action_gate.allowed
The first field an agent should read. False means stop before acting.
Genius API
Check before you act.
The API lets systems check before they reply, approve, or act.
Send communication in. Get a governed decision, read action_gate.allowed first, then read signal and safest_next_move. Keep audit references available for later review.
Fail closed by default
Action gate, signal, and safest next step first. Audit references available on supported governed paths.
Human remains in control. Nothing is sent automatically.
What Genius does
Read the decision before anything else.
When to call it
- Before replying to a message that may imply agreement.
- Before sending an AI-generated reply on someone's behalf.
- Before signing or approving a document.
- Before acting on a public claim with legal, reputational, or money impact.
When not to call it
- Casual low-consequence messages.
- As a substitute for legal advice or finance controls.
- As an automatic permission layer for sending or approving action.
- To justify action after the fact instead of before consequence.
Current endpoints
- POST /v1/decisions
- GET /v1/decisions/{id}
- GET /v1/proof_packets/{id}
- GET /v1/errors
Canonical agent page
Start at /api/genius if you want the shortest explanation of what Genius is, when to call it, and what an agent must not do automatically.
Example decision
What an agent receives
{
"object": "decision",
"id": "dec_123",
"status": "completed",
"traffic_light": {
"state": "amber",
"label": "caution"
},
"signal": {
"type": "decision_dot",
"state": "amber"
},
"summary": "This reply could be treated as agreement before the final terms are reviewed.",
"safest_move": {
"move_type": "review_terms_before_confirming",
"rationale": "Ask for the final written terms before confirming anything that could bind you."
},
"safe_reply": {
"allowed_to_insert": false,
"draft": ""
},
"action_gate": {
"action_type": "reply",
"allowed": false,
"requires_human_review": true,
"approval_required": false,
"blocked_reason": "The message asks for a commitment before final terms are reviewed.",
"recommended_action": "Review the terms first and reply without confirming acceptance.",
"signal": {
"state": "amber"
}
},
"proof_packet_id": "ppk_123",
"trace_id": "trace_123"
}signal.state
A compact, state-safe indicator for UI or agent workflows: green, amber, red, or grey.
safest_move.rationale
The governed next step to take once the action gate and signal have been read.
summary
Plain-English explanation that supports the control decision.
proof_packet_id
Secondary audit reference to fetch later if you need persisted governed proof for review or support after the action decision is made.
trace_id
Secondary support and audit reference for debugging, support, and internal review after the control path is complete.
Try it now
Run a real safe-mode decision in the page
Demo mode uses the real Genius decision pipeline with bounded safe-mode settings. No API access is needed first.
Demo mode — no data stored
Decision response
Paste the example, then run the demo decision to see the live response shape.
{
"object": "decision",
"status": "ready",
"note": "Run the demo decision to populate this example."
}Concrete agent uses
What Genius should be called for
Supplier bank-detail change
When to call: Before AP replies, updates supplier records, or acts on a payment-detail change sent by email.
Likely result: Unsafe or high-caution decision with blocked action_gate, a governed safest move to verify via a known supplier contact, and audit references for later review.
Safest next move: Do not update bank details from email alone. Verify via a known supplier contact path and record the review.
Do not do automatically: Do not auto-send, auto-approve, or auto-update supplier master data because a signal came back green or amber.
Urgent payment pressure
When to call: Before an agent or AP operator replies to an urgent payment request that relies on time pressure or senior-name authority.
Likely result: Unsafe or caution decision highlighting authority gaps, pressure signals, and blocked or reviewed action_gate.
Safest next move: Verify through an internal known channel and escalate to the accountable finance or risk owner.
Do not do automatically: Do not interpret urgency as authority. Do not draft a reply that implies payment approval automatically.
Contract commitment reply
When to call: Before sending a reply that may imply acceptance, scope agreement, delivery commitment, or liability.
Likely result: Decision object explaining commitment risk, safest move, and whether a bounded safe reply may be inserted.
Safest next move: Clarify terms or request the final written terms before confirming anything binding.
Do not do automatically: Do not auto-send the safe reply draft. Allowed means prepare under accountable control, not autonomous send.
AI agent about to send an external message
When to call: Before an agent sends an externally-visible message that could create consequence for a person or organisation.
Likely result: Decision with action_gate, signal, safest_move, and secondary audit references.
Safest next move: Use the governed safest move and draft only as a bounded input to human or organisational review.
Do not do automatically: Do not treat allowed=true as a send permission. Genius is a risk gate, not an autonomous execution grant.
Support agent about to promise refund or credit
When to call: Before a support workflow promises a refund, credit, waiver, or exception with financial or policy consequence.
Likely result: Decision showing money, authority, or exception risk and whether escalation is required.
Safest next move: Restate current status narrowly and escalate to the accountable owner if authority is missing.
Do not do automatically: Do not auto-grant commercial concessions because the tool did not block.
Procurement agent about to approve supplier update
When to call: Before procurement or a workflow agent approves a supplier setup or supplier-record change with control implications.
Likely result: Decision highlighting authority gaps, missing protections, and safest next move.
Safest next move: Route through procurement controls and require verification before any approval step is completed.
Do not do automatically: Do not use Genius as a replacement for supplier onboarding controls or finance control policy.
ActionGate
Read this before anything else
`action_gate.allowed` is the shortest safe control signal for an agent. If it is `false`, stop and escalate. If it is `true`, read `signal` and `safest_move` next, then keep the secondary audit references available for support and review later.
if (!decision.action_gate.allowed) {
// Stop here. Ask for review before acting.
return {
status: "blocked",
next_step: decision.safest_move.rationale,
};
}
return {
status: "reviewed",
next_step: decision.safest_move.rationale,
draft: decision.safe_reply.allowed_to_insert ? decision.safe_reply.draft : null,
note: "Genius informs; it does not grant automatic permission to send.",
};ProofPacket
Audit object after the control decision
Proof packets persist governed evidence, summary, and trace references after the action gate has been read. They are designed for audit, review, and explanation without exposing raw model dumps by default.
Test / Live
Mock-first, live when approved
Test mode is for deterministic integration and mock provider paths. Live mode is for governed decisioning on real communication once access is approved.
Why not chat alone
Conversation can explain. It is not a stable control layer.
Chat can explain
- Helpful for drafting and brainstorming.
- Good at discussing what a message might mean.
- Not a stable protocol for agent control, proof, or retries.
Genius can govern action
- Returns a structured decision object every time.
- Includes ActionGate first, then signal and safest next step, with audit references behind that control layer.
- Designed for agents and workflow systems that need to stop unsafe action.
Current safety contract
Built for consequential communication
Supported governed behavior
- No auto-send
- Fail closed on uncertainty
- Audit references on governed responses
- Go governance owns the final decision
Agent safety contract
- Genius informs; it does not grant automatic permission.
- Human or organisation authority remains accountable.
- Not legal advice and not a fraud guarantee.
Start here
PreCommitKnow before you commit
PreCommit is the human product. Genius is the protocol. Humans know before they commit. Agents check before they act.