name: adversarial-verification-loop description: "RIG doctrine — Adversarial Verification (non-vacuity) as a code-enforced loop. For every gate you trust: name it → plant a defect and prove the gate goes RED → restore and prove GREEN → keep the planted failure as a regression fixture → seal. A green that won't go red is theater." when_to_use: "After adding any test, gate, hook, guard, or checker; before trusting a 'green'. Say 'plant a failure', 'red-team the gate', 'non-vacuity', 'does this test actually test anything'." version: 11.0.0 owner: rig-intelligence domain: Agent engineering / operating doctrine status: Stable verified_by: jake verified_on: 2026-09-01 outcome: "agent-engineering: every trusted gate is proven able to go RED, then GREEN, with the planted failure kept as a fixture" done_when:

• "cmd:python3 bin/rig-loop test adversarial-verification-loop"
• "file:capabilities/skills/adversarial-verification-loop/loop.yaml"

great_example: examples/great/agent-engineering/iqrsqpi-v11.md quality_floor: 0.8 metadata: loop_skill: v11 engine: rig-loop doctrine_source: "RIG-Doctrine-Repo/doctrines/ADVERSARIAL-VERIFICATION.md; ~/.claude/rules/rig-build-doctrine.md rule 3" hooks: PreToolUse:

• matcher: "Write|Edit|MultiEdit|NotebookEdit|Bash"

hooks:

• type: command

command: "python3 ~/Developer/rig-capabilities/bin/rig-loop hook pre-tool --format claude --skill adversarial-verification-loop" timeout: 20 Stop:

• matcher: "*"

hooks:

• type: command

command: "python3 ~/Developer/rig-capabilities/bin/rig-loop hook stop --format claude" timeout: 20

Adversarial verification loop

G Guard → R Red (plant, expect FAIL) → N Green (restore, expect PASS) → K Keep fixture → P Proof
   ▲            │ red did not fail          │ green did not pass
   └────────────┴───────────────────────────┘  (max 3 → escalate: the gate is theater)

Step 0

rig-loop activate adversarial-verification-loop --intent "prove <gate> is non-vacuous"

Stages

Why the engine, not the prose

rig-loop plant runs both commands itself and records the exit codes. If the "red" command exits 0 the plant is recorded as red_ok: false and stage R stays red — the gate you trusted does not catch the defect. That is the finding; report it, fix the gate, loop.

Honesty notes

• Plant a real defect in the code or data the gate guards, not a false command.
• Keep the fixture in the repo's test suite; the K gate only checks you named where.