Prototype dossier
Regulatory Gauntlet
A bounded experiment in using multiple review perspectives to interrogate synthetic pharmaceutical marketing content—designed to support traceability and qualified human review, never to issue approval.
Target user
A reviewer or cross-functional team examining regulated promotional content and needing a clearer record of questions, evidence and unresolved issues.
Problem
Complex review can distribute relevant context across specialties. The prototype asks whether distinct machine-generated perspectives can help surface review questions consistently without collapsing judgement into one opaque score.
Workflow
- 01Supply a synthetic claim and its synthetic support
No real customer, patient, employee, clinical or confidential material belongs in the public evaluation.
- 02Run distinct review lenses
Each perspective examines a bounded question—for example evidence alignment, ambiguity, or missing qualification.
- 03Preserve disagreement
Outputs are assembled as traceable observations rather than averaged into an approval decision.
- 04Escalate to qualified review
A human reviewer decides what is relevant, what requires evidence, and whether the material proceeds.
Architecture boundary
The public site is a static dossier. It has no authentication, database, model endpoint, upload surface or administrative route. The retired client artifact and private services are not exposed or restored.
Evaluation method
- Define a synthetic scenario and reference answer before a run.
- Score issue recall, unsupported flags and traceability to supplied evidence.
- Record false positives, false negatives and disagreement between lenses.
- Test whether a reviewer can understand why an issue was raised.
- Treat a confident but unsupported output as a failure, not a success.
Limitations
- No public benchmark result is claimed.
- Synthetic cases cannot establish performance on real review work.
- Model output can be incomplete, inconsistent or wrong.
- Prompted perspectives do not equal independent expert judgement.
- The prototype does not validate legal or regulatory compliance.
Next validation step
Create a small, versioned set of synthetic cases with known issue classes; define scoring before execution; and ask qualified reviewers whether the issue record improves review quality, traceability or speed without encouraging automation bias. Publish only the evaluation design and results that can be safely supported.