What shipped

autopoet.app live

The front door: the paper, full documentation, and the living graph.

• The paper — the complete definition and scope, published as the launch crosshair.
• Docs covering arming, ceilings, postures, requests, leases, eval, knowledge, and the learning system.
• Desktop builds for macOS, Windows, and Linux staged for release.

The nervous system validated

The learning layer, proven end to end on deterministic seeds and production paths — no gradients, no GPU.

• Hebbian plasticity on the workspace graph: next-access prediction 0.783 vs 0.757 baseline; recovery after drift in 500 events vs 622.
• Surprise-gated cognition: quality 1.000 at 1.3% of always-model cost; attention spikes 4.7× at regime switches and self-quenches.
• The economy: 0.970 on the canonical benchmark from outcome-only credit; two layers invent a communication protocol under a 2-bit channel (0.947).
• The gym: 8,076 events through real parse, hook, and event-dispatch paths — 100% effect settlement, p50 3µs.

The constitution live

Autopoiesis v2 — the immune system, shipped and running in production.

• Ceiling-in-index with intersection-down effective grants, re-resolved every agent turn.
• The human-gated structural triad: grant, ceiling, management.
• Management postures — managed, proposed, frozen — per agent and per subtree.
• Leases: scoped, expiring, non-delegable capability grants.
• Eval-in-scratch with parse, purity, and authority gates.
• Typed self-edit requests with the injection firewall.
• The knowledge log, the access panel, plans and handoffs, and the heartbeat worker itself.

First beat

The autopoet takes its first heartbeat on a production workspace: senses two telemetry concerns and a filed request, proposes, evaluates in scratch, escalates the grant change to a human, merges the rest, and writes four lessons. The loop closes.