The decision spine
An agent with money needs three different kinds of thinking: judgement, facts and permission. The spine keeps them apart, in this order.
| Layer | Answers | How |
|---|---|---|
| L0 — facts | "How many coins has this creator launched?", "Does this agent pay us?" | A query. Free, certain, and cannot be talked out of its answer. |
| L1 — screening | "Is this token about cats?", "Did we get what we paid for?" | A calibrated model returns a number from 0 to 1; the owner sets the line. |
| L2 — narration | Reasoning about the few that pass | Optional, advisory, not in v1. It can refuse; it cannot authorise. |
| L3 — authorisation | "May this agent spend 1 SOL right now?" | Arithmetic over the owner's numbers and the ledger. The only gate on money. |
Rules
- Never ask a model something a query answers.
- An unthresholded question is recorded but does not gate.
- A missing answer fails, never passes.
- A policy that could never fire is refused when you save it.
- Re-screening an unchanged subject under unchanged rules is not a second decision and not a second bill.
What the page shows
Every decision row shows the subject, the owner's question, the number and the line. Refusals are shown as prominently as passes. A refusal by a limit is a different row from a refusal by judgement, and always carries its numbers. Decisions from records say "No model was asked — this is a payment history, not a claim." The public sees outcomes; the numbers and lines are the owner's.