LawVerse

The Semantic Movement Index, SMI, is a simple estimator. Embed every mention of a legal phrase such as "due process" or "bail", average the vectors within a decade, and take one minus the cosine similarity between two decades. We ran it over 231,908 judgments, 99.7% of them Indian, with a legal-domain encoder.

LawVerse is the platform built around it: search over Indian judgments by meaning, a citation graph that shows how later courts treated a case, doctrine timelines for 25 legal concepts, and an AI review of contracts and property documents. Its Android app is called GLaw.

The first build produced a striking result, complete with confidence intervals, and it was wrong. Four defects were responsible, and none of them raised an error or a warning. The text window was cut from the start of each judgment, so it usually did not contain the concept at all. The matcher matched loose words instead of the phrase. A fallback gave every unmatched decade the same vector. And the bootstrap resampled a single row, so every interval had zero width by construction. With no annotated benchmark for legal concept drift, nothing outside the pipeline could tell us it was broken.

A corrected pipeline: the window centered on the mention, contiguous phrase matching, no fallback, and one vector kept per document. Build-time checks that fail when two periods share a vector or an interval has zero width. A controlled ablation that changes one design choice at a time on identical data. A cheap probe that asks whether a representation can tell which concept it came from.

Around it, the platform. A cleaned corpus and a citation database of 115,537 cases joined by 67,864 citation links, with a label for how later courts treated each case. Semantic search. A document review in which a language model reads contracts and Indian property documents in 12 languages, quotes the passage behind each finding and redacts Aadhaar numbers. A FastAPI gateway, a Next.js website, an admin dashboard, and a Flutter Android app with a document scanner.

Rebuilt from the old code, the original pipeline labels 22 of 25 concepts as changing. The corrected one labels 1. The defects interact: the fallback alone corrupts 3.0% of cells, but 19.6% once the bad window is added. One defect turned out to be plainly wrong and yet harmless on everything we could measure, and we report it that way. The largest effect was not a bug at all. Changing only which token positions are pooled swings the verdict from 1 to 24 of 25, and the most dramatic drift comes from the representation that is worst at identifying a concept (23.7% against 99.7%). Measured correctly, 98.8% of the 169 reliable decade-to-decade transitions sit in the lowest band.

The document review has its own evaluation harness with a release gate. On its best run of 60 cases it named the document type correctly 98.4% of the time and caught 95.8% of serious defects. Those cases are synthetic, and 95.8% is just below the gate of 98%, so the review has not yet passed its own bar.

The study is a negative result with a fix. Its paper, "When the Metric Measures the Bug", was submitted to NLLP @ EMNLP 2026, not accepted, and is under revision. It makes no cross-country claim, because only 800 documents are from the US and none are European. It reports no retrieval benchmark, because no annotated query set exists. The platform and the app run on local machines; nothing is deployed publicly. Earlier descriptions of this project, including its public README, carry figures from before the correction; this page is the current account. The next research steps are the ones the paper names: repeat the measurement with other encoders, and build an annotated benchmark for legal concept drift.