spec-verification-criteria

SKILL: spec-verification-criteria

Domain: Agent specification verification — 7 VC criteria and independence rules
Loaded by: agent-specification-verifier
Governed by: DLMS-2026-0115 GATE_05, DLMS-2026-0114, DLMS-2026-0108 R07–R08

VERIFICATION_SEQUENCE

1. Confirm identity independence BEFORE any other step
2. Retrieve agent-instruction-record from TypeDB via agentcreation.getArtefact
3. Read deployed instruction file at .github/agents/{role_id}.agent.md
4. Retrieve agent-spec from TypeDB for content comparison
5. Evaluate all 7 VC criteria — record pass/fail for each
6. If all 7 pass: write agent-verification-record with result: PASS
7. If any fail: write agent-verification-record with result: FAIL + failed_checks + findings
8. Call agentcreation.submitVerification

Do NOT edit the instruction file under any circumstance. Return FAIL with findings.

INDEPENDENCE_CHECK (DLMS-2026-0108 R07)

Must be asserted before any VC evaluation begins.

1. getArtefact(role_id, 'agent-instruction-record') → content.author_actor_id
2. Compare with own verifier_actor_id
3. If verifier_actor_id == author_actor_id → return result: BLOCKED
   (cannot verify own work — escalate to agent-creation-orchestrator)
4. If identities differ → record independence_confirmed: true and proceed

Partial PASS is not permitted. BLOCKED is not PASS. Do not proceed past independence check if blocked.

VC CRITERIA (7 mandatory, all-or-nothing)

VC-01 vs VC-06 DISTINCTION

VC-01 asks: is the instruction file structurally valid per DLMS-2026-0114 FORMAT_A?
VC-06 asks: does the instruction file faithfully represent the approved spec?

VC-01 is a format check. VC-06 is a fidelity check. Both must pass independently.

VC-06 failures (invented content examples):

• A READS entry that does not appear in spec CONTEXT_BOUNDARY
• A tool in tools[] absent from spec TOOL_ASSIGNMENT
• A MUST rule not derivable from spec STAGE_GATES or KNOWN_CONSTRAINTS
• A NEVER prohibition omitted from spec's CONTEXT_BOUNDARY/NEVER

VERIFICATION-RECORD CONTENT

The agent-verification-record TypeDB payload written by agentcreation.submitVerification:

{
  "role_id":                       "{role_id}",
  "instruction_artefact_id":       "{UUID of agent-instruction-record}",
  "verifier_actor_id":             "{this agent's actor_id}",
  "instruction_author_actor_id":   "{author_actor_id from agent-instruction-record}",
  "independence_confirmed":        true,
  "vc_checks_passed":              7,
  "failed_checks":                 null,
  "result":                        "PASS",
  "findings":                      null
}

On FAIL:

{
  "vc_checks_passed": 5,
  "failed_checks":    ["VC-06", "VC-04"],
  "result":           "FAIL",
  "findings":         "VC-06: instruction READS block includes registry.getDocument not in spec CONTEXT_BOUNDARY. VC-04: agentcreation.writeSpec declared in tools[] with no corresponding MUST step."
}

Gate for agentcreation.submitVerification: assertAgentArtefactExists(role_id, 'agent-instruction-record') (no requiredContent check — the verification record itself is what records result)

GATE_07 SATISFACTION CHECK (VC-07 detail)

VC-07 requires all 7 DLMS-2026-0115 gates to be satisfied. Query sequence:

1. assertAgentArtefactExists(role_id, 'agent-request-record')           → GATE-01
2. assertAgentArtefactExists(role_id, 'agent-problem-analysis')         → GATE-02
3. assertAgentArtefactExists(role_id, 'agent-spec') +
   assertAgentArtefactExists(role_id, 'agent-spec-review', { result: 'PASS' }) → GATE-03
4. File .github/agents/{role_id}.agent.md exists                        → GATE-04
5. (this verification is GATE-05 — verified by VC-01 through VC-06 passing)
6-7: GATE-06 and GATE-07 are post-verification — not yet evaluable at this stage

VC-07 passes when GATE-01 through GATE-04 are all satisfied. GATE-06 and GATE-07 are downstream of verification and are not checked here.