Methodology
How the numbers are made
Groundbreak is only useful if you can check it. This page describes where the data comes from, how it is kept, how approval probabilities are produced, and what the product cannot tell you.
Sources
Groundbreak uses public records published by the cities it covers. It reads documented public interfaces only, identifies itself where a source offers an application token, backs off when asked to slow down, and never goes behind a login or a bot challenge.
- Parcels and zoning
- Assessor rolls and zoning maps from each city's open-data portal: lot area, zoning and height districts, existing use and assessed value.
- Projects and permits
- Planning cases and building permits as the city records them, normalised into comparable stages from filing to completion.
- Votes and hearings
- Legislative records of the bodies that decide land use: matters, hearings, tallies and how each member voted.
- Law
- Rules are kept as dated rule packs per jurisdiction. Each rule cites the statute, code section or ordinance it comes from and the date it took effect.
Provenance
Every record is stored exactly as it was received before anything interprets it. Interpretation can then be repeated, audited and corrected without going back to the source.
- Raw first
- Each response is saved with its URL, the time it was fetched and a SHA-256 hash of its contents.
- Traceable figures
- Every number shown in the product points to the stored record it was derived from. You can follow it to the source.
- No silent gaps
- Nothing is estimated silently or filled in from memory. Where a record has no value, the product shows that it is missing.
Reading record text
Some facts are only in prose: whether an agenda item concerns housing, how many units a file describes, which addresses it covers. A language model extracts these as structured judgments.
- What is classified
- The kind of land-use matter, the housing units involved and the addresses a record refers to.
- How it is stored
- Each judgment is saved with the model version that made it and its confidence, alongside the text it was read from.
- Review
- Classifications are reviewable and can be corrected. A correction replaces the model's judgment and is recorded as such.
Approval probability
The approval model estimates how a body is likely to vote on a kind of project from how its members have voted on comparable matters. It is a statistical summary of the record, not a forecast of any one hearing.
- Basis
- Recorded member votes on classified land-use matters that are comparable to the project.
- Calibration
- Probabilities are calibrated on recorded votes: among matters given a seventy percent chance, about seventy percent should have passed.
- Intervals
- Every probability is shown with an interval. The interval is the claim; the midpoint alone is never displayed.
- Thin records
- New members, new bodies and unusual projects have few comparable votes. The interval widens accordingly, and the product states that the record is thin.
Pro-forma and delay
Financial outputs are arithmetic on your assumptions and on durations measured from the record. The calculations are deterministic and each scenario is saved with the model version that produced it.
- Pro-forma
- Costs, revenues and returns computed from the inputs you provide. Groundbreak does not supply market rents, costs or cap rates.
- Cost of delay
- The time comparable projects took on each approval path, multiplied by the carrying cost you enter.
What this cannot tell you
A model of the record has limits that matter when money is at stake.
- The past is not the future
- Members change their minds, seats change hands and laws change. An estimate built on past votes can be wrong about the next one.
- Records have errors
- Public records contain mistakes and omissions, and they can lag events. Groundbreak reports the record; it does not correct the city.
- Judgment is still yours
- Negotiations, community positions and politics that never reach a recorded vote are not in the data.