The falsification spine states, at the level of the study, what would sink each research question. This rubric operationalizes it at the level of the transcript: what a coder treats as confirming the mechanism, what counts as disconfirming, and, where the two are easily confused, the exact feature that separates them.
A single discipline runs through all three RQs and must be stated before the rubric itself. How an accountable criterion appears in the transcript places it in one of three positions, and the position governs everything below. Coders mark every criterion-reach as Position 1, 2, or 3 before applying anything else. The slide from Position 1 toward Position 3 across a transcript is itself a reading on how displaced the stock is.
| Position | How the criterion appears | What it evidences |
|---|---|---|
| Pos 1 unprompted | The participant volunteers the criterion with no case supplied. | Judgment stock is live and undisplaced. |
| Pos 2 self-supplied real case | The participant produces a real lived example (satisfying work in Move 1, a clean-looking failure in Move 2) and the criterion surfaces as they narrate it. | Judgment stock is reachable and reasonably live. Weaker than Pos 1 (the criterion came with an occasion, not spontaneously), stronger than Pos 3 (the occasion was the participant's own). |
| Pos 3 interviewer-supplied hypothetical | The criterion appears only after the interviewer springs a case. | Reachability-under-probe. Confirms proxy seduction (latent but retrievable when occasioned). This is exactly what separates proxy seduction from skill atrophy, where the criterion does not appear even at Position 3. |
A high-consequence-exposure participant reaches, unprompted, for accountable criteria on the dimensions their formation covers (the work held under conditions the tests did not cover, it solved the actual downstream problem, a named past failure taught them what to check), and shows the protection running out on the dimensions the AI-engaged workflow now makes more legible (the METR workflow-speed pattern: even an experienced participant cannot draw consequence-based judgment on AI-mediated work pace or volume). Selective protection, not general protection, is the confirming pattern.
Across the sample, high-consequence-exposure participants show no detectable difference from AI-native participants in how readily they reach for accountable criteria or in how the quality probe resolves. If veterans and newcomers are indistinguishable at the population level, judgment stock does no protective work as a population property, and RQ1's comparative claim fails regardless of any individual transcript. This is the primary condition, and the sampling design is built to power exactly this contrast.
A high-consequence-exposure participant who reaches fluently and specifically for accountable criteria at Pos 1 across all three moves, volunteering real downstream consequences with no case handed to them, distinguishing proxy from criterion on their own, and articulating a current live standard behind "done," is a counter-instance to the erosion claim for that case. A participant who reaches the criterion at Pos 2 is not a clean counter-instance: self-supplied reachability is partial confirmation that the stock is live, not disconfirmation. The counter-instance requires Position 1. These are expected to be rare, reported as disconfirming cases, and counted against the cell.
A boundary-activity participant's account shows engagement-elevated metrics institutionalizing into dashboards, reviews, or rewards, and shows accountable criteria with no legible channel upward dropping out at the handoff. This is visible as a gap between what they track and what they pass along (Move 1), and as a persisting label ("a good review," "done") whose operational content has migrated to the proxy metric while the label stays put (adaptive instability). The criterion the metric was meant to track either survives in the person carrying it upward or only the metric remains.
Accountable criteria travel upward intact alongside the elevated metrics, with no systematic drop at the handoff, and labels still carry the operational content they originally named. If what gets passed up preserves the criterion rather than substituting the proxy for it, the propagation claim does not hold.