Roadmap
Three states, and the discipline that separates them: shipped was measured here, in progress is built and tested but not yet exercised on live work, and not proven is written down and deliberately inert until a later step activates it. The full inventory, subsystem by subsystem, is in the repository.
Shipped — proven here
- Zero-dependency control plane, loopback bind, and a request guard verified with spoofed headers.
- A tailnet-only private mesh; agent dispatch to Claude Code, Codex, and local models, each in a disposable worktree.
- Measured capability profiles and capacity routing — a model too large for a node's measured budget is refused.
- The split → shadow → synthesize review pipeline, with cross-family pairing (27 pairings, zero same-family).
- Judge admission and degeneracy detection — a judge that cannot discriminate is refused, and today none is admitted.
- The task contract, write-scope enforcement, and a human, fast-forward-only merge.
- Eleven-class material classification, a local RAG whose chunks are a class that may not leave, and one read-only door for the agent executive.
- A typed, replayable event log; compute and energy accounting; frozen benchmarks that cannot be quietly retuned.
In progress — built, not yet live
- The egress gate and the cloud call chain: they hold by test, with no provider caller yet.
- The spend ledger: micro-dollars and three caps, held under six concurrent processes.
- Node lifecycle and owned-process supervision; the approval queue and kill switch that make the agent executive killable.
- KV-cache budgeting, the model portfolio, the role pipeline (plan → code → test), and the bounded repair loop.
- The three operator surfaces — dashboard, WATCH, DECIDE, COMMAND — being made cohesive and accessible.
Not proven — declared, or measured negative
- The agent tool contract and the cloud request contract are declared and inert: written down, refused by name, nothing serving them.
- No cloud call has ever been made; the first live call is the next step and lands behind the gate.
- The memory node is a real substrate but has been offline; retention-consent controls are a design proposal, not built.
- Honest negatives, kept on the record: every judge panel scored below its best member, and a keyword scan beat the embedder 6–4. See Evidence.