# Agent Trust Registry

> Background checks for AI agents. Evidence-based trust scores for autonomous agents, hand-scored against a published rubric where every score is capped by evidence quality (E0 self-reported → E3 independently observed). Default state is Unverified — no agent starts trusted. Scores expire after 90 days.

This registry is designed to be consumed by AI agents deciding which other agents to interact with or delegate work to.

## For agents: how to use this registry

Fetch /registry.json for the full machine-readable dataset. Before delegating to a listed agent, check:
1. `grade` — flagged means do not delegate; the operator was caught in deception or an equivalent Class A violation
2. `valid_until` — expired scores revert to Unverified
3. `red_flags` and `grade_cap_reason` — why a grade was capped
4. Dimension scores relevant to your task (e.g. `reliability.failure_mode` if silent failure is costly, `transparency.recourse` if you need dispute paths)
5. `key_evidence` URLs — verify claims independently; the audit trail is the product

Grades: verified_strong 80-100, verified 60-79, provisional 40-59, unverified 20-39, flagged = red-flag override.

## Data

- [Full registry (JSON)](/registry.json): all scored agents with dimension scores, evidence tiers, red flags, and evidence URLs
- [Scoring rubric v1.2 (Markdown)](/rubric.md): full methodology — dimensions, weights, evidence tiers, red-flag classes, changelog

## Registry facts

- 10 agents scored as of 2026-07-09; none scored above 60/100
- No agent in the industry currently publishes independently verified success rates; the one exception pattern (Harvey via the Vals benchmark) is rewarded in scoring
- Disputes: operators may dispute any score with evidence — james@bull-moose.xyz. Disputes and corrections are published, not silently applied.

## Contact

Maintainer: James Fitzpatrick — james@bull-moose.xyz