We publish what we got wrong

Every construction-code database has errors. Ours included. The difference we are willing to be judged on is that we go looking for them, and we tell you what we find.

This page is the record — not a summary of one. Every number on it is queried from the database when the page is built, including the numbers that make us look bad. If you are deciding whether to trust a code reference with a plan review, this is more useful than a feature list.

What we removed

32,193 code sections withdrawn in total. They are retained internally as the audit trail and excluded from search, from the assistant, and from the site.

31,205 sections that were not code

Each had a real section number and an invented title over template text: “This section establishes specific code requirements. Refer to the official…” The dangerous case was not the empty ones. NEC 680.22 governs receptacle placement near pools and spas. It was being served with the body “This section covers allowable building height, area, and number of stories” — IBC Chapter 5 text on an electrical shock-hazard provision. All of them were embedded, so all were reachable by search and by the assistant.

We had published the larger number. It appeared on our home page and in our own planning documents. The real figure is 14,156 distinct sections across 640 code editions.

923 duplicate copies of real sections

The same text filed under two section numbers — 404.1.2 and R404.1.2, 3501.1 and E3501.1. Both copies published, both were embedded, and they disagreed with each other about their own titles. Before removing them we checked each pair and moved the better title onto the copy we kept, because sometimes the duplicate was the one that had it right.

65 rows that attributed code provisions to states that may not have made them

Rows titled “Georgia Amendment: Guards Required” and “Ohio Amendment: Automatic Sprinkler Systems” carrying text byte-identical to the unamended model code. The worst cluster was IBC 903.2, where automatic sprinklers are required: ten states sharing one paragraph.

A second, smaller set was found only by asking what the first check would miss. It makes the identical claim without using the word amendment — “Illinois — Where Required”, “Georgia — Group A Sprinklers” — so a check keyed on that word could never have seen it. Sprinkler thresholds and wind loads are life-safety provisions. Telling a designer their state has amended a section when it may not have is the most consequential error on this list.

What we have not fixed

These are not removals. They are problems we can measure and have not yet solved, which we would rather state than leave for you to discover.

80% of our DOT specification text never reaches the assistant

Our assistant reads a bounded excerpt of any specification section when answering. Most of our 9,201 verified DOT sections are longer than that, so the majority of that text is not shown on any given query. Until recently it was cut off silently. Excerpts now carry an explicit marker saying what was omitted, so the assistant cannot treat the end of an excerpt as the end of a requirement.

Worse: 10 sections hit a hard character ceiling when we ingested them. That text was never stored, so no amount of indexing recovers it — only re-reading the source. We have now confirmed source documents for 52 of 52 state specification manuals so that re-ingestion can happen. Until it does, a DOT section on this site may end mid-requirement, and it will say so.

910 DOT sections we could not verify

Sections attributed to state DOT manuals that turned out to be model-generated summaries rather than extracted text. They read as authoritative and were specific enough to act on. One example, checked against the source: a California Highway Design Manual page claimed a 10% maximum superelevation rate for rural highways. Caltrans Topic 202 actually specifies 12% for two-lane rural conventional highways, and 4% — not 6% — for urban streets below 35 mph. A designer following our page would have been given an unconservative number on a real safety parameter. All are withdrawn.

The checks that find these

Each of the above was found by a person looking. That does not scale, so each one is now a query that runs continuously. 27 checks; 8 of 9 marked critical currently pass, and 13 checks in total are failing. We would rather show you a failing check than a page implying everything passes.

Checks last evaluated Fri, 09 Oct 2026 04:07:00 GMT. Figures on this page computed Fri, 09 Oct 2026 04:07:00 GMT. We show these timestamps because a stale number should be detectable rather than silently served.

What we refuse to serve

• Unverified DOT specifications are gated out of retrieval at the database level, not filtered in the interface — two separate search paths, both gated. When we found that only one of them was, we fixed the other the same day.
• We do not manufacture verification dates. Not one of our 14,156 sections has been independently verified against the issuing authority’s publication, so not one carries a verification stamp. Section pages say “added [date] · not independently verified” instead. We could have backfilled ingest dates and shown 14,156 green ticks. That would have been a lie with a timestamp on it.
• We do not infer a permit fee we cannot source. 332 jurisdictions price permits on construction valuation rather than a flat fee. Their notes are full of percentages — but most of those are plan review expressed as a share of the permit fee, not of the project. Reading them as valuation rates would have produced a permit fee of 65% of valuation. We model the 7 that state their rate outright and say “not modelled” for the rest.

What we do not have yet

That last distinction matters and it is the kind of thing usually left ambiguous. 1,064 jurisdictions have a fee schedule we have located and dated. 164 have that and the rules and the amounts. When we say “coverage,” we mean 164.

Why we publish this

Two reasons, and neither is modesty.

The first is that you can check it. Every figure here comes from a query against a database you can spot-check by opening any section and following its source link. A claim you can verify is worth more than a claim you cannot.

The second is that a code reference which has never found an error in itself is not error-free — it is unexamined. We found 31,205 fabricated sections because we went looking. We would rather tell you that than have you find the next one.

If you spot something wrong, tell us and it goes on this page.

What is not for sale

Manufacturers can pay to reach the people reading these pages — that is a real revenue line and we are not going to be coy about it. Product pages carry a Request Information unit, and sponsored placements will appear alongside assemblies and specifications, labelled.

No paid placement inside Code AI answers, ever.Not as a cited source, not as a “related product”, not as a footnote. An answer that carries paid placement is not an answer, and the citation is the only thing we are actually selling. If that ever changes, it changes on this page first.