Data Integrity
We find our errors first, and we publish the record.
Every platform this size has defects. The difference is whether anyone is looking, how fast they're found, and whether the record is public. Ours is.
The defect registry
Every defect ever found — by an instrument, a user, or our own review — gets a permanent registry entry: what it was, when it was introduced, when it was discovered, how many days it was exposed, and what fixed it. The registry is append-forever: entries are never deleted, including the embarrassing ones.
The instruments
Defect-finding is machinery here, not vigilance. Five standing instruments watch the platform, and a change that fails its gate does not ship.
| Instrument | What it checks |
|---|---|
| Scoring guardrails | 8 checks run after every score write: recommendation coherence, a golden panel of hand-verified parcels, outlier validation, disparate-impact testing, signal-coverage honesty, net-public-value band containment (every figure carries plausible bounds), premise conformance (every benefit stream must read exactly the inputs its cited source counts), and freshness of the priority order with its equity shares reported. |
| Context-layer integrity audit | 19 checks over everything outside the scoring spine — data lineage, join integrity, vintage honesty, endpoint auth, marketing-claim drift, scope-labeled fields. |
| Agent truth sweep | A standing battery of false-denial and adversarial probes against the AI assistant — including prompt-injection and fabrication bait. |
| Persona evaluation matrix | 54 scenarios across real user seats — council members, CDC directors, land banks, residents — graded against written expectations, with raw outputs preserved. |
| Comprehensive audit | A seven-pass audit — including auditing the auditors against the defects they missed — on a semiannual clock. |
Standing disclosures
Some limits are structural, and honesty about them is part of the product — they appear in the platform's own answers, not just here.
- ▸County building-condition grades trace to the 2012 base-year assessment and can be up to 14 years old — every county recommendation carries a "verify on site" posture.
- ▸Out-of-city parcels are scored on a reduced signal set (no city crime or inspection feeds exist there); each parcel discloses its own signal completeness.
- ▸Modeled figures — costs, net public value, projections — are presented as modeled sizes, never as quotes, and every dollar figure decomposes to its formula and anchors.
- ▸Every net-public-value figure is Solid by construction: it counts recovered taxes, neighbors' property value, avoided service cost and safety — the projected community benefit that earlier versions allocated from Opportunity Atlas tract estimates was removed from every figure in September 2026, and the Atlas is now context for the equity check on ranked lists. (Earlier wording: Solid or Projected, depending on whether the long-term community benefit the model projects for nearby children is counted — the least certain part). Measured September 2026: about 45% of positive recommendations depend on that projected benefit for their sign, and we say so.
- ▸Projections assume no new decline: our decline-forecasting experiment failed its pre-registered accuracy floor and was shelved rather than shipped.
- ▸Every data layer carries its source vintage, and fixed-edition sources (like the 2021 Market Value Analysis) are cited by their published date, not our refresh date.
Why publish this?
Because the communities and institutions this platform serves have been burned before by confident numbers with quiet errors. A tool that recommends where public money goes owes its users the error record, the correction speed, and the standing limits — before they ask. If you find something wrong, tell us — user-caught defects go on the same registry, with credit.
The full scoring methodology, source list, and model derivations live on the methodology page.