Confidently wrong is the only unforgivable bug

An agent that does your data entry is only worth having if you can stop checking its work. That's the entire promise: you snap the photo, you forward the email, and you don't re-read every date it filed to make sure it got them right. The moment you have to double-check everything, you haven't saved any time — you've added a step and a worry.

So the one bug I can't ship is the one that quietly poisons that: the confident mistake.

The mistake that costs you the customer

Early on I handed Dustav a firefighter's shift calendar — a dense, colour-coded grid where the whole meaning lives in which shade each day is. Asked which days were a given shift, it produced a clean, confident, day-by-day list. It was wrong. Not "I'm not sure" wrong — here are your shifts, filed and reminded wrong, off by a couple of days, delivered with total composure.

That's the failure that ends the product. A tool that's right 95% of the time but tells you it's right 100% of the time is worse than a tool that's right 80% and tells you which 20% it wasn't sure about — because the second one you can trust, and the first one you can't trust at all. If it can hallucinate even once and hand it to you straight-faced, you're back to checking everything, and now you're checking a machine's homework, which is somehow worse than doing it yourself.

Read before you conclude

The subtler version of this bug isn't misreading a picture. It's concluding too early from too little.

A real one: a customer of my design partner had an order, and somewhere in the thread was a reference to a refund. The reader saw the word, saw the number, and concluded a refund had been issued. Confident. Wrong. Because it had only read the subject line and a one-line snippet — not the actual conversation, where my partner had asked that customer to place a different order, and the "refund" was a top-up that had already been printed and shipped. The truth was three messages deep, and the reader had stopped at message one.

The fix was not a smarter model. It was giving the reader the ability — and the instruction — to open the whole thread and read it before deciding anything. Investigate, then conclude. Cross-reference the order numbers. And if it's still genuinely ambiguous after reading everything, the correct move isn't a best guess with a confident face. It's to ask. A scheduler that occasionally asks a clarifying question is doing its job; one that fills the gap with a plausible invention is doing the one thing I can't allow.

Say what you can't read

Back to that shift grid. The honest architecture isn't "extract it and hope." It's a reading-check that runs before extraction and rates how reliably the thing can even be read — and when the answer is "this is too dense to call," it says so, out loud, in its own voice, and offers a way through: crop it to one month and I'll read that cleanly. It stops trying to be a hero on an input it can't handle and becomes honest about it instead.

That's the same instinct everywhere in Dustav: when the machine can't do a reliable one-shot read, it surfaces what it can see and lets a human close the gap, rather than asserting something it can't stand behind. The product isn't "never makes a mistake." It's "never makes a confident mistake."

Zero confident mistakes, not zero mistakes

I can't promise you a machine that never misreads a smudged photo or a weird receipt. Nobody can, and a founder who tells you otherwise is selling you the exact confident-wrongness I'm trying to kill. What I can build toward is a machine that never lies about it — that flags the shaky read, quarantines the guess, asks the question, and leaves a visible trail of what it actually did so you can check the one thing it was unsure about instead of all of them.

Every wrong-but-flagged read is recoverable in a single glance. Every wrong-but-confident read is a landmine you step on weeks later, at picture day, or at tax time.

The one-sentence version

A scheduler that's occasionally wrong and always honest, I'd trust with my week. A scheduler that's usually right and never unsure, I wouldn't trust with a dentist appointment. Dustav is built to be the first one — because for the thing it's trying to be, confidently wrong is the only unforgivable bug.