The central scientific contribution

Validation design

The key question is whether AI-assisted coding can validly measure human judgment from decision traces. The answer may be yes, no, or only under human audit.

StagePurpose
Trace collectionCapture initial decision, AI exposure, acceptance/rejection, verification, final decision, and confidence change.
Human reference codingTrain coders, establish construct-level agreement, and create a gold-standard benchmark.
AI-assisted codingRun fixed prompts and documented model settings on the same traces.
Agreement analysisCompare AI-human agreement by construct, not as one global score.
Failure-mode auditIdentify fluency bias, concision penalty, over-caution bias, verification hallucination, and model drift.

Pre-committed success criteria

Thresholds are fixed before data collection, per construct, so the design is falsifiable rather than a slogan. Agreement is judged against an adjudicated human reference on a locked validation set that is never used to tune the AI prompt.

ClassificationCondition
AI-codeablehuman–human κ ≥ 0.60 AND AI–human within-one ≥ 0.80 AND weighted κ ≥ 0.60
Human-audit-required (hybrid)meets human–human bar, but within-one 0.60–0.79 or weighted κ 0.40–0.59
Not AI-codeableweighted κ < 0.40 or within-one < 0.60
Not reliably codeable by anyonehuman–human κ < 0.60 — a boundary finding about the construct, not a pipeline failure

We also pre-register a separability test for whether AI overreliance and cognitive outsourcing are empirically distinct or collapse into one factor. On synthetic pipeline-testing data they already correlate ≈ 0.84, so the question is registered rather than assumed.


A useful outcome can be a validated instrument, a hybrid AI-human coding protocol, or a boundary-setting map of where AI measurement fails. Because the thresholds are committed in advance, "we learn either way" is a stated decision rule with defined ways to win and to lose.