Research proposes survival certificate to validate machine-extracted legal logic under parser noise
A new method certifies which logical implications survive across noisy statutory parsers, with preregistered tests on Missouri and Indian statutes and a public audit trail.
1 source · cross-referenced
- Two independently written statutory parsers disagreed on numeric-threshold presence in Missouri statutes at a false-negative rate of 0.43, motivating a formal approach to assess which extracted legal logic can be trusted.
The paper introduces a passive survival certificate for the Duquenne-Guigues implication basis derived from machine-extracted statutory contexts. The method measures per-attribute disagreement between two independent statutory parsers, then replays that disagreement against the extracted implication basis across 1,000 Monte Carlo trials. An implication is certified only when a one-sided Wilson 95% lower bound on survival probability reaches 0.95, and each certified implication includes premise spans and a minimal counterexample.
On held-out evaluations spanning 29,365 Missouri statute sections and 502 Indian central-act sections, the preregistered gate passed for 10 statute families across 7 Titles exactly, and 16 families across 11 Titles with a 5% tolerance. However, under a globally deployed error model, 93.2% of held-out chapters fell below an informativeness floor, with a factorial analysis attributing the failure to calibration-rate transfer rather than selection.
The authors release code, data products, and an audit trail that includes one retracted claim, alongside preregistration details. They caution that the certificate is usable but fragile, recommending deployment only when per-chapter calibration or error-tolerant mechanisms are in place.
- Nov 9, 2026 · arXiv cs.CL
Study finds real-world AI identification practices rely on context-dependent norms rather than fixed textual features
Trust79 - Nov 9, 2026 · arXiv cs.CL
FinEvolveKG introduces block-constrained prompt optimization for financial knowledge graph construction
Trust79 - Nov 9, 2026 · arXiv cs.CL
Study finds large language models overcompress fine-grained emotion distinctions in internal representations
Trust84