Sprint 4: Vision Contract POC - scan types, watchlists, overlays, appraisal routing

At a glance

Goals

Sprint 4: Vision Contract POC - scan types, watchlists, overlays, appraisal routing — advances E1, E2, E3, E4, E5, E6, E7, E8, E9 toward its gate(s).

Implementation — what to build

Checklist

• ○ E1 table_hunt and item_scan schemas are accepted and rendered as overlays.
• ○ E2 Provider adapter scaffold is in place (consumes the P2 boundary).
• ○ E3 Overlay/review UX validates confidence and captures corrections.
• ○ E4 room_scan: multi-image coverage, category counts, approximate dedup, and multiple sightings per candidate.
• ○ E5 mark_scan and condition_scan: OCR + normalized maker/model/serial, visible-condition grading, and follow-up-photo guidance.
• ○ E6 appraisal_prepare: readiness state and search fingerprint without estimating value.
• ○ E7 appraisal_value: comps-based ranges, adjustments, and strongest-comp selection from supplied comparables.
• ○ E8 Watchlist contract: versioned, tenant-scoped inputs whose match IDs are restricted to those supplied.
• ○ E9 Local VLM adapter with grammar-constrained output and provider routing/benchmarking.

Gates & testing

• ○ E1 table_hunt and item_scan schemas are accepted and rendered as overlays.
  CODE TEST
• ○ E2 Provider adapter scaffold is in place (consumes the P2 boundary).
  CODE TEST 💲 API COST
• ○ E3 Overlay/review UX validates confidence and captures corrections.
  CODE TEST
• ○ E4 room_scan: multi-image coverage, category counts, approximate dedup, and multiple sightings per candidate.
  CODE TEST
• ○ E5 mark_scan and condition_scan: OCR + normalized maker/model/serial, visible-condition grading, and follow-up-photo guidance.
  CODE TEST
• ○ E6 appraisal_prepare: readiness state and search fingerprint without estimating value.
  CODE TEST
• ○ E7 appraisal_value: comps-based ranges, adjustments, and strongest-comp selection from supplied comparables.
  CODE TEST HITL QA 💲 API COST
• ○ E8 Watchlist contract: versioned, tenant-scoped inputs whose match IDs are restricted to those supplied.
  CODE TEST
• ○ E9 Local VLM adapter with grammar-constrained output and provider routing/benchmarking.
  CODE TEST

Requirements

• REQ-VISION-001 accepted Table hunt and item scan schemas need accepted contracts rendered as overlays.
  Done when: table_hunt + item_scan schema examples validate; candidates render as tappable overlays.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (11, 13)
• REQ-VISION-002 accepted Vision providers need an adapter boundary (consuming the P2 boundary).
  Done when: Provider-specific code is behind an adapter scaffold returning the canonical result.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (18)
• REQ-VISION-003 accepted Vision confidence and corrections need a visible review UX.
  Done when: Overlay/review demo shows confidence and captures bounding-box and identity corrections.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (21)
• REQ-VISION-004 accepted room_scan must handle multi-image coverage, counts, dedup, and sightings.
  Done when: Multi-image coverage, category counts, approximate dedup, multiple sightings per candidate, and a missing-coverage checklist validate.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (12)
• REQ-VISION-005 accepted mark_scan/condition_scan must extract marks and grade visible condition.
  Done when: Literal OCR preserved + normalized maker/model/serial; visible-condition grading links each observation to an image; functionality is never inferred.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (14-15)
• REQ-VISION-006 accepted appraisal_prepare must report readiness and a search fingerprint without pricing.
  Done when: Returns readiness state + missing evidence + a normalized search fingerprint and never estimates value.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (16)
• REQ-VISION-007 accepted appraisal_value must produce comps-based ranges and adjustments from supplied comparables.
  Done when: Returns multiple sale-context ranges, adjustments, and strongest comps; TroveSnap computes the quantitative baseline; unsupported model memory is never treated as a comparable.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (17)
• REQ-VISION-008 accepted Watchlists must be versioned, tenant-scoped, and restricted to supplied IDs (POC enforcement).
  Done when: Result watch_id values are restricted to those supplied; watchlists are tenant-scoped and versioned.
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (8)
• REQ-VISION-009 accepted A local VLM adapter must support grammar-constrained output and routing.
  Done when: Local VLM adapter produces grammar-constrained output and participates in provider routing/benchmarking (local-vs-hosted quality delta tracked).
  Spec: docs/CC-Analysis/trovesnap-vision-scanning-spec.md (18, 27)

Verify (done when)

scan types + watchlists + overlays + appraisal routing demo on the frozen contracts.

Cost triggers

First hosted vision-provider calls turn on API spend; keep economy tier + per-scan caps (spec sections 25). Strong provider + up to 3 external comp queries per appraisal — highest per-call API cost; gate behind tokens.

Status & editing

queued   Sprint: Sprint 4 - Vision Contract POC

Edit status / sprint on the ★ Live Board → — changes are logged live with who / when / why.

Human-in-the-loop

No human tasks linked.

Findings

None recorded yet.

Concerns

None recorded yet.

Risks

None recorded yet.

Tech debt

None recorded yet.

Tools used

No tool calls recorded.

Files & artifacts

No files / artifacts recorded.

Update log

No log entries yet.