Why every takeoff line should link back to the page

When we started the alpha, we made the mistake every AI takeoff company makes: we showed estimators a beautiful list of line items and told them the model was very accurate. They asked one question: 'where did this come from?' And the demo was over.

Trust isn't a number — it's a click

high accuracy alone sounds great until it's the small uncovered fraction that loses you the bid. The estimator's job is not to trust the AI; it's to verify the numbers fast enough to ship the bid. Anything that slows verification is anti-value, no matter how shiny the AI was.

The fix isn't 'better AI' — it's structural. Every line has to ship with a back-pointer to the exact crop on the exact sheet that produced it. Click the count, see the receptacle. Click the SF, see the wall. Click the spec section, see the page.

What we ship instead

• Every takeoff line has a source_refs array — bbox, page, document, OCR snippet, confidence
• Every confidence score has a rationale string — vector grounding hit, library cross-check, trade-mismatch, etc.
• Every spec line has a link to the section number and the snippet that triggered it
• Every line you confirm or reject becomes ground truth that re-trains the recognizer the next bid

This is slower to build than 'just call OpenAI.' But it's the only thing that survives a senior estimator opening their laptop on a Friday afternoon.

We don't trust the AI. We trust the evidence. That's the whole product.