Deep mode: two adversaries

Deep mode: two adversaries

At the deep band, and only there, the draft answer is attacked by two children with asymmetric inputs and opposite defaults. One asks whether the claims are true. The other asks whether the claims are all of it. They are separate jobs and a single reviewer collapses them into one — usually into the first, because refuting a written claim is easier than noticing a surface nobody looked at.

Never run either on a light-band question. An adversarial round over a four-hit lookup costs more than the answer and produces a confident dispute about nothing; escalation is scored as a failure in the suite this family is measured against, not as diligence.

The refuter

Input	The claims and their citations. Nothing else — not the reasoning, not the children's summaries, not the question's framing.
Default verdict	REFUTED. The claim is wrong until an artifact shows otherwise.
Concession rule	It may concede only by naming the artifact it opened and what that artifact showed. "Seems right" is not a concession; it is the default, restated.
Output	Per claim: refuted with the counter-evidence, or conceded naming the artifact, or unresolved (needs X).

The withheld reasoning is the mechanism. A reviewer given the argument evaluates the argument; a reviewer given only the conclusion and the citation has to go and look.

The coverage auditor

Input	The question, and the manifests — the file listing, the package list, the route or job registries, the deploy targets. Not the answer.
Default verdict	INCOMPLETE.
Concession rule	It may only conclude "covered" after failing to name an uncovered surface. To claim incompleteness it must name a specific surface and one probe that would settle it.
Output	Named uncovered surfaces, each with the one command or observation that would cover it.

Withholding the answer is what makes it useful. An auditor that has read the answer audits the answer's own frame, which is the frame that produced the gap.

Adjudication

• Two rounds maximum. Round one runs both adversaries; round two runs only the ones with live objections.
• Evidence only. An objection is resolved by opening the artifact, not by re-arguing. If neither side has an artifact, the row is unresolved (needs X) and X is named.
• A conceded claim is recorded conceded, not validated. A refuter that ran out of objections has not verified anything; it has failed to refute. The distinction matters to whoever reopens the question, and it is the difference between "we checked" and "nobody found a problem in twenty minutes".
• Uncovered surfaces the auditor names are either covered in round two or written into what this run could not see. They do not evaporate because the round budget ended.

This is not a code review

One line, because it gets confused constantly: this gate has no pull request, no pushed branch, and no completed implementation. It attacks an answer, not a diff. The review loop over a real PR belongs to request-blocks-review and blocks; running the adversaries instead of that loop reviews a claim about code that nobody has written yet.

Degradation

A harness that cannot withhold parent context from children cannot run this design: both adversaries would see the reasoning, the answer, and each other, which removes the asymmetry that makes their verdicts mean different things.

In that case run one reviewer, keep the REFUTED-by-default posture, and say in the answer and in the record that the two-verdict design degraded to a single reviewer — so the coverage question is visibly unasked rather than silently assumed. The one thing not to do is run two children with identical inputs and present their agreement as adversarial confirmation.

Worked round

draft claim: the generated registry is authoritative at runtime  [worker/boot.ts:22]

refuter        → REFUTED by default; opened worker/boot.ts:22, which imports
                 jobs.generated.ts directly → conceded, naming the artifact.
                 Recorded: conceded (not validated).
coverage       → INCOMPLETE: the deploy pipeline's own manifests were never read.
auditor          Probe: list the manifest files in the deploy configuration and grep
                 them for the registry filename.
round 2        → probe unavailable from this checkout → row stands as
                 "unresolved (needs a deploy-configuration read)", carried into
                 what this run could not see.