[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-honesty-is-a-feature":3},"\nSomewhere right now a vendor is telling a buyer that their forecasting is\n98% accurate. The buyer will not ask on what error metric, over what\nhorizon, against what baseline, measured by whom. The claim will hang\nthere, technically unfalsifiable, emotionally reassuring, and the deal\nwill move forward on the strength of a number that means nothing.\n\nThis is the equilibrium of the enterprise AI market in 2026: every claim\nconfident, every benchmark self-administered, every case study a\nsurvivor. And it has produced exactly the buyer behaviour it deserves,\nwith leaders re-checking everything these systems produce, because\nnothing in the sales process gave them grounds not to.\n\n*In a market where every statement is confident, the only differentiated\nstatement is an honest one.*\n\n## What honest looks like as product behaviour\n\nHonesty in an intelligence system is not a tone. It is a set of concrete\nbehaviours, each of which costs something to build and pays back in\ncredibility:\n\n- **Say what is not predictable.** Some questions, on some data, have no\n  learnable signal. The honest system says so at the start: this target,\n  with your current history, cannot be forecast better than a naive\n  baseline. That sentence loses a line item and wins the account,\n  because it makes every other claim believable.\n- **Publish the misses with the hits.** Every model in production\n  carries its backtest, including the quarters it got wrong. One miss,\n  on the record with the rest, is worth more than ten case studies,\n  because it proves the record is real.\n- **Test against luck.** Run the backtest 20 times, once on the\n  real signal and 19 on noise. If the real result does not stand\n  clearly apart from the noise results, it is not a result. A finding\n  that cannot be told apart from luck should never reach a customer\n  slide.\n- **Keep the probabilities honest.** When the system says 70%, about\n  70% of those calls should land, and the reliability curve proving it\n  should be published each cycle, not asserted at purchase.\n\n> The cheapest thing to manufacture in AI is confidence. That is\n> precisely why it is worthless.\n\n## The commercial case, since it has to be made\n\nThe objection is obvious: honesty narrows the pitch. It concedes\nterritory a competitor will happily claim. This is true for one meeting\nand false for the relationship. The confident vendor owns the demo; the\nhonest one owns the renewal, because their claims were the ones that\nsurvived contact with production.\n\nThe harder commercial edge cuts deeper. An honest-limits posture\nconverts the sales conversation into an evaluation the buyer can run:\nagree a metric and a baseline in week one, run on the buyer's data with\nthe buyer's users, and measure. A vendor who publishes misses can offer\nthat without flinching. A vendor who cannot, cannot, and buyers\nincreasingly notice which is which.\n\nThe deepest reason, though, is not commercial. Enterprises are about to\ndelegate real decisions to these systems. The ones that get trusted\nshould be the ones that earned it, and the earning mechanism is a\npublished record with the failures left in. Any market that punishes\nhonesty about limits will eventually be corrected by its own incidents.\n\n*Trust the system that tells you what it cannot do. It is the only one\nwhose other claims mean anything.*\n\nHonest limits are built into how Prophesee works. Unpredictable targets\nare named as such, every published backtest includes the predictions it\ngot wrong, and the calibration record is published every cycle. Ask it\nwhat your data cannot support. [Start here](/contact).\n",1786786820103]