Sketch Space: An Evidence-Aware Computational Pipeline for Residential Development Feasibility

Abstract

Sketch Space assembles a first-pass residential development screen from a street address. It merges public parcel, footprint, assessor and permit records into one record with per-field provenance and explicit unknowns. It places a detached accessory dwelling unit (ADU) under California's state-law path (Gov. Code §66323(a)(2)) with a rasterised search whose candidates are re-checked against the encoded setback and separation rules. It also builds an owner-confirmed model of the existing house, permit steps and a budget range. We evaluate the screen at a fixed commit, under protocols committed before each run. On approved permits for new ADUs, the screen finds a placement on 37 of 40 San Francisco development lots and on 195 of 200 Los Angeles lots (97.5%; 95% CI 94.3–98.9), a city not used in development; the pre-registered transfer criterion failed only because recall was higher. It also places a unit on 81.5% of random San Francisco RH-1 lots and 93.5% of random Los Angeles single-family parcels, so recall cannot show that it rejects infeasible lots; precision is unmeasured. Where Los Angeles permits state the approved footprint, it fits the screen's region on 45 of 63 lots (71%); on the rest, the screen is stricter than the approval. A reference search exact in translation finds 4 of 57 "no" answers to be search misses. End-to-end runs from 101 addresses found the correct parcel in 95 cases and exposed a geometry-format defect that keeps the solver out of one product path. Status: preprint, version 1, not peer reviewed. The supplementary archive contains the experiment protocols, result reports, aggregated and pseudonymised per-lot results, the literature table and a claim-traceability table. It contains no source code or addresses. Competing interests: the author is the founder of Sketch Space, the commercial product evaluated. AI use: an AI assistant (Anthropic Claude) carried out much of the code audit, experiment code and runs, literature search and verification, and drafting; see the paper's statement on AI use.