All Models Are Wrong
Status: concept — the operating philosophy behind consensus.
George Box's observation — all models are wrong, but some are useful — is not a caveat PRISM apologizes for; it is the design premise. Every model in the ensemble errs routinely, and the architecture exists so that this does not matter.
Wrong on purpose
Each PRISM model is a deliberately narrow specialist: trained on one pool's patient histories until continuing the six-column table is the only thing it can do. Individually, these models are noisy. One will occasionally fire on a patient with nothing to find; another will stay silent on a pattern it should recognize; all of them are sampling from a probability distribution, not consulting a fact.
Trying to make a single model reliably right would mean fighting that statistical nature head-on. PRISM does not fight it. The engineering effort goes into making the errors uncorrelated rather than nonexistent: five models trained on disjoint pools, with no shared data, weights, or adapters, fail independently. Independent failures rarely coincide. Shared truths do.
Screening medicine already works this way
A system that is wrong most of the time sounds unacceptable until you notice that this is how screening medicine already operates. Mammograms, colonoscopies, and routine blood panels all return far more negative results than positive ones — deliberately. When a condition is uncommon but an early catch is valuable, a low positive-predictive-value program is not a flaw; it is the standard, accepted shape of the trade. Nobody scores a screening program on the accuracy of its individual results. It is scored on whether the catches justify the searching, because any instrument that finds needles must examine a great deal of hay.
PRISM's suggestions are that kind of instrument. On real data, most flags will not yield a diagnosis, and the business is built to expect exactly that: PRISM is paid only when a suggestion leads to a documented early detection, so the inefficiency of negative screens is priced in rather than hidden.
Most investigations rule things out
The same tolerance for wrongness runs through diagnosis itself. Differential diagnosis is a process of listing possibilities and eliminating most of them; the majority of tests a physician orders come back negative, and that is the method working, not failing. A negative result narrows the differential — it has value.
PRISM enters this workflow as one more reason to consider a test, addressed to a physician who keeps full discretion. The asymmetry is architectural: the system can add a possibility to the list but has no pathway to remove one, and its silence is never a clearance. A wrong suggestion costs one ruled-out possibility; a suppressed one could cost years.
Consensus turns many wrong models into one trustworthy flag
Useful wrongness only becomes a usable signal through voting. If errors are independent, agreement among models is improbable except where something real drives it — so requiring agreement filters noise without touching signal.
The 2026 synthetic proof of concept demonstrated this directly on its NOPE controls: patients whose records contain nothing that could legitimately predict the test, so any fire is pure noise. Requiring unanimity collapsed NOPE false fires twenty-fold, to 1 in 1,500, at no cost on the unanimously-fired GOOD carriers. The full grid, misses, and ablations are in the prototype results.
What this does not license
Two honest limits. First, those numbers come from deliberately clean synthetic data; they are an upper bound on the mechanism, not a forecast for real claims. Second, the philosophy depends entirely on independence being maintained: the pools were generated heterogeneously on purpose, and any sharing of data or weights across them would silently re-correlate the errors that voting is supposed to cancel. Useful wrongness is a property of the ensemble's structure, not a blanket excuse for noise — a single model's output remains meaningless on its own, which is precisely why no PRISM output is ever a single model's.