01 · The verification backbone
- Verification protocol — how every Missouri DWI proposition is checked against primary sources before it ships, and the errors that process caught.
- mo-dwi-authority.md — the shared, verification-gated authority file: statutes, 19 CSR regulations, and case law, pin-cited and dated, with unresolved questions marked
⛔ OPEN rather than asserted.
- Changelog — release and correction history. Corrections to legal authority, deadlines, or classifications are recorded there and in the commit history.
02 · Definitions
Verified authority
A proposition checked against the primary source (Revisor / SOS regulations / CourtListener) with a pin cite and a verification date recorded in the authority file.
Open question
A legal question we could not verify. Marked ⛔ OPEN; the skills are instructed not to assert it.
Synthetic example
An example built from invented facts, labeled as synthetic. No real client or arrest data appears anywhere on this site.
Unverified proposition
Anything not yet through the protocol. It does not ship in a skill as an assertion.
Issue-spotting output
A list of questions and categories for the attorney to investigate against the record.
Legal conclusion
A determination about a specific case — admissibility, suppression, classification, or outcome. The skills never make one; the attorney does.
03 · Correction history
The v0.2.0 verification pass corrected substantive defects found by adversarial review — including a mis-seeded hearing-request citation (corrected to § 302.530) — and published both the fixes and the errors made finding them at /verification-protocol. Future corrections follow the same pattern: fixed in the files, recorded in the changelog, announced in the RSS feed.
04 · Reporting an authority error
Found a wrong citation, deadline, or classification? Use the authority-correction issue template (or the deadline/classification template). Synthetic or redacted facts only — never post real client or arrest information. If the error could affect a live case, report it privately per the security policy.
05 · Research roadmap
Planned benchmark topics — planned, not completed; no results are claimed and nothing here has been peer reviewed:
- Generic prompting vs. controlled skills on the same DWI fact patterns.
- Deadline-recognition accuracy (notice date → hearing-request deadline).
- Criminal vs. administrative-track separation in output.
- Refusal vs. test-result classification accuracy.
- Offender-level classification against the § 577.001 definitions.
- Foundation-issue recall against an attorney-built answer key.
- Hallucinated-authority rate with and without the verification-gated authority file.
Scope. All benchmarks will use synthetic case files. Issue-spotting output is measured; legal conclusions are not automated and not scored as if they were.