nx_cuad_lib.nx

nx_cuad_lib.nx -- THE CUAD CLAUSE TAXONOMY (41 types) as a typed, benchmark-aligned playbook. WHY (July 2026 SOTA, researched): CUAD -- 13,000+ expert labels over 510 commercial contracts from EDGAR, covering 41 clause types -- is the benchmark essentially every contract-AI product trains or evaluates against. Our clause layer had a TWO-clause playbook, so "retrieval works" was true and meaningless. This enumerates the real target so coverage becomes a number with a denominator. u2605u2605THE HONESTY MECHANISM, BUILT INTO THE ORGAN: declaring 41 constants would let us claim 41 types of coverage while only a handful can actually be detected -- exactly the coverage-gaming this house has been burned by. So the taxonomy is split in two and the split is enforced in code: DECLARED -- all 41 types exist as named constants (the complete, greppable denominator) CUED -- the subset with real discriminative detection cues cuad_coverage_permille scores against CUED types ONLY. A type that is merely declared can NEVER contribute to a coverage number. The organ literally cannot flatter itself; adding a constant does nothing until you add cues to go with it. u2605THE CONTRADICTION TOOTH: "Uncapped Liability" and "Cap On Liability" are SEPARATE CUAD types, and a contract asserting BOTH is internally inconsistent -- one clause caps exposure while another blows the cap open. Reviewers miss it because the two live pages apart and each looks fine alone. Detected here as a first-class contradiction rather than as two independent findings. u2605FAIL-CLOSED: unrecognised text is CUAD_UNKNOWN, never the nearest type. In diligence, "I could not classify this" is actionable and "probably a notices clause" is how a change-of-control provision gets skipped. license_tier: ORIGINAL LIB.