A worked example: running a medical writing QC pass through a Council.

Picking up from yesterday's stack post. Theory is one thing, let's walk a real use case. Medical writing QC, specifically a late-stage draft of a clinical study report section before it goes to a medical reviewer.

A worked example. Running a medical writing QC pass through a Council.

Picking up from yesterday's stack post. Theory is one thing, let's walk a real use case. Medical writing QC, specifically a late-stage draft of a clinical study report section before it goes to a medical reviewer.

The input. A 12-page efficacy results section. Study already unblinded, statistical outputs locked, first-draft narrative written by a medical writer. The brief to the Council: flag inconsistencies between narrative and tables, check claim calibration against the data, surface anything a reviewer would want challenged.

Stage 1, individual responses. Four models run in parallel against the same source pack. Claude flags two instances where the narrative overstates a subgroup finding relative to the confidence interval. GPT catches a numeric mismatch between an in-text percentage and the corresponding table. Gemini picks up tonal drift in the discussion of adverse events, language softer than the data warrants. The specialist model queries whether a secondary endpoint claim is consistent with the SAP.

Stage 2, anonymised peer review. Each model receives the other three responses stripped of identity and ranks them on rigour, specificity, and defensibility. The numeric mismatch ranks highest across all four reviewers, everyone agrees it's the hardest finding. The tonal drift ranks lowest, flagged by only one model and contested by two others. This is the signal we want, convergence on high-confidence issues, divergence on judgement calls.

Stage 3, Chairman synthesis. The Chairman produces a consolidated QC memo. Three tiers: must-fix before reviewer handoff, recommended for discussion, noted for reviewer awareness. Every item traces back to the originating model and the ranking it received. The medical writer sees a single document, the audit trail shows the full deliberation.

What the human reviewer gets. A prioritised QC pack instead of four overlapping AI reports. A clear view of where models agreed and where they didn't, which is itself diagnostic. More time on the judgement calls, less on the mechanical checks.

What the pattern does not do. Replace the medical reviewer. Catch everything a domain expert would catch. Remove the need for QC SOPs. It compresses the mechanical layer of review so human expertise lands where it's most valuable.

The economics. Four model calls per pass, roughly 3 to 5x the cost of a single-model QC. Against medical writer and reviewer time, the ratio works, usually by an order of magnitude. Against the cost of a mismatched number reaching a regulator, it's not a conversation.

Single-model AI gives you a draft reviewer. A governed Council gives you a QC function.