07 — Capability Verifier Evidence Review (SUPERTRACK G)

#	Requirement	Design answer	Verdict
G.1	Capability is not function-exists	"no … branches on capability identity"; capability = behavioral+operational (11,10)	INTENT-OK (explicitly rejects existence proof; answers FIX4/FIX5)
G.2	Capability is not free-text proof	FIX5/FIX6 lineage: behavior+evidence, never free-text; FIX7 "normalized capability measurements" (10)	INTENT-OK; measurement schema not specified → NEEDS_CLARIFICATION
G.3	Requires verifier identity	"manifest/FK principal/action/tier contracts"; verifier-bound evidence	INTENT-OK; verifier-FK schema not specified
G.4	Requires behavioral test	implied by "behavioral … operational evidence"	NEEDS_CLARIFICATION — the actual behavioral test (keyset/resume/perf) definition is not written for FIX7
G.5	Requires operational evidence	"operational evidence", "fresh post-activation evidence" (10,checkpoint)	INTENT-OK; evidence row schema + "fresh" definition (time/epoch window) not specified
G.6	Directus cannot self-attest verified=true	runtime/Directus zero DML on controlled evidence (02)	INTENT-OK (target); LIVE-FAIL — has_table_privilege(directus, capability_operational_evidence, INSERT)=true
G.7	Missing keyset/resume/perf evidence blocks readiness	FIX6: typed evidence empty → 0/3 → scale NOT_SAFE; FIX7 inherits	INTENT-OK; live confirms qt001_capability_operational_evidence=0 → 0/3 today

Live state

qt001_capability_operational_evidence = 0 → capability scores 0/3, scale = NOT_SAFE today. This is the correct fail-closed baseline. But Directus can INSERT into that table (proven), so under the current ownership the "0/3" can be flipped by the app role — i.e. capability is self-attestable today. FIX7 closes this only after owner cutover + REVOKE.

Is there fake capability proof?

Target design: no fake proof — capability requires a behavioral test result + operational evidence + verifier identity, under owner-only storage; existence and free-text are explicitly rejected. This is the correct answer to Codex's FIX4 "to_regproc/to_regclass existence only with literal-true negative test" finding.

As written: unverifiable + one live hole. The behavioral-test definitions (what a passing keyset/resume/perf test actually asserts, against what scale, producing what measurement), the measurement normalization, the verifier-FK, and the "fresh post-activation" window are not specified. And live, the evidence table is Directus-writable, so capability is self-attestable until cutover. Required correction: a capability spec defining each measurement (id, behavioral test, threshold, evidence schema, verifier-FK, freshness window) and the owner-only storage + REVOKE that prevents self-attestation.

Capability verdict

No fake capability proof in the design intent; capability is self-attestable live until cutover; FIX7 capability spec is incomplete. → NEEDS_CLARIFICATION + LIVE-FAIL-until-cutover. Real scale/keyset/resume/perf evidence remains a 2.6B prerequisite (consistent with all prior memory). Contributes to the block.