Retrieval-augmented compliance assessment

Compliance findings
you can check.

Verity reads a research or academic document, works out which regulatory frameworks it touches, retrieves the sections that actually apply, and reports what is missing — quoting the regulation behind every finding.

The step that matters is the last one. Every quote the model produces is checked against the passage it was shown before it reaches you. A finding whose citation cannot be found is labelled unsupported and excluded from the score, rather than printed in the same typeface as a real one.

Run an assessmentTry the retrieval playgroundRead a finished report

No account, nothing to install. Rate limited, because it spends real money on a real model — and the finished report is there whether or not you have a run to spare.

What is in the index

ADA/Section 508 · Export Controls (EAR/ITAR) · FERPA · GDPR · HIPAA · IRB. SOC 2 and ISO/IEC 27001 are classified but not retrieved against — their text is copyrighted and cannot be redistributed here, and Verity says so rather than paraphrasing them.

413
sections of regulation text
1,147
retrievable passages
6
frameworks with a corpus
768d
embedding dimensions
gemini-embedding-001

Retrieval quality, measured

Six configurations compared on the same labelled set, with the one tunable parameter fitted on the development slice alone. The full report includes the configurations that lost, and why.

90.6%
Recall@10, held out
32 queries · Hybrid + rerank
0.741
MRR@10, held out
0.776
nDCG@10, held out
118
labelled queries
76 dev · 32 test
Read the full evaluation →

How a run works

  1. 01

    Classify

    One call decides which of eight frameworks the document actually touches, and must say what in the document triggered each one — so a wrong answer is visible rather than merely plausible.

  2. 02

    Decompose

    The same call emits short concerns: the specific things in this document that might create an obligation. Those become the retrieval queries. Embedding a whole document as one vector retrieves nothing in particular.

  3. 03

    Retrieve

    Each concern is embedded and matched against the corpus, filtered to the framework in question. BM25 and rank fusion are implemented and measured too; the evaluation explains why they are not the default.

  4. 04

    Assess

    The model sees the document and the retrieved passages, and must quote both for every finding it makes.

  5. 05

    Verify

    Both quotes are checked against their sources — exact, near, or unsupported. Unsupported findings do not count toward the score, and are shown as unsupported rather than quietly dropped.