What this covers, and what it does not

2,690 of 2,895 rules can be answered against a file

The rest are stated, cited and readable, and the atlas will not pretend to check them. Each one carries a written reason, because “no test” with no explanation is how a gap becomes permanent.

Settled before occupancy

11

Fixed at site plan approval or in the affordable housing agreement, so they are over before there is an occupied unit to carry facts. A developer needs them; a compliance file cannot answer them.

The agency acts, not you

14

Where the actor is the state or the city. Testing one would report you out of compliance for something the agency had not yet done — a verdict against the wrong party.

We need a fact we don't collect

180

Real obligations of yours that could be checked, if the atlas asked for the right fact. This is the only one of the three that is a backlog, and it is the one still flagged internally until it is closed.

Limits: 10 of 16 tables are loaded, and 6 stay empty

The atlas holds 126,638 published limit values across 10 of the 16 tables it describes, so for those it answers a rent or income question directly rather than pointing you at a chart. 6 tables hold nothing, and 44 current rules test against them. For those the only truthful answer is insufficient_evidence — not pass, not fail — and the row for each says why it is empty, rather than leaving you to assume it is a backlog.

One of the empty ones is empty because it does not exist: HUD publishes MTSP income limits and no MTSP rent table at all. The atlas derives the LIHTC maximum rent from the income limits under IRC 42(g)(2), which imputes 1.5 persons per bedroom — and an odd bedroom count imputes half a person no published table has a column for. Agencies do not interpolate alike, so the derived cap is reported as a band, and a rent inside that band comes back unresolved rather than decided.

5 of these are not a backlog item at all. Utility allowances, PHA payment standards and Live Local market studies have no central publication to ingest — they are elected per agency or commissioned per property, which puts them in the property layer described below rather than in any queue.

Tell us which limit table you needed bound →

Depth is uneven across jurisdictions

93 jurisdictions have at least one rule ingested. The range is wide: Arizona has 8, Florida has 137, and the median is 18. A jurisdiction near the bottom of this list is not a jurisdiction with fewer obligations — it is one where less of the agency's manual has been read into the corpus. 1 sit below half the median and should be treated as a starting point rather than a survey.

Ask for depth in a specific jurisdiction →

Depth is far more uneven across programs

The corpus is, in practice, a Low-Income Housing Tax Credit corpus: 1289 of 2895 active rules (45%) belong to it.

The jurisdictions column is the one to read before trusting a resolve. A program with a national rule count but a single-digit jurisdiction count has a federal floor and almost no state overlay, so resolving it anywhere else returns the floor and says so.

Request a program →

3740 questions the agencies' own materials left open

These are recorded per jurisdiction on state_coverage.open_questions during ingestion: points where the agency's published manual was silent, ambiguous, or could not be retrieved. They are not rules with low confidence — they are places where no rule was written because the source did not support one. Counting them is the point; a corpus that recorded no open questions would be one that was not reading carefully. Data status prints every one of them, grouped by jurisdiction and most-open first.

Deliberately out of scope

These are not backlog items. They are boundaries of what this corpus is.

How fresh the sources behind this are

Every rule cites a dated document, so staleness is queryable rather than guessed. Two different clocks matter and they are not the same thing: when we last verified the document was still there and unchanged, and whether the document itself is older than its own publication cadence implies it should be. A source can be verified this morning and still be an obsolete edition.

• 0 sources have never been re-verified since ingestion, and 0 are overdue against their own declared cadence. The oldest verification on record is 2026-08-22.
• 289 sources have no publication date recorded, which means the age test cannot run on them at all — they are neither fresh nor stale, they are unmeasured. 0 declare no re-check cadence.
• 23 source URLs did not resolve on the last archive pass, and 18 documents are editions the agency has replaced but still serves. Both are recorded rather than removed, because a broken citation you can see beats one you cannot.

Data status ranks the full refresh queue, source by source.