Fail-closed valuation: no evidence, no number

A value estimate is only useful if the person reading it knows how much to trust it. Vendor valuation APIs give you a number and a vague confidence score; you cannot see the comparables, the filters or the reason a sale was included. For a solo operator deciding whether to make an offer, that is the wrong shape. So the Keystone platform computes its value range from public county records in the open, and refuses to produce one when the evidence is thin.

The pipeline

1. Source. The county assessor publishes sold sales near a parcel inside a fixed study window. Sales after the cutoff do not exist in that source, so the report states the window instead of implying recency.
2. Arms-length, closed-world. Only the county’s validity code X counts as arms-length. Any other or unseen code is treated as not arms-length. Guessing generously here is how family transfers and foreclosure resales end up pricing a house.
3. Row validation. Each row must fit a closed schema: price, a parseable date, living area, no person keys, no duplicates. Rejections carry a fixed reason so the operator sees exactly what was thrown out.
4. The subject’s own sale is not a comparable. This was a real defect: one subject’s prior sale ranked as its top comp and pulled the band wide. Excluding it moved the result from a low-confidence $392k–$722k to a moderate $395k–$633k. The number changed because the evidence got cleaner, not because anything was tuned toward an answer.
5. Similarity scoring. Distance, recency, living area, age, style, exterior, basement, bedrooms and baths, plus a bonus when the county itself designated the sale as a comparable for this parcel. The top six are selected; every candidate keeps its inclusion or exclusion reason.
6. The number. Score-weighted median $/sq ft times the subject’s living area for the point; weighted P25–P75 for the band, widened by ±0.25 × the coefficient of variation; rounded to the nearest thousand.
7. Confidence as named factors. Comp count, $/sq ft dispersion, maximum distance, recency against the window end and subject-facts completeness each have a threshold for the high tier, and the factors print with the number. A moderate tier tells you which factor held it back.

Where it fails closed

Fewer than three accepted arms-length comps means status insufficient and no dollar field — not a wide band, not a hedged number, nothing. The same rule applies when the comparables belong to a different subject, when a ZIP straddles a jurisdiction boundary, when the county search returns two plausible parcels, or when the source refuses runtime access. In each case the response names the step that stopped and the evidence that would unblock it.

One presentation detail mattered more than I expected. The lookup’s status is step-level — “every source step resolved” — and can be complete while the value is insufficient. The first version of the console showed a green “complete” banner over an empty value card. Now the composite state has its own amber banner: evidence capture complete, value insufficient. Same data, honest reading.

What the number is and isn’t

It is internal decision support over candidate-grade evidence, for offer-range thinking. It is not an appraisal and it is never published. Interior condition is unverified, the study window is a structural staleness limit, and the confidence factors say so on every result. The value of the tool is not that its number is right; it is that you can see exactly how right it can be.

You can step through the recorded cases, including the failures, on the Keystone platform page.