Multi-persona AI review models the receiver, not the commission

Several AI agents took on different stakeholder personas and reviewed a paper I was about to send. They told me to strip a list of named external vendors from one paragraph because the names looked like a turf risk for one of the readers. I did. Then a second-level check surfaced a question I should have run first: did the sponsor commission those names? Back in my kickoff notes from two weeks earlier, the sponsor — the person who had actually commissioned the paper — had named those vendors specifically as items two through four on the original Asks. I had not just strayed from the brief. I had executed a persona recommendation directly against the sponsor’s instruction.

The strip was reverted within the hour. The lesson was sharper than the reversal. The persona lens had identified something real — the names would register as turf-adjacent for one of the readers, and that prediction was accurate. What the lens could not know was that the sponsor had already priced that risk in and wanted the names anyway. The persona models the receiver. It does not model the commissioning history.

This is the same shape as a problem I have hit with multi-LLM councils. Load three frontier models with a paper and an evaluator prompt, and the council will overrule constitutional decisions you have already made — naming choices, structural choices, deliberate compressions — because the council cannot see why you made them. The protection turned out to be passing the rationale file alongside the paper. Once the council could see why a choice was made, it stopped trying to undo the choice. Persona-based review has the same gap at a different surface, and the fix is the same: pass the commissioning record alongside the persona profile. The receiver model needs to know what the commissioning model already decided.

The discipline that took longer to learn is the one underneath the fix. Lens output is diagnostic, not authoritative. The persona correctly flagged a risk shape, but “this is a real risk” does not mean “change the body to remove it.” The risk might have been priced in by the sponsor. The risk might warrant a verbal pre-brief rather than a body edit. The risk might be a feature, not a bug. Persona review tells you what shape the risk has. It cannot tell you whether to absorb it, route around it, or proceed through it. That call depends on commissioning history the lens cannot see, so when the persona recommendation contradicts the sponsor instruction, hold against the lens, not against the sponsor. Flip-flop between the two is the symptom of having anchored on the lens output rather than the actual commission.

Multi-persona review is genuinely useful. It catches risk shapes a single reviewer would miss. But the value is diagnostic, not directive. The lens identifies the receiver’s likely reaction. The commission tells you what to do about it. Both halves matter. Either alone produces confident wrong answers.