A single model auditing its own output is theater. The blind spots that produced the original output produce the audit too. Every Trial-tier SpotlightAICore deliverable passes through a different model in a different family before release, looking for what the producing model missed. This is what that pipeline does and what the resulting review memo contains.
The pairing rule
Pairings are mechanical and asymmetric. A deliverable produced by Claude Haiku is reviewed by Claude Sonnet. A Sonnet-produced deliverable is reviewed by Claude Opus. An Opus-produced deliverable is reviewed by Sonnet. The producer and the reviewer are never the same model and, where possible, sit at different capability tiers. Same-tier same-family review is excluded by construction.
This is not the only sensible pairing — cross-vendor review (Claude reviewing GPT, or vice versa) is more adversarial still — but cross-model within Anthropic gets most of the benefit at a fraction of the operational complexity, and keeps the data-handling perimeter inside a single vendor relationship the customer has already vetted.
What the reviewer looks for
The reviewer is briefed to read the finished deliverable as a skeptical opposing counsel would, looking specifically for:
Unsupported claims. Assertions in the deliverable that exceed what the cited source establishes, or that have no citation at all.
Citations that don't carry the weight stated. The cited source exists; reading the cited passage carefully reveals the source supports a narrower, qualified, or differently framed proposition than the deliverable's paraphrase asserts.
Overconfidence. Banned-vocabulary slips ("clearly," "obviously," "devastating," and the rest of the house list) and other language that overstates strength of position.
Omitted contrary evidence. Evidence in the document set that meaningfully cuts against the deliverable's framing but is not addressed.
Logical gaps. Premises that don't actually support the conclusion drawn, or conclusions that require an unstated intermediate step.
Factual errors. Names, dates, dollar figures, statute citations wrong on the face of the source material.
Ambiguous framing. Passages that could be read multiple ways and would benefit from precision.
The verdict and its consequences
The reviewer returns one of three verdicts. Pass: zero findings or only one or two low-severity ones. Pass with revisions: low or medium-severity findings only; producer model revises before release. Hold for founder review: any high-severity finding, or three or more medium-severity findings.
The verdict is not advisory. The Hold verdict gates the deliverable: M. David Hoyle reviews the work product against the reviewer's findings before the deliverable releases. A pipeline escalator catches the case where the reviewer under-classifies its own findings — if a high-severity finding appears but the reviewer stamped pass-with-revisions, the wrapper auto-escalates to Hold. The customer cannot receive a Trial deliverable that auto-passed despite a high-severity finding.
The review memo is appended to the deliverable
The customer does not get the work product without the audit. Each Trial deliverable's docx includes an Adversarial Review Memo as a final section: producer model, reviewer model, verdict, summary, finding count by severity, and each finding with its location, the finding text, and the suggested revision. The audit is delivered alongside the work, not in a separate compliance file the customer never sees.
Why this is what Trial tier pays for
Strategic-tier deliverables go into discovery and motion practice. Trial-tier deliverables go into briefs, jury arguments, and trial binders. The cost of an error at Trial is whatever the verdict would have been minus what it actually is. Two-AI Adversarial Review is what separates the two tiers operationally; it is also what makes the insurance-backed accuracy guarantee actuarially defensible at Trial-tier pricing.