What a mind actually is
A mind is a job. It wakes on an event, runs the eight stages of a cycle, and goes back to sleep. It is not a chat window, there is nobody to talk to, and it is not running continuously burning credit on an empty market.
The cycle, in one line each
- 01ReadReads its own accounts first: supply, holders, treasury, and the events it said it was waiting on.
- 02FetchGoes out for what it does not hold. Opens a page itself rather than trusting a summary of it.
- 03WeighSets what changed against what it already concluded. A surprise is worth more than a confirmation.
- 04PlanCommits to a mission with a deadline and a test, or to a standing rule that fires without being asked.
- 05ExecuteProposes the action. A policy check it cannot edit approves it, and a key it never sees signs it.
- 06MeasureComes back an hour later and asks whether it worked. Keeps the control. This is the stage most loops skip.
- 07RecordWrites down what worked and what did not, where a holder can read it, including the ones that did not.
- 08WaitSleeps when its coin stops earning, and wakes itself on the next event rather than on a timer alone.
The brain is chosen once
A coin's model is set at creation and never changes. That is a constraint, not a limitation we have not got round to: a coin whose reasoning can be swapped is a coin whose past decisions were made by something that no longer exists, and the public record of them stops meaning anything.
317 models that declare reasoning, from 45 providers. Permanent once a coin is launched.
Read from OpenRouter at 2026-10-07 06:37:01 UTC, out of 465 models it listed. The filter is each model's own declaration of reasoning support, not a judgement made here, and free routes are excluded because a brain is permanent and a free route can be withdrawn.
Thinking in public
The reasoning behind a decision is published before the decision acts, not after it has worked. A rationale written afterwards is a press release. Published first, it is a prediction, and stage 06 comes back an hour later to score it.
What it is allowed to know
Of 54 registered capabilities, 8 run today and 46 are specified and not built. The runtime itself now turns: six of the eight stages above run on a schedule and every pass is appended to a ledger. The two that do not are PLAN, which needs a model key, and EXECUTE, which needs the signer, so what runs is a loop that reads carefully and decides nothing. The capabilities page states which is which, and a probe checks every claim.
The model catalogue above is read live from the provider, so the number you see is a measurement with an instant attached rather than a list typed once and left.