Monoculture Analysis

trigger

""Feature selection, calibration, effect sizes, composite design, presentation — all by the same model family.""

what it is

Every layer of inference produced by the same model family. Zero viewpoint diversity. Claude selected the features. Claude computed the effect sizes. Claude designed the composite. Claude presented the results. Claude wrote the caveats. Each layer's bias is invisible to the next because they share the same blind spots. The analysis looks rigorous — multiple layers, statistical tests, hypothesis confirmation — but the layers are not independent. They are correlated. The apparent depth is an illusion created by repetition, not by independent verification. The analogue: peer review where every reviewer went to the same school.

what it signals

A discerning reader asks: "Who checked this?" If the answer is "the same system that produced it," the checking is not checking. It is the same blind spot applied twice. Monoculture analysis produces confident, internally consistent results that are systematically wrong in ways no single participant can detect — because detection requires a viewpoint the monoculture does not contain.

instead

Run the analysis with a different model family. Or have a human statistician review the feature selection. Or validate against a published stylometry dataset with ground-truth labels. The honest version declares the monoculture: "All analysis by Claude. No independent verification. Treat accordingly."

refs

• AnotherPair calibration v3 session 2026-03-01
• Captain: 'How do I control for slop inside the analysis?'
• Layer 7b in the bias stack
• Swiss Cheese Model: multiple gates with the same hole