Change control for AI systems
Ship AI. Know what changed.
Your agent does the engineering. Epok reads the source where you connect a repository, ties the change to the requirements it touches, and keeps the decision of the person who is accountable.
One change, from agent to decision
Step through what Epok records, and what it refuses to assume.
Agent
Reads the system
- Document classifier ยท accepted state 4
- classifier model-a
- training-set v4
- Authority to decide: none
The agent calls epok.get_system_context. It sees the accepted state, never a way to decide.
Five ways of knowing, kept apart
Most records blur what was claimed with what was checked. Epok labels every item with how it became known.
- Agent reported
The agent said so. Kept, but never counted as proof.
- Attached
A record arrived. Nobody has checked it.
- Integrity checked
A digest matches the bytes supplied. The bytes are proven; the claims inside are not.
- Observed in source
Epok read it from the repository itself.
- Human confirmed
A named reviewer confirmed it.
Connect in minutes
Epok speaks MCP. Create a scoped credential, paste one line into Claude Code, Codex or any MCP client, and test it from the browser. Revoke it any time.
Create a workspaceclaude mcp add --transport http epok \
https://your-workspace.epok.work/mcp/ \
--header 'Authorization: Bearer ${EPOK_AGENT_TOKEN}'What Epok does not do
- It does not certify a system, approve it, or declare it compliant. People decide; Epok keeps the record.
- It does not treat a passing test or a reported metric as validation.
- It does not assume a regulation applies. Where no classification is recorded, applicability is unknown.
- It gives agents no way to record a decision.