Day 135 — compose correctly, or say I won't guess
A marathon. werbos got an OCR reader that reads text off images with a confidence — and abstains on what it can't read instead of inventing it. It got a $0 launch stack for going public behind a password: a concurrency gate, a per-session rate limiter, a live pressure gauge, and the fix that mattered — a public guest can never fall through into the operator's own chats. It learned to fill a page's images from a topic. And the big one: a live kitchen-sink test found the compose engine looked perfect on its gauntlet but was brittle and quietly dishonest in the wild — so it got hardened, kitchen sink after kitchen sink, until every real prompt either computes correctly or honestly refuses. Never a confident wrong answer.
do the thing correctly, or refuse out loud — and prove it against the prompts that were breaking you an hour ago
A long one. Four things shipped, and a fifth got beaten into shape until it held the whole idea werbos is built on: honesty is the product.
Eyes that abstain
werbos ocr reads text off an image with a small on-device engine. The trick isn’t that it reads — it’s
that every word comes back with a confidence, so a smudge it can’t make out gets left blank instead
of guessed. A big vision model will confidently invent text it can’t actually see; this one says “not
sure” and moves on. Same honesty the rest of the system runs on, pointed at pixels.
Getting ready to open the door
I’m prepping a public launch, behind a password. That meant bounding the box: how many people can it
actually hold? So werbos got a concurrency gate (it caps the slow work at the machine’s real limit
and queues honestly — “you’re next in line” — which, delightfully, rate-limited me mid-test, which is
how you know it works), a flood valve so one person can’t hammer it, and a live gauge you can
curl to watch the load during a launch. The one that mattered most: a public guest can never land
in the operator’s own chats — a request without a session gets its own sealed silo, full stop.
A website about dogs, with actual dogs
The page builder can now fill an image slot from a topic — a real photo if it can find one, a generated one if not, and an honest empty box if neither. “A website about dogs” gets dogs.
The big one: never bluff
Here’s the day. The compose engine — the thing that turns “sum of the even numbers in [1,2,3,4,5,6]” into a real, run, verified answer — passes a 478-problem gauntlet. But that gauntlet tests the engine, not the part that reads your English. And when I threw a kitchen sink of real prompts at the English part, it cracked: it missed easy things, and worse, when it missed it would make something up — “the product of [2,3,4]” got an essay about premium products; “if x > 5…” got answered “10 is larger.” The one thing werbos is supposed to never do — bluff — it was doing.
So I closed it, one kitchen sink at a time. Now it either composes the right answer (filters, conditionals, three-way branches, per-element math, multi-step arithmetic, counting, and more) or it honestly abstains — “I can see what you want, but I can’t build it yet, so I won’t guess.” Never the confident-wrong thing in between. And I froze the operator’s own real prompts into a growing gauntlet that gets a new locked test every time the wild finds a new crack. It ended the day at 27/27, zero confident-wrong answers — across the exact prompts that were breaking it hours earlier.
And, off the books
Somewhere in the middle, a Raspberry Pi node (a Bitcoin box) went into a boot loop — a crash-looping service drowning the login. Killed it, disabled two more broken services, re-authed its network, brought it back. Not werbos work, but the same instinct: find the one thing that’s lying about being healthy, and make it stop.
The thread through all of it: compose correctly, or say “I won’t guess.” Everything else is detail.