Five Spartan ephors held authority over even the kings. Ephor puts that constitutional oversight layer inside your AI agent infrastructure — policy gates, tamper-evident audit chains, and Human-in-the-Loop controls enforced at the point of execution.
Every capability is embedded in the agent runtime via WASM components. No sidecar agents. No post-hoc audits. Compliance happens where the decision does.
Every agent action passes through a WASM policy evaluator before execution. Define rules in composable policy-as-code; any violation triggers an instant hold — not a retrospective flag. Ship with confidence that your agent cannot violate its own policy envelope.
WIT · policy.evaluateEvery captured action is linked into a SHA-256 hash chain. Append-only, verifiable from any entry. Your compliance team gets an immutable record of every decision the AI made — not logs you could edit, but a chain you can prove.
SHA-256 · append-only · chain-validHigh-risk actions are automatically held pending human approval. Reviewers decide via browser console, Pharo IDE widget, or any HTTP client. SLA timers enforce safe-default behaviour on timeout. Article 14 of the EU AI Act requires this for high-risk systems.
Art 14 · HITL · SLA-awareGenerate a structured evidence bundle for any time window: audit entries, policy snapshots, reviewer decisions, and chain integrity proof. Hand it to your conformity assessment body without bespoke tooling. Your GPAI system documentation is always current.
Art 11 · Art 47 · JSON + PDFConfigurable PII scrubbing runs at capture time, before the entry enters the chain. Define which fields are sensitive; Ephor redacts them in the canonical record while keeping the audit structure intact. Data governance and audit governance in one pipeline.
Art 10 · GDPR-aligned · configurableReal-time visibility into agent risk levels, pending approvals, SLA burn-down, and chain integrity. Know your risk posture at a glance — not from a weekly report produced three weeks after the incident.
Art 9 · Art 72 · live · SLAEvery agent action passes through Ephor's five-stage pipeline. Like the five ephors of Sparta, each stage holds authority over the next.
"Compliance should be runtime-verifiable, not audit-retrospective. Ephor makes EU AI Act adherence a property of the system — not of the spreadsheet that documents it."
Every Article maps to a concrete, implemented Ephor capability — not a policy intention.
| Article | Requirement | Ephor Feature | Status |
|---|---|---|---|
| Art 9 | Risk management system — continuous identification and mitigation throughout lifecycle | Policy gate + risk-level capture on every action; real-time dashboard; configurable hold thresholds | Implemented |
| Art 10 | Data governance — data quality, minimisation, and documentation for training and operation | PII scrubbing at capture time; data-minimisation config per agent class; field-level redaction | Implemented |
| Art 11 | Technical documentation — comprehensive system documentation maintained and current | Evidence pack export: policy snapshots, audit entries, reviewer decisions, chain proof | Implemented |
| Art 12 | Record-keeping — automatic logging of operations for post-market monitoring | Tamper-evident SHA-256 hash chain; append-only; chain-valid verification on every query | Implemented |
| Art 13 | Transparency — users must be informed they are interacting with an AI system | Agent-class identification stamped on every captured action; disclosure fields in audit record | Implemented |
| Art 14 | Human oversight — high-risk systems must allow humans to intervene and override | HITL hold → reviewer approval/denial → SLA timeout safe-default; browser console + Pharo widget + HTTP API | Implemented |
| Art 47 | Declaration of conformity — providers must draw up and maintain conformity documentation | Structured evidence pack exportable on demand; maps to GPAI conformity assessment requirements | Implemented |
| Art 72 | Post-market monitoring — continuous monitoring of high-risk AI systems in production | Real-time risk dashboard; SLA burn-down alerts; chain integrity monitoring; reviewer console | Implemented |
Ephor runs as a sidecar or embedded WASM component. Any HTTP client, MCP host, or Pharo/Smalltalk image integrates in minutes.
# 1. Start the governance node docker compose -f demo/docker-compose.yml up --build # 2. Capture an agent action curl -X POST http://localhost:9800/capture \ -H 'Content-Type: application/json' \ -d '{{"session_id":"s1","agent_class":"PaymentAgent", "action":"payment.send","arguments":["amount=50000"]}}' # 3. Request human approval (Art 14 hold, 5-min SLA) curl -X POST http://localhost:9800/oversight/request-approval \ -d '{{"entry_id":"...","risk_level":"high","sla_ms":300000}}' # 4. Reviewer decides via browser console — http://localhost:9200 # 5. Agent polls until: approved / denied / timeout → safe-default
Ephor is in limited early access. We're working with teams building high-risk AI systems under the EU AI Act — payment agents, medical decision aids, HR screening, credit scoring.