Six interoperating components for verifiable agentic AI. This repo holds the suite-level reference architecture and specifications.
New here? Start with QUICK-TOUR.md — a 5-minute reading order across the three documents that show the architectural thesis.
The Evidence Suite is a set of six open-source components for building agentic AI systems whose execution can be independently verified. It targets architects and engineering teams shipping AI agents into regulated, audited, or high-stakes environments — where after-the-fact reasoning ("did this run actually do what it claims?") is a hard requirement, not a nice-to-have.
Implementation lives in the six component repos linked below. This repo (agentic-evidence) holds the cross-cutting documents: reference architecture, specifications index, standards interop, glossary, roadmap, and versioning policy.
| Component | Lane | Repo |
|---|---|---|
| Receipts | Cryptographic provenance evidence — signed, hash-chained records of what an agent did. | agentic-receipts |
| Policy engine | Deny-by-default enforcement at agent-action boundaries; emits decision receipts. | agentic-policy-engine |
| Eval harness | Bypass / injection / determinism scenarios with regression gates. | agentic-eval-harness |
| Artifacts | Manifest schemas and provenance rules for integrity-preserved agent outputs. | agentic-artifacts |
| Trace CLI | Create, sign, redact, verify, and export verifiable agent traces. | agentic-trace-cli |
| Evidence viewer | Drag-drop UI for inspecting traces, receipts, policy decisions, and artifacts. | agentic-evidence-viewer |
The full reference architecture is at REFERENCE-ARCHITECTURE.md. It defines the three trust boundaries, the six components, the data flow through a single agent run, and the architectural positions taken for the v0.1 release line.
Specifications live in their respective component repos. As they stabilize toward v0.1, this section will index them directly: the receipt schema and canonicalization rules, the policy decision-receipt format, the artifact manifest format, and the evaluation-scenario contract.
The full standards-mapping document is at INTEROP.md. It defines the integration shape for OpenTelemetry, SLSA, W3C Verifiable Credentials, the Model Context Protocol, and CloudEvents, with field-level mappings and a per-version compatibility matrix.
A long-form treatment of the design rationale — why receipts, why deny-by-default, why a separable evaluation harness — is available as PDF and HTML source. v1.0, April 2026.
Cross-suite vocabulary is defined in GLOSSARY.md. It distinguishes the terms most commonly conflated (receipt vs. log, trace vs. OTel trace, provenance vs. data lineage).
SemVer rules, breaking-change definitions, and deprecation policy are at VERSIONING.md. Companion to ROADMAP.md, which tracks live status; this document carries the rules that govern it.
The suite opens to external contribution at v0.1.0, once each component repo carries a tagged release, conformance vectors, and stable specifications. In the interim, issues are tracked publicly in each component repo's v0.1 milestone for visibility and design discussion.
MIT. See LICENSE.