A LOCAL-FIRST POLICY EVIDENCE PROTOTYPE

Evidence before
answers.

A small, inspectable workflow for dated policy clauses, evidence display, abstention, and human review routing.

Explore the source See how it works

Respect the date

Filter policy versions to those active on the selected date.

Show the clause

Return a matching clause with policy version and clause identifiers.

Route uncertainty

Abstain on no match; flag ties and selected high-impact terms for human review.

UNDER THE HOOD

A deliberately small
reliability experiment.

The Python demo uses deterministic token overlap against a bundled JSON fixture. It filters by effective date, applies a simple score threshold, and exposes its routing decision in the result.

Python standard libraryLocal JSON fixtureNo model callsNo saved query history

Run the local demo

FROM THE LOCAL INTERFACE

Inspect the workflow.

Real captures of the prototype running on the developer's machine. The hosted page is a static presentation.

Local prototype showing a matched synthetic meal expense clause with version and evidence details — 01 / EVIDENCE MATCHLocal browser capture · synthetic rule

Local prototype showing an unsupported query routed to abstention and human review — 02 / ABSTENTION + REVIEWLocal browser capture · synthetic rule

SCOPE AND LIMITS

For exploration,
not policy decisions.

Travel and expense examples are invented synthetic fixtures.
Matching and risk values are simple heuristics, not calibrated confidence.
No real policy corpus, document ingestion, LLM, authentication, or audit trail.
The local server binds to loopback. Hosting it as a public backend would change its privacy boundary.