# Method

<!-- https://antecedint.com/method -->

Method

# Where the right answers come from

A search over law returns documents. It does not say whether they were the right ones. Antecedint is built to say so: for each kind of question lawyers ask, we find where the public record already decided the answer, turn those decisions into labels, train retrieval against them, and publish the benchmark so anyone can rerun it.

The record supplies the labels. The people who search it supply the corrections: when you or your assistant find a wrong or missing record, the MCP Feedback program puts an accepted fix into the next build.

## Four strata

1. 01RawThe record as the government publishes it.
2. 02ResolvedEvery reference pointing at the exact text it was applied to.
3. 03JudgedThe questions the record already answered, kept as labelled pairs.
4. 04TunedRetrieval trained on those pairs, with the eval attached so it is re-measured when anything changes.

## Judged sources the record contains

Each corpus has a place where a person with authority already decided which document answered the question. That decision is the label. The method is the same in every row; what changes is who decided and where the record keeps it.

| Corpus | Who decided | Status |
|---|---|---|
| PatentsLive, benchmarked | Examiners, in art rejections | Live, benchmarked |
| PTABNext | The Board, in institution and final written decisions | Next |
| Case law, cite-checkFoundation ingest underway | Courts, in citations, pin cites and quotations; sanction orders on bad cites | Foundation ingest underway |
| Case law, authorityPlanned | Judges quoting precedent | Planned |
| Codified lawFoundation ingest underway | Point-in-time versions from OLRC and eCFR | Foundation ingest underway |

## What the first corpus shows

Patents were first because the judged source is dense and public: every art rejection names the reference an examiner applied to a claim. Retrieval is trained on 609,723 of those citations and measured on a 97,913-query public holdout, family-disjoint from the training set. The v1 result on that holdout is 45.3% recall at 100 and 74.2% at 1,000, bake-off fingerprint `6fd3de94f3b9`. The harness is published, so the number is one anyone can produce again.