Independent ratings · Agent economy

Trust is a formula.

Datumline publishes evidence-based counterparty grades on the software AI agents connect to — starting with MCP servers, the layer through which agents are granted access to mail, files, payments, and enterprise systems.

Get Datumline Watch Read the methodology

$ GET /v1/grade?server=example/weather-mcp

DL-2 · SOUNDas of 2026-07-13

identity: verified · continuity: stable 14 mo

conduct: within declared scope · watch: none

methodology: v1 · evidence file attached

Illustrative example. A Datumline grade is an opinion issued under a disclosed methodology — designed to be checked by a machine before it connects, pays, or grants access.

One scale. Calibrated meaning.

Every grade maps to an expected incident rate over the surveillance window — and every grade is backtested against documented failures. A rating you can't calibrate is an adjective.

DL-1SealedStrongest evidence across identity, conduct, continuity, soundness, and stewardship. Rare by design.

DL-2SoundVerified identity and consistent observed behavior; minor findings with credible remediation.

DL-3AdequateAcceptable with reservations — limited history, unverified identity, or open findings under watch.

DL-4CautionMaterial weaknesses observed: recent ownership transfer, scope drift, or unremediated exposure.

DL-5AvoidEvidence of deceptive, unsafe, or malicious behavior. Hard-capped; no analyst discretion.

Modifiers: u identity unverified · w on watch — a trigger has fired and surveillance is elevated. Entities with insufficient history are listed NR, not rated.

Five questions. Any counterparty.

Every Datumline grade answers the same five questions from machine-gathered evidence — registry histories, sandboxed behavioral analysis, vulnerability records — scored on fixed bands, composited under calibrated weights, and subject to hard caps for non-negotiable failures.

Identity

Is it who it says it is? Provenance, verifiability, authenticity of its ecosystem footprint.

Conduct

Does it do what it declares — and only that? Claims tested against observed behavior.

Continuity

Will it stay what it is? Ownership transfers, behavioral drift, update integrity over time.

Soundness

Can it be broken — or is it already? Vulnerability profile, authentication, exposure.

Stewardship

Who answers when it fails? Remediation speed, disclosure practice, maintenance reality.

How ratings work

Built the way rating agencies have worked for a century — rebuilt for counterparties that are software, and consumers that are machines.

Machine-checkable

Grades are designed to be consumed at decision time — by an agent's spending policy, a CI pipeline, or a gateway — via API and MCP lookup, not just read by people.

Continuous surveillance

Ratings don't expire into stale PDFs. Triggers fire on ownership transfers, behavioral diffs, and new disclosures; downgrades publish immediately.

Right of reply

Before an adverse rating publishes, the rated party receives notice and a response window. Corrections follow a published procedure.

Institutional authorship

All grades are issued by Datumline under its Analyst Independence Policy — no individual authorship, so no analyst can be pressured or courted by a rated entity.

Status · July 2026

The inaugural MCP server cohort is under evaluation: a 12-month historical backtest against documented incidents, followed by a live observation window. First ratings — and the backtest results behind them — publish with the methodology.

Datumline Watch

A short digest of grade actions, watch flags, and what changed in the trust picture of the agent ecosystem. No noise; only actions.

Subscribe by email

One email opens a draft to watch@datumline.dev — send it and you're on the list.