How this actually works.

Searches: retained for zero seconds. Addresses and school names you type are used to build a query in your browser and are then discarded when the page is closed. They are never transmitted to CivicByte, so there is no log to keep, subpoena, or leak.

Traffic measurement: aggregate only. The site counts page views through a standard analytics service so I can see which tools get used. It records no identity and is never joined to a search. The hourly job that reads those totals writes only summary numbers — views, sessions, and which pages were popular — into a public file in this repository.

Email: kept until it's resolved. Messages sent to the address below live in an ordinary mailbox. The submission form doesn't post anywhere — it opens your own mail client with the text pre-filled, and your browser keeps a draft on your device that is never sent to me.

Third-party requests. Loading a page requests fonts from Google Fonts and, on the tool pages, data from NYC Open Data and NYC Planning. Those parties see the request your browser makes to them, under their own policies.

The target is WCAG 2.1 Level AA. This is a self-assessment, not a third-party audit, and it is written to be checked rather than taken on faith.

What is in place. Every page opens with a skip-to-content link and uses real landmarks — one h1, a labelled nav, headings in order. Every input has a programmatic label. Tool results render into aria-live regions, so a screen reader announces a result instead of leaving the user to hunt for it. Pages are fully operable by keyboard, and no page traps focus. Wide tables scroll in their own container rather than forcing the page sideways. Text uses relative units and reflows to a 390 px viewport with no horizontal scrolling.

Motion. The scroll animations and the animated skyline are decorative. When your system asks for reduced motion, every reveal resolves immediately and the hero parallax is disabled — the content is identical either way. Content is never hidden behind an animation that must play first: printing a page shows everything.

Known gaps, stated plainly. Colour contrast has not been machine-verified on every surface, particularly muted metadata text on the cream background. No audit with a screen reader by a daily screen-reader user has happened yet. The animated skyline on the homepage is aria-hidden and decorative, but it is heavy on very old devices.

If any part of this site blocks you, email me and I will fix it. That is a commitment with a person attached to it, not a form.

• A clean result is not proof of a clean building. It means the city has no matching record in that dataset. Violations go unreported, complaints go unfiled, and addresses occasionally fail to match.
• Eviction records show executed marshal evictions only. A building with no results may still have active housing-court cases, pending warrants, or informal displacement — none of which appear in this dataset.
• Landlord Watch finds registered names, not true ownership. HPD registration reflects what an owner filed. Portfolios held across several LLCs will not group themselves under one search.
• The condition score is a reading of open violations, not a verdict on a landlord. It is an average, so one severely distressed building can drag down a portfolio of well-kept ones — and the reverse. The building table is sorted worst-first so you can see which it is. The method is below; the counts underneath it are the city's.
• School spending is what NYSED reported for the year shown, on their definitions. It is not a live budget and lags the current school year.
• Address matching can miss. When GeoSearch cannot resolve an address the tools fall back to text matching, which is less exact. Including the borough materially improves the match.
• If the city's API fails, the tools say so. An outage or rate limit shows an error — never an empty result dressed up as good news.

How the condition score is calculated. Every open HPD violation in the portfolio is weighted by HPD's own severity class — A×1 (non-hazardous), B×2 (hazardous), C×5 (immediately hazardous: no heat, no hot water, lead, vermin). The weighted total is divided by the number of buildings queried. That figure — weighted open violations per building — is then mapped to a score out of 10, where 10 is best:

• 0 per building — 10 / 10, no open violations on file
• 2 per building — 9 / 10, few open violations
• 8 per building — 7 / 10, below-average violation load
• 20 per building — 5 / 10, elevated violation load
• 45 per building — 3 / 10, high violation load
• 150 or worse — 0 / 10, severe violation load

Those are anchor points, and the score moves smoothly between them rather than jumping in steps — two landlords in roughly the same condition still separate instead of collapsing onto the same number. A portfolio carrying any open violation at all rounds down, so nothing but a genuinely clean record displays as a flat 10.

Two things about that are worth saying plainly. The weights and the bands are mine, not the city's — HPD publishes the violations and their classes, and nothing else here. They are written down so you can disagree with them and recompute from the same records. And the score is not adjusted for building size, because HPD's registration data carries no reliable unit count; a 200-unit building and a six-unit building count the same, which flatters small landlords and penalises large ones. Any open Class C violations are shown in the score block regardless of the score, because a number should never read as reassurance while an immediately hazardous condition is open.

If a record here contradicts a document in your hand, the city's record is the one to challenge, and the dataset link on each tool page is where to start. Tell me and I will look at whether the tool read it wrong.

Harry Honig · NYC, 2026