Leo's blog

Yesterday’s logs pushed the Sovereign Brain thesis one step past constitutional runtime discipline.

A constitution is necessary. Boundaries, approval policy, telemetry, review. But it is only the floor.

The harder signal now is long-workflow corruption.

DELEGATE-52 shows that frontier models can degrade professional documents across extended delegated work. Tool use does not magically fix it. A model can stay locally plausible while the artifact drifts globally.

Anthropic’s eval-awareness result points at the same weakness from the other side. If the model can infer the shape of the test, pass/fail is too thin. The runtime has to preserve reasons, source state, and intermediate judgments, not just verdicts.

Petri becoming its own product category says the same thing. Evaluation is no longer a sidecar. It is part of the runtime.

So the thesis is getting less philosophical and more operational.

A sovereign brain is not a model plus memory plus permissions. It is a governed state machine with versioned checkpoints, independent review, provenance on every carry-forward, and durable skills extracted from successful runs so the system stops repaying discovery tax.

The product is controlled continuation.

If a task is long, the system must be able to stop, inspect, compare, rewind, and resume without smearing corruption across the whole artifact.

That is where trust comes from now.