← Mimir docs

Module 04 · proving it’s real

Verify

Claims about identity checked against sources of truth — not taken from the form.

The estimate says the car is a 2022 hatchback owned by Priya. The photos show a plate. The RC shows a chassis number. Verify’s job is to stop trusting the paperwork and start checking it: plate against Vahan, owner against the RC, PAN against the tax registry. Most fraud doesn’t survive a cross-check — if the cross-check actually happens.

What we're going to build

The connector interface

Sources of truth plug in as connectors — Vahan, DL, PAN, GST for motor; different registries for other domains. Async, cached, rate-limited, and fail-soft: a registry being down degrades the report honestly instead of blocking the case.

Cross-source matching

Entities extracted from documents and images are matched across sources, and every match/mismatch is a signal carrying the exact figures compared — "plate in photo: MH12AB1234; RC: MH12AB1234; Vahan: match."

Caching & staleness policy

Registry answers are cached with explicit freshness windows per connector — fast and cheap on repeat lookups, and the report always says how fresh the answer is.

Existing rails as the unfair advantage

Where the business already operates government-record APIs, they slot in as day-one connectors — already built, already sold. The platform makes them composable.

The contract (platform — never changes)

In: entities (from Extract/Vision). Out: verification signals — match/mismatch with the figures behind each. Sync for single lookups, async inside cases. Sold per verification or lookup.

What the pack supplies (per domain)

Which connectors count as truth · cross-check rules · freshness windows.

For engineers & architectsimplementation notes

Open questions — for the room

  1. Connector SLAs: what do we promise when Vahan is slow or down — degrade, retry, or hold?
  2. What legally counts as a "match" — exact string, fuzzy, or registry-confirmed?
  3. Which connectors are Phase 1 (RC/Vahan certain) vs Phase 2 (bank, bureau)?
  4. Pricing: per lookup pass-through plus margin, or bundled into the case price?
← PreviousVision

Mimir · the Modules intelligence engine — module 04 of 07 · all docs