Numbers you can defend
Assay holds every number to rules that survive a review. The same page also states the limits.
01
Unknown is not zero
When Assay lacks a price, it records the usage without a cost. $0.00 would claim the work was free. Every total keeps the unpriced share visible.
$0.00 is a claim.
It asserts that something cost nothing. That is a strong statement to make about a model you cannot price.
"Unpriced" is a fact.
It keeps the gap visible on every screen it affects.
02
A wrong figure is corrected, never rewritten
Assay adds a correction and keeps the original beside it. Reconstruct last quarter's report, then see why the current figure differs.
The original stays
Reproduce the report as it appeared when you sent it.
The correction sits on top
Read what changed and why.
Every page of the console closes by saying where its numbers came from - the corrections folded in, what was left out, and the rates it was priced against.
03
A measured zero and an unmeasurable one stay apart
A measured zero and an unknown are different facts. Assay keeps them separate on every screen and in every ratio.
A measured zero
is a finding. It was looked at, it came back empty, and it stays in the denominator of every ratio.
An unknown
is an absence. It is excluded, counted, and printed next to the ratio it was excluded from.
04
Two honest answers beat one convenient one
Product cost has two defensible readings: an overlapping upper bound and a partitioned estimate that adds up. Assay prints and labels both.
05
Every ratio states what it left out
Every waste figure, outcome cost and percentage carries the count and share it excluded. Shares divide by the whole window.
06
Every number carries its date
Every export states when Assay captured it, what it covered and what it excluded. It also says that the figure does not update.
This site follows the same rule. Every screen is a demo capture; no page exposes a figure from a real ledger.
The limits we state instead of hiding
Each limit appears beside the capability it bounds. Read both before you act on the number.
- You supply the value. Assay records spend, work and result in one place you own. You decide what the result was worth.
- Invoice checks require an invoice. Subscription usage has none, so Assay reports the resulting gap.
- Attribution measures attention. It records what a session worked on. It makes no claim about authorship.
- Missing records leave past costs low. Assay refuses to replace them with a corpus average.
- Erasure is bounded by your own backups. It is immediate in the live store and becomes irreversible only once your backups have rotated past it. Quote your retention window, not an exit code.
- Automatic labelling favors grouping over nuance. A larger local model improves detail while keeping the text on your machine.
- Interactive work produces no outcome. Assay records its spend, transcript and label, then excludes it from cost-per-outcome ratios.
- The repository is private today. The substrate's MIT grant is already in place.
The limits are the product.
Anything can print a confident total. What makes a number worth acting on is knowing exactly what it left out.