Auditing Brain-LLM Similarity Claims via connection_audit: A Methodology-Export Study

Bridge Paper B imports the Watabe-Claude Method's connection_audit framework(PL-21, extended by PL-24's REFUTED verdict) — built and tested entirely onWCM's own cross-domain claims (Bridge Papers C-Z) — and applies it unmodifiedto the brain-LLM similarity literature (neuroscience of language). Stage 1 audits six claims (B-1 through B-6) drawn from Schrimpf et al. 2021,Caucheteux & King 2022, and related work. None reach STRUCTURAL status: twoare REFUTED, three UNDETERMINED, one NUMERICAL. Stage 3 audits the field's own null-model methodology instead of claims aboutneural computation, and finds two STRUCTURAL (4/4) connections: fMRIstatistics (post-2016 Eklund et al. cluster-failure crisis) and WCM-PL(PL-16, 2026-07-04) independently converged on the identical fix(structure-preserving permutation nulls) for the identical disease (naiveshuffling inflating significance under autocorrelation). One item (C-3) isdeliberately left open rather than resolved from secondary sources. This is secondary analysis: no new brain or model data was collected. Files included:- bridge_paper_b_v1.pdf: Full paper (v1.0, Methods/Results/Discussion/Conclusions)- generate_bridge_paper_b_v1.py: PDF generation script (reportlab)- bridge_paper_b_design.md: design document (Stage 1-3 plan, pre-registered predictions)- audit_brain_llm_claims.py: Stage 1 connection_audit code (B-1 through B-6)- audit_stage3_claims.py: Stage 3 connection_audit code (C-1, C-2) GitHub: https://github.com/watabe-masanori/cds-polytope (AGPL-3.0)