Method: the pipeline every file travels
Every brief passes the gates or it does not ship.
Most data products describe their coverage. We publish our checks. This page is the pipeline a case file travels before a firm ever sees it, including the parts where a human says no.
Five steps, in order.
- 1. Discovery
The full public record is read across HPD violations, DOB filings, and OATH/ECB case data. Nothing is sampled; the whole set is scored.
- 2. Quality-first selection
Cases are ranked on open hazardous load, penalties of record, and recurrence, then screened: institutional, public, and receivership buildings are excluded by a gate with a versioned decision ledger, not by eyeball.
- 3. Verification
Each selected case is checked against the live city record: the address must match the city page, contact fields must belong to the named party, a second contact path counts only if it is genuinely distinct, and the count on the city link must meet or exceed the count in the brief.
- 4. Render through the claim gates
The brief renders through automated claim-honesty checks before a file is written. If one check fails, there is no file.
- 5. Human final read
A person reads every case file before it ships. This step is permanent. It has caught what the automation missed, which is exactly why it stays.
The claim gate: 32 automated checks against overclaiming.
Before a surface ships, its output is scanned by a gate built for one purpose: catching claims the record does not support. A sample of what it refuses:
| Banned money language | Dollar figures are penalties of record or pending matters. Words that assert a live running balance are banned tokens; the build fails if one appears. |
| Invented inference | No sentence may assert owner intent, predicted outcomes, or holding history from the absence of a record. Absence proves nothing, so it claims nothing. |
| Basis gating | The phrase penalties of record renders only when the figure is an adjudicated amount from the city record. Estimates say estimated. |
| Surface coherence | The same case must make the same claim on screen, in the PDF, and in a letter. Divergence fails the build. |
| Provenance wording | A recurrence signal keyed on a registered office is described as exactly that, never as deed ownership it cannot prove. |
The findings ledger: we publish our own mistakes.
Every defect we find in our own claims gets a numbered finding, a fix, and a permanent regression test. A sanitized sample from the ledger:
-
Found, fixed, tested
A data field that read "No" was treated as a yes by a truthiness check, rendering an ownership transfer that never happened. Fixed with an explicit yes-parse; a test now asserts negation strings can never gate a claim.
-
Found, fixed, tested
A freshness stamp could be refreshed without regenerating the data it described. The stamp is now written only by the build that produced the artifact and carries the artifact's fingerprint.
-
Found, fixed, tested
Two date windows that should have been identical differed by one leap day, silently under-counting a sliver of history. Caught by the baked-versus-live gate; the windows now derive from one source.
A vendor that cannot show you its error ledger is asking you to assume it has none.