LandClear

Residential entitlement — the process of determining whether a piece of land can actually be developed, and on what terms — is a workflow built around two questions, asked in order. First: is this site even worth pursuing? Then: if so, how should I price the risk? The first question is answered today by hours of attorney time and zoning research that mostly rules sites out. The second is a memo most entitlement professionals write by hand, parcel by parcel.

LandClear takes a parcel address and produces structured output for both stages: a screening verdict (go, proceed-with-caution, or no-go) with rationale, and a pre-acquisition due-diligence memo with approval path, key risks, unknowns, and deal-term implications. The goal is not to replace the human decision — it’s to compress the parts of the workflow that are mostly information retrieval, so the human time goes to judgment.

Two-stage decision model, four-layer architecture: facts → decisions → risks → output. Facts stay separate from judgments; risk stays separate from memo output. The separation is what makes the system reviewable — a partner can read the assessment and trace any conclusion back to the underlying facts and the rule that mapped one to the other.

Hybrid data: a static Postgres reference layer for things that change slowly (jurisdiction zoning codes, general-plan designations, overlay districts), and real-time Claude web search for things that change on a regulatory or state-bill cycle. The static layer is the source of truth where it exists; the dynamic layer fills gaps with explicit attribution.

That last point is one of the design choices worth flagging on its own. Stage 2 output validation includes guards against the model fabricating quantitative specificity not present in input data — the kind of failure mode that’s invisible in a demo and corrosive in production.

• Next.js front end, Supabase (Postgres) for storage, Claude Sonnet via the Anthropic API for generation.
• A normalized schema covering all four layers: jurisdiction, zoning reference, general-plan reference, parcel, five categories of environmental and regulatory overlays, assessment, risk items, and the due-diligence memo — with full constraint coverage (cascade deletes, uniqueness, CHECK constraints).
• Provider layer: real Claude as the primary path; a context-aware mock fallback when no API key is available, useful for offline development and reproducible demos.
• Output parser validates verdict enum, risk level, and all five memo section keys; rejects on missing required keys or hallucinated quantitative specificity.

The end-to-end pipeline runs against real Southern California parcels with live Claude generation. Two parcels are validated through both stages — one walks the full pre-seeded narrative path, the other proves the system can handle a parcel that’s only seeded with basic facts and let the model do the rest.

Beta-1 closed. The system is currently in private beta with development partners; the next build direction depends on validation feedback on whether the decision model and output structure match how entitlement professionals actually make decisions. Not externally available.