Elliot Murphy

elliot-murphy.bsky.social

15 September 2026 19:12

Importantly, the best bracketing decoder can be the wrong mechanism. Ordered role-filler laws decode well partly because they retain daughter order that Merge should discard. More decodable information is not automatically more mechanistic truth (similar to 'surprisal' debacle with LLM-brain work).

Elliot Murphy

elliot-murphy.bsky.social

15 September 2026 19:11

For syntax, the target is unusually restrictive: preserve non-associative grouping; factor daughter order out of the composed object; remain recursively closed; keep substructures addressable; enforce type/role compatibility; and support headedness.

Elliot Murphy

elliot-murphy.bsky.social

15 September 2026 19:11

NAP's basic rule is simple: a neural mechanism is admissible only if it preserves the defining invariants of the computation it claims to implement. Neural data should then adjudicate among admissible mechanisms - not merely locate correlates of linguistic variables.

Elliot Murphy

elliot-murphy.bsky.social

15 September 2026 19:11

Much contemporary work in language neuroscience asks which LLM features, embeddings or surprisal measures best predict brain activity. NAP starts elsewhere: formal invariants first, neural mechanisms second (see: arxiv.org/abs/2609.03160).

Elliot Murphy

elliot-murphy.bsky.social

15 September 2026 19:11

This paper develops the Neural Admissibility Program (NAP): begin with what linguistic computation must preserve, then ask which neural dynamics could actually implement it. Many other domains of neuroscience have long adopted this logic, yet language has not been embraced in the same way.