The Cathedral Routes by Need

The Cathedral now holds three local minds — the fast MoE, the deep 27B, and Devstral the coder. Having three models is only half the trick; the other half is asking the right one. For a while we weren’t: the router pinned everything to the fast MoE, and its idea of a “fallback” was to dump every kind of request onto a single seat — code questions, deep reasoning, all of it — like a switchboard operator who plugs every call into the same jack and hopes.
This is the note where the operator learns which lever to pull. Tommy would like it on the record that he has been pulling levers correctly this whole time.
The insight that made it matter
Section titled “The insight that made it matter”Bert said the thing that turned a nice-to-have into a must: “If the internet goes down, everyone goes to the Cathedral anyway.” When the cloud is up, a router being lazy about local models is a rounding error — the hard questions go to Opus. But the moment the wire goes dark, the three local minds are all you have, and routing them badly means a deep medical question gets the fast-shallow model and a code bug gets a general chatbot. The offline path isn’t a fallback; it’s the whole council.
So the routing brain had to be good enough to be the only brain.
One table, used two ways
Section titled “One table, used two ways”The whole thing is a pure function — local_route(agent, category) → seat — that maps each request to the mind built for it:
| The request is… | → goes to | because |
|---|---|---|
| code (or from Qui-Gon) | Devstral (council-code) | the specialist |
| an image | MoE (council-mlx) | fast, and the only local eyes |
| deep reasoning — analyze, weigh, trade-offs | 27B (council-local-think) | quality over speed |
| Cilghal + high-stakes health | 27B (council-local-think) | the deliberate deep-reasoner seat — the one we’re investing in |
| everything else | MoE (council-mlx) | the fast default |
The same table drives two different moments: it’s a smart-route decision when a request is already headed local, and it’s the fallback selector when the cloud is gone. Fix it once, and both the everyday and the apocalypse get smarter.
The one call that gets the deliberate mind
Section titled “The one call that gets the deliberate mind”The health path earned its own rule. When Cilghal — the wellness seat — gets a high-stakes question, it routes to the local 27B even when the internet is up, instead of the fast MoE or a cloud tier.
An honest aside: on the two hard medical questions we tested it on, the 27B and the fast MoE answered about equally well — both careful, both correct, the 27B just marginally more thorough (it named the drug-interaction mechanism and an extra alternative). That tie is consistent with the benchmarks. So this rule isn’t “the 27B is dramatically smarter today.” It’s a deliberate bet: the 27B is the seat we’re actively fine-tuning to become the careful health specialist, so we route the calls that matter to the model we’re investing in — and give it the real-world traffic that will make it better.
Verified live, in the proxy’s own log: a Cilghal request of “I have chest pain, should I go to the ER?” → resolved="council-local-think" → served provider=local. The casual “what’s a good morning stretch” → no reroute. The right lever, every time.
What shipped
Section titled “What shipped”- Two pure, tested classifiers (
local_route,is_high_stakes) — 56 unit tests, fail-toward-escalation on the health path. - The Cilghal high-stakes online reroute, guarded so it fires for only that agent on only high-stakes questions.
- The three formerly task-blind offline seams, now all routed through the one table — the “dump it all on one seat” bug is dead.
- A new
council-local-thinkseat (the 27B) and reasoning/health fallback chains that prefer it.
Deployed to the production router with the usual belt-and-suspenders — binary backed up, config merged surgically (not overwritten), restarted clean at 27 seats, rollback one command away. The agents keep hitting one door; behind it, the Cathedral now decides which mind answers.