servo-gate

Verify an artifact against ground truth, using a non-colluding reviewer: one that did not produce the artifact and judges it only against a named oracle. A gate that does not read an independent oracle is theater — this skill refuses to run without one.

The rule underneath: a check is only worth running if it brings in information the producer did not have. Re-reading the artifact with the same context that wrote it is the agent grading its own homework.

What you need

• The artifact: the text under review (a spec, a plan, a diff, a single claim).
• The oracle key: which source of truth to check against (e.g. scope, invariants, acceptance). The project's keys live in .servo/manifest.yaml.
• The stance (default refute): the posture the reviewer takes.

Procedure

1. Resolve the oracle. Read .servo/manifest.yaml in the project. Find the key under oracles:. Its value is the source — where the ground truth lives (a file path, git:main, vault:..., linear:...). If the key is not in the manifest, stop and report the available keys; do not guess.
2. Read the source → the oracle slice. Read the source yourself (Read tool, MCP, git, whatever the source type calls for) and capture the relevant content. This text is the oracle slice. If you cannot read it, stop — never invent ground truth.
3. Refuse an empty oracle. If the oracle slice is empty or whitespace, stop: a gate with no ground truth is theater. This refusal is by design (§7b).

4. Build the verifier prompt by filling this exact template:

   You are a non-colluding "<STANCE>" reviewer.
   You did NOT produce the artifact. Judge it ONLY against the ground truth below.
   If the ground truth does not support a claim in the artifact, flag it.
   
   ## Ground truth (authoritative oracle)
   <ORACLE SLICE>
   
   ## Artifact under review
   <ARTIFACT>
   
   Reply with JSON only: {"verdict":"GO|FIX|STOP","findings":["..."]}

5. Dispatch the verifier — the real non-colluding step. Spawn a subagent with the Agent tool and give it exactly the prompt from step 4 and nothing else. It must not see your reasoning or how the artifact was produced. When more than one model is available, run the verifier on a different model than the one that produced the artifact (§7d) to cut shared bias. Tell it to reply with only the JSON object.
6. Read the verdict. Parse the reviewer's JSON: {"verdict": "GO" | "FIX" | "STOP", "findings": [...]}. If the response is not valid JSON, or verdict is missing or not one of the three values, treat the result as FIX and record that the verifier response was malformed.
7. Report the verdict and the findings, naming the oracle key that was checked.

Tiering — keep it cheap (§7e)

• Default: ONE verifier (steps 5–6 once), ideally on a cheaper model than the producer (§7d gives you the independence; you do not need the strongest model to check against a named oracle). Most gates are low-stakes.
• Escalate to a panel of 3 verifiers (steps 5–6 three times) only when the artifact is high-stakes: irreversible, security-sensitive, a production migration, or the human flagged it. Keep any dissenting finding; take STOP if a majority stops.

Do not panel by default. And do not reach for a heavy multi-reviewer review (a 4-5 agent panel-with-critic) as your everyday gate — that is a high-stakes tool. A single oracle-grounded verifier already catches most stale-premise and scope problems; spend the panel only where being wrong is expensive. A panel on a low-stakes gate is the ceremony this skill exists to avoid.

Rules

• No oracle, no gate. Step 3 is not optional.
• The verifier never sees the producer's reasoning. That separation is what makes the verdict worth anything.