What citation integrity means for Verify
Our first research paper separates fake authorities, misquotes, and unsupported propositions. Verify runs on matter documents in the workspace; here is how the paper maps to it.
We published Citation integrity in evaluation as technical report TR-2026-01. The claim is simple: one headline accuracy number hides failure modes lawyers actually care about. Inventing an authority is not the same as citing the wrong case. Drifting a quote is not the same as asserting a proposition the passage does not support.
If evaluation collapses those errors into a single score, product teams optimise for the average and firms get a number that explains nothing when something goes wrong. The paper is methodology for separating those modes. It is not a leaderboard.
What Verify is aiming at
Verify is the product surface aimed at those checks — citations, quotations, and propositions against stored sources. It runs on matter documents in the signed-in workspace today.
The paper’s axes are how we evaluate systems and how Verify exposes checks in the product: citation resolution, authority fit, quote fidelity, proposition support, and retrieval. Each axis answers a different question. Mixing them into one “accuracy” figure is exactly what we refuse in the report.
Benchmarks stay honest
On Benchmark you will see suites for citation work, redaction, and real work tasks. Citation resolution is Internal at v0.3. Redaction is Internal and under active work. Real work tasks are Planned. Comparable scores and manifests publish only when a suite freezes — not before.
That publication rule is the same one in the research paper. Internal drafts are not public leaderboards. If we publish a score, it comes with a frozen suite, a version, and a way to replay. Quiet rescoring of a named public figure is not acceptable.
What you can use today
The shipping grounding surface is Search over UK judgments Obiter holds: in the workspace today, and publicly at /search once it opens. Stable stored case pages are what verification will need to check against. Without reproducible retrieval, citation integrity has nothing solid to stand on.
Read the full report on Research for the failure-mode taxonomy and protocol sketch. Check Benchmark for suite status. Use public Search when you need judgments. Verify sits on top of that same stored source — it does not invent a second corpus, and it does not blur matter files into the open index.