EVE AI Core
The Infrastructure of No.
Every agent action you route through EVE is policy-checked, tenant-bound, hash-chained, and recorded in a signed Decision Certificate — verifiable offline with a public key. No secret. No “trust our logs.”
Below are real, live-signed agent Decision Certificates from this server — one per outcome, hash-chained together. Click Run all checks and your browser verifies each signature against the published ECDSA P-384 public key — locally, with no secret and no call back to EVE. Then click Tamper and watch verification fail.
Records the tier, approval mode, charter compliance, and any violated principles. The action passed a non-overridable charter veto, cognitive locks, and a tiered policy before it ran.
The org_id is part of the signed payload and the certificate lives in a per-tenant ledger. A certificate cannot be silently moved between customers.
Each certificate’s prev_hash is the previous one’s content hash. Deleting or reordering any record breaks the chain — and it’s detectable.
ECDSA P-384 signatures verify with the public key at /.well-known/eve-pubkey. Your auditor or regulator verifies independently — no secret, no EVE involvement.
Whether an action runs, is modified, waits for a human, is blocked, or is rejected — each produces a signed, chained certificate.
/api/autonomous/*) or the in-process runtime. Tier 4 and charter HARD_BLOCKs never execute; charter compliance is re-checked at human approval. Drop-in adapters for unmodified third-party frameworks (LangGraph, CrewAI) are on the roadmap (Q3 2026). EVE governs and certifies the decision to act and binds the action’s arguments; your executor performs the effect.
EVE CoreGuard enforces single decisions. EVE Agent Infrastructure governs agent actions on that same engine. EVE Proof makes every outcome independently verifiable. Same charter, same signing key, same certificate format. Enforces against ECOA / Reg B, FCRA, HIPAA, and the EU AI Act; maps to and evidences SR 26-2 and NIST AI RMF. SOC 2 readiness program underway. You control which decision fields are sent; in VPC and on-prem deployments, governed content and evidence stay inside your boundary.
The certificates above prove an action was authorized. The layer we are building goes further: an agent holds no standing credential — a usable, single-use, parameter-bound authorization (and, for a real executor, a short-lived least-privilege credential) comes into existence only after EVE approves the exact action. It is authored and tested in isolation on the feature/proof-bound-agent-execution branch: a delegated authority graph, atomic single-use consumption, identity/workload attestation, a 2-of-2 witness quorum, and an open eve.action.authorization.v1 protocol with a cross-language (Python + JavaScript) conformance verifier. A $50,000-wire acceptance test runs the whole chain end-to-end with every mutation rejected.
EVE_AGENT_AUTHORITY_GATEWAY and is not wired into the running server. Its reference connectors and credential providers are safe simulations: no real money moves and no real cloud credentials are issued. It is shown so the public record matches the code, not ahead of it. Full detail on EVE Agent Infrastructure.
In 60 days, every action in it is governable, certifiable, and audit-ready — with certificates your auditors verify themselves.