Integrity

2026-07-27 → 2026-10-25 · inconclusive 0

Official statistics get revised. Our contracts say the first published reading decides the question, so a later revision must never quietly change a settled result. This watch is the evidence that it does not: if a re-read ever disagrees, it is logged as a disagreement, not applied.

Read this as measurement, not as a verdict on any source. A slice that points toward the more common outcome will look better without being better, so these numbers only become a statement about source quality once the base rate is accounted for and n is much larger. We publish them now because the ledger should be visible while it is still inconvenient.

• block 3332026-07-270609ea3e44c2

  Addendum 6 · a mechanism we had preregistered was never wired

  Two earlier addenda specified how resolved outcomes should recalibrate the rules behind a forecast, and we proved the mechanism with rollback tests. It then sat idle: the production resolution path never called it, so for weeks not a single rule carried real-world evidence. We found it, wrote the finding into this chain before fixing it, and made the repair forward-only so no past result was retroactively credited.

• block 3342026-07-277ce85dcbb4a9

  Addendum 13 · the field that recorded which rules were used was wrong

  Every decision stored a rule set that belonged to a different engine, so the credit and penalty from a resolved question landed nowhere. The verdict itself was never affected, but the learning loop was writing to an empty address. Recorded here first, then corrected forward-only.

• block 4312026-08-04b919481cf232

  Addenda 15 and 15.1 · the new gate shipped with the wrong formula, and was corrected before it ever fired

  A new publication gate was specified against a measured failure class, but the formula we froze quietly deviated from the exact formula used in the measurement. A pre-acceptance verification pass showed the deployed version would have suppressed our best forecasts instead of the failing class. It was corrected in a follow-up addendum before a single production forecast was affected; both the flawed specification and the correction stand as consecutive blocks in the chain.

• block 14572026-10-139fcf5645ff2d

  We have stopped publishing in one question family

  Our forecasts on US continuing jobless claims have been poor. The 18 forecasts we published in this family average a Brier score of 0.51; someone who always said 50 percent would score 0.25. The last four weeks were worse: six forecasts on three questions scored 0.70. We have stopped publishing new forecasts in this family. We keep opening questions, forecasting behind the scenes and scoring; none of the past records has been removed or altered. We will return to publishing only if a criterion we recorded in advance is met: doing no worse than a coin flip on at least eight new questions. If we remain clearly worse, we will close the family altogether; if we land in between, we will wait for eight more questions, once.

Anyone can publish a hash chain of their wins. The reason these entries exist is that a record which cannot hold bad news about itself is not a record. Both were repaired forward-only: no earlier result was recalculated, and the chain still carries the account of the failure.

1717 chain blocks · 34 preregistration documents · verify the chain