Every Assay number is supposed to survive a review. These rules are how.

Unknown is not zero

If Assay cannot price a model, it records the usage without a cost. $0.00 would claim the work was free. Every total keeps the unpriced share visible.

A wrong figure is corrected, never rewritten

Assay adds a correction and keeps the original beside it. You can reproduce last quarter’s report, then see why the current figure differs. A model the table still cannot vouch for gets no correction. Nothing is estimated from an average.

A measured zero is not an unknown

“This cost nothing” and “we cannot measure this” are different facts. Product reports keep them in separate buckets. Merging them would print a hole as a finding.

Two readings, both named

Where a question has two honest answers, Assay prints both and labels which is which.
  • Inclusive — an upper bound that does not partition.
  • Weighted — the estimate, and it does partition.
It never picks one silently. See Products.

Every ratio names what it excluded

Waste, cost-per-outcome, token yield, and ROI compute over priced spans only. Each report carries the count and share it left out. ROI with no declared value is undefined, not −100%.

A zero is rendered with its reason

$0.00 recoverable, a 0% token yield, a $0 cost-per-outcome — each is a result conditioned on something. The surface says what.

A snapshot says it is a snapshot

A static export stamps the capture instant, the store, the version, the window, and the filters. It ends with: these numbers do not update.

A label states its source

Declared, judge, or heuristic. A reader can see which, and the rank is always: declared > judge > heuristic Assay validates that a declared label is in the vocabulary. It does not check that the label is true. A producer that marks every sitting build:feature gets a clean, useless aggregate — reported faithfully.