Why we show the page

The reasonable first reaction to AI-generated estimation is suspicion. A points schedule is a commercial commitment: price from a wrong one and you either lose the tender or win it at a loss. An output you cannot verify does not reduce your risk, it relocates it somewhere harder to see. So the first design decision in Tenderfly was not about extraction at all. It was that every number must carry its receipt.

What checking used to cost

An estimator reviewing a colleague's schedule has one real option: re-read the pack. That is why checking rarely happens in full. A 900-point schedule built from 80 documents cannot be re-derived inside a tender deadline, so review becomes sampling, and sampling means the errors that survive are the quiet ones: the double count, the missed exclusion, the plant priced down the wrong control route.

What we do instead

Every point Tenderfly generates traces to its origin. Click a point and it opens the source document at the right page with the evidence highlighted: the row of the schedule, the label on the drawing, the sentence of the specification that produced it. Where a requirement came from spec prose rather than a tagged item, the sentence itself is the citation.

This changes what review means. Verifying an item takes seconds, so spot-checks become cheap and disagreements become lookups. When the generated count differs from the client's own schedule, or from your estimator's instinct, the question "why" has a clickable answer instead of an afternoon of re-reading.

It keeps us honest too

Citations are not only for the estimator's benefit. Extraction can be wrong: a smudged drawing, an ambiguous clause, a tag convention we have not seen before. A system that shows its sources gets caught quickly and precisely when it errs, and the correction lands in exactly the right place. A system that hides them gets trusted until the day it should not have been, which is the worst possible failure mode for a tool that writes numbers into bids.

The principle

An estimate is an argument, not just an output. It says: this equipment exists, it needs these points, they cost this much. Arguments come with evidence. Software that wants a place in that process owes its user the same standard a careful colleague would meet, which is to answer "where did that come from" by opening the page.