How BuildNYC predicts plan examiner objections

Before you file, the Plan Exam check shows you how NYC DOB actually interprets and enforces the building code — flagging the objections your filing is most likely to draw. It is built on the department's complete record of plan-examiner objections, so it reflects the code as it is applied in review, not merely as it is written. Every figure on this page comes from a held-out test — filings the model never saw while it was built — so nothing here is asserted; it can be checked.

1. What it predicts

Plan examination returns objections — items an examiner marks as missing, incomplete, or non-compliant, each adding days to weeks of re-submission. The check predicts which objections your specific filing is most likely to draw, ranked by confidence.

It deliberately does not try to predict who reviews your job. DOB assigns examiners by workload, not by anything on the filing, so the reviewer is effectively a lottery. Predicting the person is a losing game; predicting the objections is not.

2. Data sources

Every figure comes from DOB's own records, queried from BuildNYC's database (never a live public API at read time):

• DOB NOW objections — over 588,000 recorded plan-examiner objections, each parsed into the underlying requirement, the code cited, and how it was ultimately cleared.
• Continuously updated — the record grows as DOB posts new objections, so the enforcement pattern the check reflects stays current, with no retraining.

3. How the prediction works

The forecast comes from DOB's enforcement record — not from a model's opinion about the code. Every one of the 588,000 objections is parsed into the requirement it enforces, the code section behind it, and how it was resolved. When your filing runs through the check, it identifies which of those requirements the department is most likely to cite it on — ranked by how consistently that requirement is applied and how strongly the signal is corroborated. The output is a confidence-ranked list: a trustworthy top, and a broader "also verify" tail.

4. The code as it's enforced — not as it's written

The building code is one thing; how NYC DOB examiners actually apply it is another — and the gap between the two is what costs filers time. A requirement can look satisfied on paper yet still get cited in review; another can pass where the letter of the code might suggest otherwise. Because the check is built entirely on what examiners actually cited — the department's real, requirement-by-requirement enforcement pattern — it reflects the code as it is practiced, not just as it is stated. That practiced interpretation is exactly what you can't get from reading the code, and exactly what decides whether your filing clears.

5. How we validate it

Back-test numbers are only worth the discipline behind them. Ours were produced on a held-out set of 79 filings — chosen so that none of them were used while building or tuning the model. Because these are separate filings, each reviewed by whoever DOB happened to assign, the set spans 51 distinct examiners and every plan-exam discipline — so the numbers reflect the department broadly, not one reviewer's habits.

For each held-out filing we generate the prediction, then compare it against the 273 objections those filings actually received. A prediction counts as correct only when it covers the same substantive requirement as a real objection — paraphrasing is allowed, naming the wrong requirement is not. That yields two numbers:

• Recall — of the objections a filing truly drew, what share the check surfaced.
• Precision — of the items the check flags, what share are real objections.

6. Results

7. Why it matters — objections are what delay approval

A prediction is only worth acting on if objections actually cost time. They do — and the effect is large. Using DOB's own filing-to-permit dates, we measured how long filings took to be permitted, grouped by how many objections they drew — controlling for construction floor area (a far more reliable size measure than declared job cost) to rule out the obvious confounder that bigger projects both draw more objections and take longer.

Median days from filing to first permit · DOB NOW filings, 2022–2025 · 550,000+ permitted filings.

That is the whole point of the check. Find the objections before you file, fix them, and file clean — so your filing clears on the first pass instead of cycling through re-review. The relationship is correlational, but the mechanism is direct: every objection triggers a resubmission and another round of review.

8. What it is — and isn't

• It predicts objections, not the examiner. Reviewer assignment is workload-based and effectively random, so naming your specific reviewer is out of scope by design.
• The strict number understates real value. Ground truth is each filing's recorded objections; a correct flag that reviewer skipped is scored as a miss.
• Subtle, filing-specific items are the frontier. Objections that hinge on a single dimension or a one-off detail are the hardest signal to read, and where the model is weakest today.
• It improves on its own. The forecast is outcome-based, so it sharpens automatically as new objections are recorded — no retraining required.