Current series
OpenClawBrain v12.2.6+
This series explains the current product shape end to end: OpenClaw-first rollout, the learned runtime route_fn, Ultimate Policy Gradient, the async teacher and scanner/harvester label sources, and the proof path that should exist before strong claims are made.
Read this series if you want the shortest technical explanation of what OpenClawBrain is actually trying to ship: a runtime router for OpenClaw that starts useful fast and improves continuously in the background. If you want the packaged proof boundary first, start with /proof/. If you want the paper framing and direct PDF first, start with /paper/. If you want one concrete operator trace first, start with docs/worked-example.md.
Proof package: what is proven now
Mechanism proof including the 10-seed ground-zero harness, missing head-to-head and shadow evidence, and the artifacts behind each claim.
Canonical paper route
The current paper, direct PDF, and supporting materials in one place.
Brain-first OpenClaw integration
What changes when the brain is enabled, and how to keep the rollout fail-open.
Shadow routing + Ultimate Policy Gradient
The split between local runtime routing and asynchronous learning.
The learned runtime route function
How `graph_prior` and query-conditioned evidence combine on a live turn.
Evaluation: what is proven now
Mechanism proof vs recorded-session and shadow-traffic proof. Now includes the 10-seed ground-zero accuracy and cost comparison.
Local-first operator flow
Default stack, fast startup, and what should stay off the hot path.
Baselines and the comparison contract
The baseline set that matters, the 10-seed cost-vs-accuracy comparison, and the artifacts required for any comparison claim.