Acceptance test · Supplement Part C · 2026.06-hb1389

Engine Calibration

Fed only public data, the engine must reproduce Cirrus's already-adjudicated verdicts — rank the winners high and fire the correct hard-fail on each kill. Right answer for the wrong reason is a bug.

Tests passing

7/7

calibration green

C.100-43-44-XX-XX-XXX-0000

Tests: End-to-end underwrite + ranking · expect: Ranked, top decile

Rankedrank #1

C.2C2-0-DALINDA-LN

Tests: Eligibility filter (§5.1) · expect: Killed — residential zoning

KilledResidential Zoning Lla Inapplicable

C.2C2-FAU-RESEARCH-PARK

Tests: Tenure-type exclusion · expect: Killed — ground lease

KilledLeasehold Or Restricted Authority Land

C.2C2-ARTISTS-ALLEY-DELRAY

Tests: Residual solve as a kill · expect: Killed — ask ≥ MaxBasis₉

KilledAsking Exceeds Max Basis Residual Kill

C.2C2-9924-HAPPY-HOLLOW-DELRAY

Tests: Overlay / carve-out layer · expect: Killed — Ag Reserve overlay

KilledAg Reserve Overlay Contested Conversion

C.2C2-15445-S-SR7-UNINC-PBC

Tests: Jurisdiction resolution (GIS, not postal) · expect: Not ranked — resolved unincorporated

ArchivedBelow 71 Unit Missing Middle Gate

contrastPBC-INSTITUTIONAL-2023

Tests: Basis-spread beats raw yield (§6.4) · expect: Ranked, below the family-trust lead

Rankedrank #2

Exact numerical reconciliation (Supplement Part A: 108 units, ~0.97 rent ratio, MaxBasis₉ ≈ $1.19M) runs in the engine pytest suite — uv run pytest.