note-sta...

## Stage 5 / AI Decomposability — Open Research Angle **The claim under scrutiny:** Dreyfus Stage 5 (Expert) involves tacit pattern recognition with "no rule decomposition" — the expert acts from holistic recognition, cannot articulate the underlying rules, and the performance cannot be decomposed into explicit criteria. **Two-component analysis of the claim:** *Component 1 — Phenomenological / tacit (Polanyi):* Expert practitioners genuinely cannot articulate the rules governing their performance. Introspective access to expert cognition is structurally limited. This component is grounded in phenomenology (Heidegger, Polanyi's "tacit knowledge") and is not obviously falsified by AI advances — the human expert still cannot tell you what they know. *Component 2 — Technical / external decomposability:* In 1980, there was no method to externally reconstruct the pattern structure underlying expert performance without going through the expert's own articulation. This component was a contingent technological limitation, not a logical necessity. **Andreas's premise:** Pattern recognition is a precondition for rule decomposition. AI has massively advanced external pattern recognition capability. Therefore the technical constraint on Component 2 is no longer what it was in 1980 — the axiom can be attacked on that dimension. **What AI decomposition produces:** Not traditional rule sets (IF condition → action). Instead: probabilistic behavioral signatures derived from performance data (FDM, simulator recordings, etc.) that approximate expert behavior without requiring expert introspection. These signatures can be learned from flight data at scale, aggregating across many expert performers. **Critical open question:** Does AI-derived decomposition need to be human-readable to be pedagogically useful? Two distinct use cases with different requirements: - Automated competency assessment: human-readability not required — the model evaluates and scores - Instructor-mediated feedback: decomposition must be interpretable by a human instructor to be actionable in a learning conversation **Connection to existing framework:** - Directly challenges the Stage 5 blindspot of the OB instrument (note-obs-proxy-struktur-grenzen) - If behavioral signatures can be extracted at Stage 5, it potentially extends OB-equivalent proxy coverage — partially addressing the latent variable inference problem (note-systemdesign-zwei-fundamentale-probleme) - Does not resolve Component 1 (tacit phenomenology) — expert cognition remains partly inaccessible to the expert themselves **Status:** Open research angle. Not yet established whether AI-derived behavioral signatures constitute a genuine form of "rule decomposition" in the Dreyfus sense, or whether they are a different epistemic category altogether (external statistical approximation vs. internalized rule structure).

Permalink: http://localhost:3000/sensemaking?noteId=note-stage5-ai-decomposability-challenge