nx_clause_lib.nx

nx_clause_lib.nx -- CONTRACT TEXT REASONING: clause retrieval, redline deviation, playbook risk. u2605THIS IS THE QUALITATIVE GAP. Every other legal organ here computes over facts a caller supplies. This one READS THE DOCUMENT: it locates a clause in raw contract text, RETRIEVES the closest standard clause from a playbook corpus, measures how far the draft has drifted, and scores the risk. That is the deterministic core of what Ironclad's Jurist and DocuSign's IAM agents do during negotiation. u2605TWO DIRECTIONS OF DRIFT, MEASURED SEPARATELY -- because they are different problems and a single "similarity" number hides both: RETAINED : how much of the STANDARD clause survives in the draft. Low retention = protections were DELETED. This is what silently strips an indemnity or a limitation of liability. NOVEL : how much of the DRAFT is language the standard never had. High novelty = terms were INSERTED -- an added carve-out, an uncapped obligation, a new termination right. A redline can score 900 on one and 100 on the other. Averaging them into one figure is how a dangerous edit passes review looking "95% similar". u2605RETRIEVAL IS BY MEASUREMENT, NOT BY LABEL. cl_best_match scans the playbook and returns the clause with the highest retention against the draft. It never trusts a heading: a paragraph titled "Limitation of Liability" that actually reads like an indemnity will match the indemnity standard, which is exactly the mislabel a human reviewer skims past. u2605FAIL-CLOSED: an empty draft, an empty playbook, or a clause that matches NOTHING above the floor returns CL_NO_MATCH -- never "matches the first entry". An unmatched clause is escalated, not guessed. Thresholds are DATA (rule 11) as named constants in per mille, never literals in branches. STRUCTURE: pure text functions, no storage, no I/O. license_tier: ORIGINAL LIB.