Know your roof's number before any salesman calls.

Measured agreement with commercial provider reports across 553 properties. Uncalibrated engine (the non-circular measurement): 8.19% median absolute area error, 85.2% within 20%. Shipping engine (what a report prints): 7.83% and 85.7%. Typical pitch error 0.3 rise-per-12 on both. Re-runnable: uv run python -m app.calibration Remaining gate: Stays N/A. The rubric requires signed on-roof physical audits; provider reports are explicitly not promotable to field truth. 0 roofs field-measured. Against the bars (area WAPE <3%, pitch within 0.5/12 on >=90%), today's engine is 2.7x over on area and at 63.6% on pitch — collecting the cohort now would score RUNG 0, so close the gap before spending a one-shot sealed cohort.

Structure gateway with address-point containment; the orchestrator now READS the anchor and re-anchors the LiDAR crop when a candidate contains the address point — closing the program's signature 'computed but never read' defect class. Reachable via POST /v1/measure. Remaining gate: Signed v3 place/parcel/footprint contract still absent. Re-anchoring is proven on 1 property and depends on a live gateway with no fallback. Next artifact: persist gateway payloads across N>=20 addresses including known-wrong-geocode negative controls.

First real LiDAR-derived 3D result in program history: 1,823-triangle mesh from 1,028 vertices, glTF 2.0 GLB independently re-parsed (chunks aligned, indices in range, zero degenerate triangles, recomputed area matches). Now wired into the product path, so the report renders real faces rather than a bare point cloud. Remaining gate: B4 now PASSES 8/8 on 6740 (collection date 2019-04-17 resolved to the day from GPS time in the point records; release_gate_pass True, orderable). The stated blocker is void. Remaining: n=1 property, and the raise below is NOT yet adversarially audited.

sha256 pre-read tile verification with a typed failure vocabulary; the manifest size-drift trap documented (would have false-rejected the whole 409G vault); every reading carries a derived-sigma method string; the circular accuracy feedback loop is closed — the engine now reads a reviewed frozen snapshot, so re-running the harness can no longer silently change published claims. Remaining gate: No offsite backup and no restore drill. Cloud checksum verification unproven.

Always five Letter pages with a band on every metric and a new Evidence & Confidence page; rendered 3D from real mesh faces. Remaining gate: The measured-3D result now passes B4, so a passing fixture is reachable. Facet error is 50% median (the 75% figure was partly a corpus ceiling: ga_all_3d caps at 5 facets against an EagleView mean of 13.1). Tier-1 lineals now FIRE, on 29% of clouds at 37-88% WAPE — far from the 5% bar. Raise NOT yet adversarially audited.

Streaming chat with route receipts, tool timeline with inline approvals, conversation persistence/search/export, prompt library, usage card, three-state readiness ladder — all live, reachable routes. Remaining gate: Moderated homeowner/estimator studies and a full assistive-technology audit.

No change. There is no timing instrumentation in the measurement path. Remaining gate: Rejected on audit: a single ~1,000-point crop is not product speed, and the figure sits inside a range v23 already counted. Needs real instrumentation, a production SLO, and load-test history.

Usage accounting that cannot break a completion, typed failure vocabularies, in-band stream errors with guaranteed termination, Scout ledger. 279 automated tests green. Remaining gate: Own-tests-marking-own-homework is not product reliability. Production dashboards and sustained end-to-end monitoring absent; accuracy telemetry has been dead since 2026-08-07.

Pairing-token auth with no loopback trust; SSRF default-deny including the tailnet CGNAT range; approval boundary with execution-time re-check and revision pinning; 0600 credential storage; clean secret scan. Remaining gate: Restore proof still absent.

The operator UI half of the gate is closed and reachable: 8-agent supervised roster with grounding enforcement, tools-in-chat that search and cite, approval-gated proposals, prompt library. Remaining gate: The usefulness-evaluation half is open. Needs a recorded, replayable evaluation over K>=10 real estimator tasks with a non-mutation assertion.