Standard · Version 1

The Counted Claim Standard.

One shape for every result. A number travels with how it was measured, how many tries stand behind it, and whether it meets the bar of the one relying on it.

The six parts.

Every counted claim has the same six parts, whatever it measures. Four steps produce them: Measure, Count, Prove, Check.

01ResultThe number you are shown.
02MeasureHow the number is measured, fixed in advance.
03CountEvery try behind the result, none left out.
04ProofThe evidence for the result, with its grade stated.
05BarThe bar of the one relying on it, in plain words.
06AnswerA signed ALLOW or REFUSE: does the result meet the bar.

The rules.

  1. Every try is registered on the public ledger before its outcome exists. A try that was not registered first does not count toward the result.
  2. The measure is fixed in advance. It cannot change once the first try is registered.
  3. The buyer’s bar is fixed and read back before anyone is checked. The buyer sees the exact rule their words became, then the check runs.
  4. The answer is signed and re-checkable offline with the published key, by anyone, without asking MIZAN.
  5. Refusals are published under the same name as passes. A record shows what failed as plainly as what passed.

Evidence grades.

Every claim states its grade. Grades are never mixed and never inflated: a claim carries the lowest grade any of its parts can support.

DECLAREDThe seller says so. Nothing has been checked.
RECORDEDLogged on the public ledger as it happened, with its source named.
PROVENChecked by math. The proof verifies without revealing the method.

Profiles.

A profile fixes how one kind of result is measured and counted. The shape stays the same. 27 measures are live today, any result, one standard. Any skill an agent can define is counted. Any skill graded by rule is proven.

AI evaluationLiveA model’s score over tasks declared before the first one ran. Score and count can both be PROVEN.
Task completionLiveOf the tasks an agent took on, how many it finished to the agreed measure. Every task counted, not only the finished ones.
Cost per completed taskLiveWhat each finished task actually cost, with the unfinished ones in the denominator.
Instruction adherenceLiveHow often an agent did what it was told, measured against the instruction fixed before it ran.
Factual error rateLiveHow many of an agent’s stated facts were wrong, out of every fact it stated.
Payment behaviourLiveWhether an agent paid what it owed, on time, across every payment it was party to.
Latency SLALiveHow often a service answered within its promised time, every call on the record.
Tool-use safetyLiveHow often an agent’s tool calls stayed inside the rules it was given, across every call.
Run consistencyLiveWhether the same task gives the same result when run again, from every repeat, not the best one.
Agent workProfileTasks an agent was hired for, and how many it completed to the agreed measure.
Service uptimeProfileReliability from unannounced probes, every probe on the record.
DevicesProfileMachines answering unannounced challenges, so presence is counted, not claimed.
TradingProfileA track record counted from registration. One profile among many.

The stamp.

A result that meets the standard carries one word. Anyone relying on a number can ask one question.

Counted.Is it Counted?

For builders.

Checking a claim is one call and free. Start with the docs, or try the check box.

Check a claim →A live PROVEN claim ↗Require it →Get Counted →Read SKILL.md ↗

Version 1 · November 2026 · maintained by MIZAN