner recognizes xfail_ as special.
• what-was-tried: none — the naming convention has been sufficient so far.
• why-left: a first-class xfail manifest field is a schema/runner change.
• closes-when: the manifest schema gains an explicit xfail field and the runner reports xfail cases distinctly from ordinary failures.
• status: tracked-defect

RG-010 — Skills static cases never compare expectations.violations

• id: RG-010
• evidence: c.expectations.ok (src/cli/index.ts:1323) — a static skills case’s passed field is derived solely from ok === c.expectations.ok; any expectations.violations a manifest author writes is never compared against the actual lint findings.
• what-was-tried: none — the static runner’s ok-only comparison predates this mission; no comparison logic for expectations.violations has been attempted despite fixtures already declaring violation shapes.
• why-left: adding a violations-shape comparison is a behavioral change to the static skills runner.
• closes-when: runStaticSkillCase additionally compares c.expectations.violations (when present) and fails the case on a mismatch.
• status: tracked-defect

RG-011 — MUSTER_BASE_URL deprecation is skills-only

• id: RG-011
• evidence: usedDeprecatedAlias (src/cli/index.ts:1376-1380) — only the skills adapter’s endpoint resolver emits the deprecation warning; every other adapter reading MUSTER_BASE_URL does so silently.
• what-was-tried: none — the warning was added only at the one call site (resolveSkillsBehavioralEndpoint) when the alias was introduced; no attempt has been made to thread it through the other adapters’ endpoint-resolution paths.
• why-left: adding the same warning to every other adapter is a behavioral change.
• closes-when: the deprecation warning is emitted from one shared helper every adapter’s endpoint-resolution path calls.
• status: tracked-defect

RG-012 — SOP static-drift findings are always severity: "warning", never "error"

• id: RG-012
• evidence: checkRuleTextPresence (src/adapters/openclaw-sop/manifest.ts:422-441) hard-codes severity: "warning"; f.severity !== "error" (src/adapters/openclaw-sop/index.ts:156) means static drift can never by itself make a manifest run fail.
• what-was-tried: none — this is the current, shipped severity classification.
• why-left: promoting static drift to error is a severity-policy decision that would need its own justification.
• closes-when: a muster FR either affirms warning-only as intentional or promotes drift findings to error severity.
• status: tracked-defect