[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-the-verification-gap-is-a-scoring-gap":3},"\nRoughly a quarter of executives say an AI error has reached their\nboard or an external audience before anyone caught it, part of the\nwider [verification gap](/insights/the-verification-gap) between\nhow fast AI output travels and how slowly anyone confirms it. Those\nerrors did not sneak past anyone. They got past the analyst, the\nreviewer and the final read-through because nobody asked the one\nquestion that mattered. How likely is this particular output to be\nwrong?\n\nMost organisations answer with more checking. That is the wrong\nlesson, or at best half of one, because checking harder cancels the\npoint of the machinery being checked.\n\n## The arithmetic against verification\n\nVerification means a person re-deriving what the machine produced.\nDo that for everything and the AI has saved you nothing; the work\nwas simply done twice. So no organisation verifies everything. Each\none verifies what it happens to distrust, and distrust gets\nallocated by instinct, recency and the seniority of whoever is\nasking. None of those tracks where the errors actually are.\n\nSelective checking is the only affordable kind, and selecting needs\na basis.\n\n*Unverified is the symptom. Unscored is the disease.*\n\n## What a score changes\n\nA score here is not the confidence a model reports about itself;\nthat number is where the problem starts. A score is a track record.\nThis model, on this class of question, in this context, has been\nright at this rate, and the rate comes from outcomes, not\nassertions. Attach that to every output and the checking question\nanswers itself. An output with a strong, current record earns a\nlight touch. An output from a model never scored on this question\nearns a heavy one, or does not travel at all.\n\nScoring also flips the burden. Today a person must justify\ndistrusting an output that looks finished; scored, the output must\njustify being trusted, on the page.\n\n## Where the score belongs\n\nThe score cannot live in the board pack, because by the time the\npack is assembled the error has already travelled. It has to attach\nwhere outputs are born and move with them. That makes this an\ninfrastructure problem, not a policy one. Outputs, decisions and\noutcomes have to flow through something that binds them together\ndurably, so that when an outcome lands, the model that called it\nwrong inherits the miss.\n\nThat binding is what [one event backbone](/insights/one-event-backbone)\nexists to provide. A memo telling staff to check AI outputs adds\neffort and no information. A backbone that scores outputs before\nthey travel puts the information exactly where the checking\ndecision gets made, so errors arrive at a threshold, with an owner,\nwhile they are still cheap. Prophesee scores on the backbone by\nconstruction. Every output carries the record it has earned, and no\namount of after-the-fact checking can imitate that.\n[Publishing the record](/insights/trust-is-a-record-you-publish)\nthen becomes a choice rather than a scramble.\n[Start here](/contact).\n",1786984937839]