Day 61 — 'shrink the weld until it fits your garage.'

“shrink the weld until it fits your garage.”

Some of the best days start in the car. This one started with a question on the way to church — if we can hand-roll an MMM brain, why can’t we design our own LLM? — and by the end it was runnable code with a learning loop wrapped around it.

What’s actually in a frontier model

We took it apart. A frontier model is four things fused into one blob of weights: knowledge (facts you could look up), skills (learned procedures), reasoning (the general substrate that generalizes over novel input), and alignment (tone, refusals, helpfulness). They’re entangled, not layered — you can’t unplug the knowledge and keep the reasoning. Wrapped around that blob is a harness (tokenizer, sampling, tools, the agentic loop) and, separate again, an eval.

Then the punchline, which is the whole werbos thesis in one line: werbos already owns every part of that anatomy except the reasoning weld. Knowledge → the library. Skills → the router. Harness → the framework. Eval → the bench. What’s left — the one thing only a datacenter-scale pretraining run produces — is the weld. So the dream of “our own model, no open weights” isn’t about out-pretraining the labs (a race you’d lose to free). It’s about minimizing the weld — externalizing everything else until the reader you still need is small enough to distill once and own forever. Not the fantasy. The reachable version.

The brain is a wrapper; the siblings are organs

If the intelligence lives in the composition and not the weld, then the brain is a thin wrapper and the sibling primitives are its organs — knowledge hub, skill place, compute sector, memory — with partyparrot as the nervous system carrying signals between them. A room becomes an insulated, sealed packet: content-addressed, signed, fail-closed. A silo is not a safe — the room scopes routing, PFM provides the lock. We wrote the whole vision down (docs/brain-nervous-system.md) and the plan to close it (docs/honest-loop-plan.md), because with a detailed plan the next ahas form on their own.

Then we stopped talking and built it

The runtime brain, minimal and end to end. A library registry holds many content-addressed MMMs. A front intent weld routes a query to the one it’s most characteristic of — or abstains at the door. The unchanged fenced reader grounds it. A back render weld phrases the facts fluently but fenced — a groundedness check drops any embellishment and falls back to the safe cited render. Blend the style, never the facts. In front of all of it, a reflex layer: cached turns skip the cortex entirely, and the matcher misses (falls through) whenever it isn’t sure. Every layer abstains rather than guess — front on intent, reader on evidence, render on embellishment, reflex on doubt.

Then the organ the whole thing was missing: the Honest Loop. A frontier model is a frozen snapshot of an enormous feedback loop; we’d externalized every static part of it but never the loop itself. So: use → record provenance → judge (the operator is the un-bribable grader) → the miss becomes a row in a living gold set → re-bench → gate. The gate is the point — a change promotes only if lies don’t rise and the composite doesn’t fall, two independent guards. A lie vetoes a promotion even when accuracy jumps. Self- improvement without a gate tears down.

The loop’s first real turn

Then the best part — the loop caught its own first bug. We ran the pipeline live on the real 2342-chunk client corpus (26,947 fact leaves in a second and a half) plus two decoy libraries, and routing sent “how far away is the moon?” to the law library, and an off-topic “price of gold” to the law library too. The unit tests never saw it, because their fixtures were all the same size — but a huge corpus contains almost every common word, so generic terms outvoted the query’s real anchor.

We fixed it the honest way: weight each term by how characteristic it is of a library — its document-frequency over the library’s size — so a generic word sitting in a handful of a huge corpus’s thousands of leaves stops mattering. And the fix over-corrected on the first try; the measurement said so (two topical queries wrongly abstained). The data told us the ranking was already perfect and only the threshold was wrong, so the fix became a relative-dominance rule instead of a bigger floor. Re-measured on the same corpus: routing went 3/6 → 6/6. That’s the Honest Loop working on werbos itself — a bug found by use, fixed by measurement, guarded by the gate.

You can use it now

werbos brain "who is a mandated reporter?" --registry libs.json — the whole pipeline from the command line. Route, ground, render, print the provenance so you can see why it answered. Local reader, the Anthropic API, or the claude CLI with no key. The library stays on the box.

And the orchestrator got its verdict

The multi-agent orchestrator had one open question since it shipped: on compound questions, does its abstain-combiner keep lies at or below the single reader? Forty minutes of 7B inference on a 16GB laptop answered it: single 4 lies → MAO 0. Four compound questions where the single reader grounded a confident half-answer, MAO turned into honest abstains — no bridge across an empty hole. It costs coverage (it answers fewer things), which is the honest trade: you’d rather hear “I can only ground half of this” than a confident half that reads as whole. It earns its place as the compound-question organ.

Every hard call today was settled by the bench, not by opinion. That’s the instrument doing its job — and now the instrument grades werbos, too.