Product teams remember what they shipped.
Not what they believed.

Bet began as a “decision journal.” The deeper I took the concept, the less convincing that container became. The effort is front-loaded, the payoff distant and probabilistic, and the core promise is difficult to validate, since decision quality and outcome quality are only loosely coupled. It’s the profile of a tool people admire and quietly stop using.

The durable idea survived the container. Product organisations keep excellent records of what they did, and almost no memory of why they believed it was right. Most of the product stack records actions, artifacts, and outcomes. Very little preserves the reasoning behind a decision in a form that can be revisited when reality answers. Bet is an attempt at that instrument.

Everything turns on the unit. I tested several — hypothesis, assumption, forecast, decision — and each missed something: too formal to speak aloud, too passive, too binary. “Bet” was the only one that combined belief, commitment, and accountability in language product teams already use. That combination matters practically: you can’t measure whether someone’s judgment is well-calibrated using objects that were never actually risked on.

A bet carries a claim, the confidence behind it, the assumption it rests on, the alternative it passed over, and the part most decisions never get — a resolution condition, set in advance.

It never sets your confidence.

The system can read a claim from your text and infer the assumption underneath it — but confidence is left blank for you to own. This adds friction, and I kept it: a calibration score is meaningless if the confidence being scored was the model’s guess rather than yours.

Confidence is structured, not free-form.

I used a small fixed scale rather than free-entry percentages, choosing consistency over precision — a “73%” you can’t actually feel isn’t more honest than a band that means the same thing every time. This is one of the assumptions I’d test first with real users; the specific scale is a reasoned bet, not a proven one.

Where the criterion is explicit, settlement is anchored to it.

The bet resolves against the threshold set in advance, rather than against a retrospective story. In the checkout example, the redesign lifted completion by 1.4 points against a 3-point criterion: not met, cleanly, even though it lifted small-cart completion sharply. That small-cart result is real, but it’s a new finding — the seed of a new bet — not grounds to relabel a failed one as a partial win. Letting the flattering subset soften the verdict would reintroduce exactly the bias the product exists to remove.

The AI reads structure out of prose written for other purposes, and it can propose possible links between a current bet and earlier ones, with its reasoning visible so you can confirm or reject the connection. That kind of cross-time recall becomes increasingly difficult to maintain by hand as decisions accumulate and language changes.

Its role is bounded on purpose. It doesn’t set your confidence. It may calculate whether an explicit threshold was met, but it doesn’t interpret ambiguous evidence or redefine success after the fact. And it proposes tensions for you to judge — it doesn’t assert them. A product about calibration has no business being miscalibrated about its own conclusions, which is why the record view, with only a handful of settled bets, says plainly that there isn’t yet enough to say.

Four things I haven’t validated. Whether the fixed confidence scale matches how people actually hold uncertainty. Whether the tension-detection stays trustworthy as history grows, or starts surfacing coincidences that read as insight. The cold-start problem — the structural value works from the first bet, but the cross-bet value only arrives with volume. And adoption: whether extraction genuinely lowers the capture cost enough to beat the friction that killed decision journals in the first place.