Someone will read the werbos architecture — a small owned model, knowledge kept in cards instead of weights, an honest-abstain gate, a router that scopes retrieval, a tongue that reads instead of remembers — and think: I could just prompt an AI to build that.
You can’t. Not because it’s secret. Because you can only ask for the destination once you know it exists — and knowing it exists is the entire climb. werbos is the answer to a question most people don’t know to ask.
So we drew the climb. The werbos trail, at /trail/: 111 days
read straight off the build logs, as a hike.
The map is honest about the walls
The trail isn’t a triumphant line to the summit. It’s:
- Runs — the stretches where it all just worked (the day the model went “pure magic”; the day “move the knowing into the library” finally clicked).
- Walls — “the door said no.” “the library is the ceiling.” The reader wall, where scaling the model up did nothing and a 1M-parameter reader beat a 57M one.
- Dead ends — the dashed spurs we hiked into and turned around from. “Memorize the knowledge in the weights” — everyone’s first instinct — is a dead end that ends at the reader wall. “Bigger is better” is a dead end: our 30M model beat our 125M one. “Trust the loss number” is a dead end: it once ranked four models exactly backwards.
- False summits — twice we planted a flag and then, being honest, climbed back down. We “beat a 14B at coding” — then fixed a bug in our own grader and it was a tie. We shipped a trainer and celebrated — then noticed the model on top of it was still weak.
The dead ends are the proof
Here’s why the trail matters more than any single benchmark: the backtracks are what make it un-promptable. “Cards not weights” wasn’t a clever idea someone typed. It’s what was left standing after the memorize-in-the-weights approach hit a wall and fell over. You arrive at it by walking into the wall — not by being told the wall is there.
A straight-line version of this trail doesn’t exist. If it did, someone would have one-shot it already.
Generous map, hard territory
We published the whole thing — the trail, the day-logs, the training telemetry, the architecture. The map is deliberately generous. And it still doesn’t hand you werbos, because a map isn’t the territory. Following it still takes the walking: the taste to know when a result is lying to you, the stubbornness to keep knowledge out of the weights when every tutorial says to stuff it in.
The moat was never the model — the model costs about $150. It was never even the framework. It’s the judgment that produced the framework, and judgment is walked, not prompted.
Walk it
You can walk the trail at werbos.netlify.app/trail — every waypoint is a real entry in the logs, the walls in red, the dead ends dashed, the summit still dashed at the top because the base is at ~90% and we haven’t planted it yet.
It’s the most honest thing on the site. Not a pitch — a map of an actual expedition, backtracks and all.