Planisphere — the evidence API for AI actions

Planisphere is the evidence layer around your AI's consequential acts: gate holds the act until a human signs off, the sign-off seals into a tamper-evident receipt, and anyone can verify that receipt forever — offline, without trusting you or us.

Gate an action right now against the live API:

curl -sS -X POST https://api.planisphere.ooo/v1/gate \
  -H "X-Planisphere-Key: ps_test_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"pack":"law","proposed_action":"File this motion with the court."}'

You get a decision, an action_key, and a sealed evidence_packet you can hand to any auditor.

Prompts instruct. Gates decide.

Prompt engineering tells a model how it should behave. A gate controls what actually executes.

Every agent runs a loop — read context, evaluate, act. Planisphere hooks that loop at the one step with consequences: the moment a proposed action is about to touch the world. Before your workflow files, sends, exports, deploys, or deletes, it posts the proposed_action to /v1/gate. The decision comes back from code running at the execution boundary, not from hoping the system prompt held.

• Gate — execution control. Policy evaluates the act deterministically: allow proceeds, block stops, and needs_review pauses the action and routes it to the responsible human — a resumable object with its evidence attached, not a dead end.
• Seal — the record of what happened. The decision, the review, and the policy references seal into a tamper-evident receipt. Raw prompt and document content stays out of durable evidence by default.
• Verify — no trust required. Anyone can check a receipt's seal offline, without access to your infrastructure or ours.

Why not just write a stricter system prompt?

	System prompt (probabilistic)	Runtime gate (deterministic)
Enforcement	model attention weights	code at the side-effect boundary
Cost per rule	more tokens in every call	zero prompt tokens — policy lives server-side in the pack and your tenant overlay
Prompt injection	injected text can rewrite the instructions	injected text can change what the agent proposes, not what is allowed to execute
Audit trail	a transcript you assert is complete	a sealed receipt anyone can verify offline

An agent that has been talked into "delete the production records" can still propose it. With the gate at the boundary, that proposal pauses at needs_review and waits for a named human — the model's context has no vote in the decision.

Where to go

• Quickstart — five minutes to a verified evidence record, key to offline verify.
• Integration guides — SDKs, middleware, agent frameworks, review channels, webhooks.
• API reference — the full interactive Swagger surface; the schema itself is at /openapi.json.

No key yet? Get a free test key — work email + organization, no card, 100 sealed actions.