The formula underpricing framing by six and a half times
An audit of our estimating engine found a framing formula far below real cost, and we refused to patch it with a guess.
We audited the estimating engine, testing dozens of its pricing formulas against a corpus of real bids. Two live bugs turned up, along with a pile of rates never compared to a built cost. The framing lumber formula alone ran about six and a half times below real cost on an actual house, which is a swing near ninety thousand dollars on a mid-sized home.
My ruling: a formula with no real evidence behind it stays as it is. We do not patch it with a guess. Of everything reviewed, eighteen formulas were flagged as impossible to calibrate from the evidence on hand, and we left them alone rather than quietly repair them with a made-up figure.
The standing rule is blunt. With no corpus evidence there is no calibration, and no invented numbers. Output generated by the engine itself does not count as evidence, because that would be circular. We did make some rate changes, and two of them turned out to be accidental duplicates that would have quietly overridden real pricing tiers. We caught those and fixed them where they stood.
On the single project that had a real built cost on file, the estimate was fourteen percent off actual before the patch and four percent off after it. The test suite passed in full at every step.
The honest repair for a broken formula is to say which parts you cannot fix yet, never to produce a better guess. The engine had been wrong for some time without anyone knowing, and I treated that as a system lacking real evidence, not as a person to blame.
Story details
| Project | Kanopi |
|---|---|
| Type | Mistake and repair |
| Year | 2026 |
| Firm telling | on the Common Ground wiki |