Methodology

How Groundline reaches an answer

Written so a reviewer can disagree with a specific step, rather than with the product as a whole.

The evidence model

Every value Groundline shows is one of exactly four things. There is no fifth kind, and a language model cannot create any of them.

Confidence propagation

A calculation inherits the weakest confidence of its inputs. One provisional standard downgrades everything computed from it, all the way to the report cover. Confidence is never averaged upward, because an answer is only as good as its weakest link.

Geometry

Parcels are projected into a local tangent plane anchored at the parcel centroid using the WGS84 radii of curvature at that latitude. Web Mercator is never used for measurement — its area error at mid-latitudes exceeds 50%.

Setbacks are applied by half-plane offset on convex parcels, which is exact. On non-convex parcels an infinite half-plane derived from a short interior edge would clip the whole polygon and can understate the envelope threefold, so Groundline instead solves the correct predicate — every point at least the required distance from each lot line segment — on a 0.25 m grid and extracts the boundary. The method used is reported with the result, and irregular parcels carry reduced confidence.

Yield

Groundline reports the binding constraint alongside every capacity figure. Knowing the maximum floor area is 4,200 sq ft is useful; knowing it is capped by FAR rather than by the setback envelope is what tells you whether a variance would change anything.

Use qualifiers written in prose in the ordinance ("on corner lots having not less than 75 feet of frontage") are stored as machine-checkable preconditions and evaluated against the parcel. A use whose precondition fails returns zero units with the reason, not a theoretical density figure.

Cost and returns

Construction estimates are parametric and labelled AACE Class 5 — screening accuracy of −30% / +50%. Construction draws follow an S-curve rather than a straight line, because a straight-line schedule overstates early interest by 10-15% on a typical build. IRR is solved by bisection over a bracketed range rather than Newton's method, because development cash flows are frequently non-normal and Newton can converge on a spurious root; when no sign change is bracketed Groundline returns "undefined" rather than a number it cannot stand behind.

Valuation

Four approaches, each declaring its own reliability from the quality of its inputs, reconciled with weights that are shown to you. Residual land value is reported separately as the maximum supportable land price rather than blended into an as-is value, because the two answer different questions and blending them produces nonsense when the development programme does not pencil.

The Copilot

The model interprets the question, selects tools, and writes prose. It cannot calculate. Before any answer is shown, every number in it is checked against the values the engines produced during that turn; an ungrounded figure blocks the answer. Prohibited-claim checks are negation-aware, so the Copilot can say "this is not a certified appraisal" without tripping the control that stops it claiming to be one.

Jurisdiction readiness

A jurisdiction is not published because a shapefile loaded. It is published when a set of manually reviewed golden parcels reproduce the expected district, overlays, setbacks, buildable area, use statuses and unit counts, and every citation in the bundle resolves to a section that exists. Current suite: 65 assertions across 12 parcels for the demonstration jurisdiction.

Versions

Platform

0.1.0

zoning engine

1.2.0

envelope engine

1.3.0

yield engine

1.1.0

cost engine

1.0.0

proforma engine

1.2.0

valuation engine

1.1.0

comps engine

1.1.0

planMatch engine

1.0.0

copilot engine

1.0.0

report engine

1.0.0