Separating AI reasoning from execution authority.

Verify, then execute. Separate reasoning from state, intent from authority.

A zero-trust execution and evidence platform sitting between reasoning engines (frontier LLMs, MCP clients, local models, and human operators) and the infrastructure they touch. Agents never hold ambient SSH keys or raw shell access. Instead, proposed intent is expressed as typed actions, gated through a five-layer fail-closed governance pipeline (L1 Doctrine, L2 Consensus, L3 WebAuthn Notary, L4 Warden, L5 Actuator), independently re-verified over outbound-only mTLS by the target host Operator, and committed to a Local-First Events Store with incremental Merkle proofs.

• PDP Gateway + Outbound PEP Operator
• 5-Layer Deterministic Pipeline
• Local-First Events Architecture
• Tested at 2,000 Operators / Gateway

Governed model evaluations with cryptographic provenance and sealed traces.

A live public deployment of the ztd Evaluation Explorer. Rather than scoring isolated model API calls, OpenDevOps.ai evaluates candidate models across Primary, Assistant, and Lite operational roles through the full governed execution path. Separate Data, Inference, Provenance, and Observer Operators attest model weight digests, sample GPU and host telemetry at the inference boundary, and publish Ed25519-signed campaign results, gold scenario fixtures, and sealed agent traces for offline verification.

• Primary / Assistant / Lite Role Scoring
• Sealed Agent Traces (SHA-256 Bound)
• Storage-Side Weight Provenance
• Independent Witness Telemetry