AgentMeasure

# AgentMeasure

> Open measurement standard for AI agent usage and outcome claims. It defines what
> counts as one operation, what a result may be called, and what evidence supports
> what strength of claim. Verdicts are PASS / FAIL / UNPROVABLE.
> Local-first. No network code in the tool. MIT.
>
> Last reviewed: 2026-09-19

## What this standard is

AgentMeasure is the **semantics layer** for agent measurement. It says what a claim
*means*. It does not judge whether a record is authentic (that is a signing layer)
and does not judge whether money is owed (that is a billing-reconciliation layer).

Three integrity layers are orthogonal and composable:

- tally integrity (is the amount right) — billing reconciliation tools
- record integrity (is the record authentic) — signing / transparency layers
- **classification integrity (what may the outcome be called) — AgentMeasure**

## Canonical documents

- [standard/CORE.md](standard/CORE.md): objects, grains, observability states, invariants
- [standard/METRICS.md](standard/METRICS.md): metric contracts M1–M6
- [standard/SETTLEMENT.md](standard/SETTLEMENT.md): AMS-1 settlement statement clauses S-1…S-8
- [standard/QUALITY.md](standard/QUALITY.md): quality model, use profiles, claim discipline
- [standard/TRUST.md](standard/TRUST.md): evidence profile axes, including causality V0–V4
- [standard/scope.md](standard/scope.md): what this standard is not
- [standard/hard-questions.md](standard/hard-questions.md): answers to the hostile questions
- [standard/known-limits-unprovable-by-surface.md](standard/known-limits-unprovable-by-surface.md): expected UNPROVABLE rates by topology
- [extensions/COMMERCIAL.md](extensions/COMMERCIAL.md): economic semantics (experimental)

## Machine-readable single sources of truth

- [registry/vocabularies.yaml](registry/vocabularies.yaml): every enum, generated into schema / TS / Python
- [registry/metrics.yaml](registry/metrics.yaml): every metric contract
- [registry/billable-unit.yaml](registry/billable-unit.yaml): billable units
- [schemas/payloads/](schemas/payloads/): payload schemas per observation type

## Verification

- [conformance/](conformance/): vectors, runners, evidence cases
- [conformance/README.md](conformance/README.md): check families mapped to invariants
- Run: `python3 conformance/runners/run_metrics.py`, `run_outcome_audit.py`, `run_delegation.py`

## Tooling

- `pipx run agentmeasure check` — local transcript check (Codex, Claude Code)
- `pipx run agentmeasure check --audit` — adds operation-resolution, cache, token checks
- `pipx run agentmeasure conformance --fixture events.jsonl` — fixture conformance
- `pipx run agentmeasure settle --effects effects.jsonl --statement --price 0.99` — negotiable statement

## Boundaries (do not treat this as a substitute)

- Not an observability platform — use OpenTelemetry, Langfuse, Arize
- Not a signing or transparency layer — use PEAC, Glacis, IETF SCITT
- Not a billing-amount reconciliation tool — use a billing audit layer
- Not a quality scorer — it does not judge whether an answer was good
- Not a payment rail — measurement never sits in the payment critical path

## Optional

- [proposals/](proposals/): standards-change proposals (AUP)
- [CI-DISCIPLINE.md](CI-DISCIPLINE.md): tests-before-push rule