Module 03 · reading the images

Vision

Photos and video become findings — in the vocabulary the pack defines.

Priya’s eight photos and walkaround video show a dented front bumper, a scratched fender, and — importantly — a completely intact rear. Vision turns pixels into findings: part, damage type, severity, region on the image. But here is the discipline that keeps Mimir generic: Vision does not know what a "bumper" is. The motor pack taught it.

What we're going to build

Pack-defined feature schemas

Vision returns features in the schema the pack supplies — parts/damage/severity for motor, condition grades for fleet, something else entirely for the next domain. "Damage" never appears in engine code. This is the R4 guard in practice.

Detection + severity models

Damage detection is commoditising — we integrate and benchmark models through the gateway rather than claiming detection accuracy as our pitch. Our work is the harness: consistent features, calibrated severity, evidence regions on the image.

Video as frames with intent

The walkaround video is sampled against the capture script — do we have all eight angles, does the plate appear, is the lighting usable? Video backs up photos; it does not replace structure.

Compare mode (before/after)

Two sessions in, a diff out: what changed between handover and return? This one operation turns a claims tool into a fleet, rental, warranty and logistics tool — the cohort multiplier.

Identity cues for Verify

Plate and chassis read from the images become signals that Verify cross-checks against the registry. Vision reads; Verify proves.

The contract (platform — never changes)

In: images/video (or two sessions for compare). Out: feature signals — regions, classes, scores — per the pack’s schema. Sold per inspection or comparison.

What the pack supplies (per domain)

Feature schema · vision model bindings · severity calibration · capture-script linkage.

For engineers & architectsimplementation notes

Vision emits feature signals whose value shape is the pack schema — engine code treats them as opaque typed payloads. A CI lint fails any engine source that names a domain feature (the R4 guard, mechanised).
GPU-heavy inference runs on its own queue with batch-friendly workers; compare mode is a pure function over two sessions’ stored features — no re-inference when both sides are already processed.
Video pipeline: sample frames keyed to capture-script steps, dedupe near-identical frames, run the same feature models — video adds coverage, not a second model stack.
Failure modes: partial angle coverage emits a coverage_gap signal (Decide can route on it); model-version drift is pinned per case so a re-run is reproducible.

Open questions — for the room

Build vs license the motor damage model for Phase 1 — and which do we benchmark against?
Part taxonomy: adopt an industry standard or define our own (and map at the edges)?
Severity calibration: who signs off that "moderate" means the same thing across insurers?
Compare mode: Phase 2 as planned, or does a fleet design partner pull it into Phase 1?

← PreviousExtract Next →Verify

Mimir · the Modules intelligence engine — module 03 of 07 · all docs