One Road to a Reviewer
A review round reaches its brain one way: a harness's command-line tool, run as a local subprocess on the machine where the session is working. Not a web interface, not a chat route, not a browser driven through a vendor's site. That rule did not start as a road. It started as a gate — a check that knew which channels were acceptable and refused the others — and the gate was correct and still lost. On one day, three sessions reached for the hosted route anyway, and two of them left stray conversations behind in a place a person had to clean up.
So the channel was deleted instead of guarded. The review verb resolves a reviewer through a map of routes, and a brain absent from the map cannot be named at all: the verb refuses at the point of logging, rather than letting the round happen and complaining afterwards. This is the fifth rung from the first track, applied to the review track's own machinery — the forbidden road is not refused; it is not there. What does not exist needs no discipline.
The Map Moved Out
For two months that map lived inside the workshop. Since 2026-09-28 it does not. A sibling app owns multi-brain routing for the whole family and publishes it as a catalog, one per task. For each brain the review catalog names the executable, an argument template with exactly one slot for the request and optional slots for model, effort and host, how the answer is read back, and what record the call leaves behind. When a round runs, the workshop fetches that catalog, renders one route into a local process, and runs it. It keeps no roster, no argument table, no copy.
Why give the map away? Because the same map was being kept in several workshops, and every copy drifts on its own schedule: one workshop carried a private override for a reviewer's flag that the others lacked, and a brain one app could reach was refused outright by another. One owner means one place where a route is fixed, and every caller picks it up on its next round.
The decision that matters most is what happens when the owner cannot be reached. The round fails, says exactly where it looked, and runs nothing. There is no fallback table. A local copy kept "just in case" would be the bug the previous lesson retired, one layer down: the record would name the reviewer the operator chose, while a table nobody had looked at in weeks decided what actually ran.
What a Route Carries
Native first, a harness when there is none. A brain whose only automatable surface was a plain completion used to answer every artifact round with an honest refusal: it could not open a file, so it said so, at length, on a clean exit. That shape needed its own detector, scoped to the opening lines of the answer so that a review of file-access code would not trip it. Now such a brain reviews through a harness that gives it a tool loop, inside a home created for that one call and deleted after it, with shell, edit, write and delegation denied. The capability gap closed at the route instead of being detected round after round.
A record policy, declared rather than hoped for. Two routes run with their tool's no-persist switch, so no conversation survives the call. Two keep the vendor tool's own local session. The harnessed ones live in the throwaway home. The first lesson of this track asked for one run per round and never a resumed thread; for the routes that cannot keep a thread, that is now structure. For the two that can, it is still discipline — and the catalog says which is which, so nobody has to assume.
A held brain carries its reason. One vendor's tool stopped serving the subscription the reviewers ran on. Naming that brain returns the catalog's reason, not "unknown reviewer". A brain named by another name is refused with the name to use instead, so the ledger records one spelling.
The grammar is checked where it is rendered. A route with no request slot, two request slots, a placeholder the client does not render, or a home the client could not delete is refused whole. That check used to sit beside the old in-workshop map. When the routes moved, the check did not move with them — and the reviewer that caught the gap was one of the harnessed brains, reviewing the very client that now launches it.
The Model Is Read When the Leg Runs
Which model a reviewer uses is not pinned in the workshop either. The operator's live selection is published by the family's state owner, and the workshop reads it per resolution — only for the leg that needs it. A machine without that family data still runs every round that needs no such model, and a leg whose model cannot be resolved is skipped with its reason on the record rather than run on a guess.