A deployment-document intake agent that knows which of its own answers to trust.
Infrastructure projects run on paper. Permits, notices to proceed, interconnection agreements, inspection reports — each one read by a person and hand-keyed into the system of record. It is slow, and it fails quietly: a transposed permit number or a wrong expiration date surfaces as a stalled build six weeks later.
Extraction is the easy part of fixing that. Any current model reads a permit. What decides whether anyone keeps using the system is whether it knows which of its own answers to trust — and hands the rest to a person before they reach the record.
Modelled, not measured. The figures move with the assumptions above and with the calibration. The last line is the one that matters in a review: throughput bought by loosening the bar is paid for in errors nobody catches.
Turn the verification pass off and watch the fee field. It posts $1,240.00 — confidently, plausibly, and wrong. The document's total is $2,225.00; the extraction took the first line item. No confidence score catches that, because the model is not uncertain. Only a second pass reading the source is.
That case is seeded deliberately in this run and should be read as an illustration, not a failure caught in the wild. Whether a particular model misreads a particular fee table varies run to run, which is precisely the argument: a pipeline that only works when the model happens to be careful is not a pipeline. Extraction is a commodity. Knowing what not to trust is the product.
Sit with the people doing intake today and take the real inventory.
Run a labelled sample and set the bars against measured behaviour.
Into the system of record, with the queue where reviewers already work.
python3 plancheck.py sample_permit.txt · --no-verify to see what escapes · --sweep for the calibration curve.