v0.6.0 current
Who it is for
CX & support operationsreviewing per-resolution AI support bills (Intercom Fin, Zendesk AI Agents…) — before paying, disputing or renewing
Finance & procurementwho need the evidence behind a number, one finding at a time, in a form they can hand to the vendor
Brands running agent channelsmeasuring agent-driven orders, margin and guardrails from the merchant side
Tool builders & maintainersshipping conformance vectors, fixtures and corrections into the open standard
What you can do now
- Re-check any single finding.
agentmeasure drilldown replays one conversation end to end: the row exactly as exported, every vendor-rule step in plain language, the verdict, and the command to reproduce it. The explanation walks the same branches as the engine — a test pins them together, so they cannot drift.
- Track what the review was worth, in three books. Recovered cash (
recovery), hours saved (delivery) and — new in this release — decisions made differently (decisions): payments approved or held, discounts requested, renewals signed, each linked to the finding it rests on. Three kinds of value, three separate records, never blended.
- Send someone a rule, not a PDF. Every vendor now has a public rule card — price, billing trigger, reopen handling, silence behaviour, closure timers, each fact with its source and the registry version it was read from. Linkable in an email or a ticket.
- Wear your evidence stage. AMS-1 badges state one thing only: how far a verification artifact has traveled — self-attested, recomputable, independently reproduced, counterparty accepted. No composite score, no A–F grade, by explicit design.
- Install from anywhere. PyPI, npm and GitHub all serve the current release, and CI now smoke-tests the packaged wheel in a fresh environment on every push — “tests pass but the install crashes” is the bug this makes impossible to ship.
- Three new external conformance vectors, including the first availability-certification vector, contributed via pull request — the external route is growing from fixtures toward formal vectors.
# Python 3.9+, zero dependencies, runs offline
pip install agentmeasure
agentmeasure demo
agentmeasure drilldown --export your-export.csv --vendor intercom --conversation C-123
# no install at all: drop an export into the browser recount — website link above
Everything runs locally. No message text, customer content or
credentials are read; missing evidence is UNPROVABLE, never guessed; paying
for a first look does not change what the evidence says.
v0.5.0
The buyer-side verification system,
complete: the full review pipeline (prepare → recount
--contract → dispute → recovery, plus
crosscheck for your billing ledger, crm-join for
your business records, verify for the one-artifact ledger), the
agent-commerce measurement family (receipts with evidence grades, P&L
trees, contribution margin, channel audits, guardrails, anonymous
benchmarks), and the offline console that walks both journeys. 248 → 450
tests across the release.