Why we pulled our own record.
We published a track record, re-measured it against our own stated bar, and pulled it. The case dataset behind it — named markets where our attention flag appeared to lead the crowd’s repricing — was withdrawn on 2026-07-15, and the empty list was left in place as the receipt. This page is the full account, told in the order it happened: what we published, what our re-measurement found, what we decided, and what survived it.
What we published
The dataset was a list of case studies: for each one, a market we had flagged, the date the flag fired, and the repricing that followed. It was presented as evidence that the flag led the move. As of its last published state (2026-07-10), the list holds 0 cases — because of what came next.
What our own re-measurement found
When we re-measured the published cases with the same instruments we point at everything else, the claim did not hold. The recorded reason, exactly as the API serves it:
Internal re-measurement showed flag-to-window lead times of 0-2 days on all published cases — below our own 3-day publication bar. The cases were withdrawn rather than relabeled. This is the honest state, not a temporary absence.
In plain terms: the flags were arriving with the news, not ahead of it — under the lead-time bar we ourselves had published as the standard a case must clear.
The decision
There were softer options. We could have relabeled the cases, moved the bar, or let the dataset quietly go stale. We withdrew it instead, dated the withdrawal (2026-07-15), and left the empty dataset in place — the API still serves the retirement notice, not a silent absence. The retired case list is intentionally empty and will not be repopulated under the old selection rules.
What survives
One measured result cleared re-measurement and stands: the population-level flagged-vs-control repricing regression. Its one-sentence form, rendered from the same committed result every surface reads:
Measured association: flagged markets repriced materially at 1.90x the rate of matched unflagged controls (95% CI 1.64-2.26, shock-days excluded, controls reweighted to the flagged cohort; n=1523 flagged vs 45864 controls, window 2026-03-29 to 2026-07-22, as of 2026-07-22). An attention-leads-movement association — never a directional claim.
That is an association across the whole flagged population — adjusted ratio 1.90× (95% CI 1.64–2.26), 1523 flagged markets against 45864 matched controls, window 2026-03-29 to 2026-07-22. It says flagged markets move more often than controls. It does not say which way, and it never became a directional claim — that is the caveat the sentence carries, permanently. Full context: the measurement on The System page.
Reproduce it yourself
The surviving regression ships with a reproducibility bundle: the inputs summary, the exact cohort and control definitions, the reweighting and shock-day rules, the measurement window, and an honest list of what the committed data can and cannot support. Download the reproducibility bundle (JSON) or read the machine record at /api/record-data.
The standard going forward
Three rules came out of this, and they bind every future record. A case publishes only when it clears the stated lead-time bar under mechanical re-measurement — the instrument decides, not the author. Population-level claims carry their confidence interval, their cohort sizes, and their non-directional caveat everywhere they appear, in one canonical sentence that cannot fork. And withdrawals stay published: the record includes its own retractions, dated, at the same addresses where the claims once stood. The historical selection rules the cases were measured against are preserved on the methodology page.
Informational only — not financial, legal, or investment advice. Prediction-market prices are shown as a signal of what the crowd believes.