The record — methodology (retired)

The case-study Record is retired.

On 2026-07-15 we re-measured every published case against the selection bar below and found flag-to-window lead times of 0–2 days on all of them — under our own 3-day minimum. We withdrew the cases rather than relabel them. The case list is intentionally empty and will not be repopulated under these rules. What survives is the population-level measurement — the flagged-vs-control repricing regression, served in the record’s data artifact — and the calibration scorecard, which keeps grading every published probability against reality.

The rules below are preserved as the historical record of how cases were selected, and as the bar any future case product would have to clear before publishing.

Selection rules (historical)

  • Minimum move: The prediction market's implied probability must move at least 20 percentage points across the reprice window. Smaller moves are not shown as evidence.
  • Lead window: Our flag must precede the START of the crowd's reprice by 3 to 45 days — measured to the move's first qualifying day, never to when the move completed (a July 2026 external audit found our earlier cases displayed completion time as notice; all were retired). Longer “leads” on permanently-watched entities are not treated as notice.
  • Thesis relevance: Detection-day coverage must speak to the market's actual question, not merely name the entity. Entity attention plus an unrelated market move is an association, never a documented case.
  • Baseline history: An entity must have been watched for at least 14 days before its flag counts — a flag raised near the start of our data has no baseline to be measured against and is excluded as a left-edge artifact.
  • Cross-domain breadth: The flag must have appeared across at least 8 of our coverage domains on the detection day.
  • On-subject evidence: At least one detection-day headline must actually be about the entity (word-boundary match), so the attention evidence is auditable on the page. Keyword mentions scattered across unrelated coverage do not qualify — entity ubiquity is not signal. Stated honestly: this establishes entity-level linkage (abnormal attention on X preceded movement in an X market), not an event-level causal chain — the detection-day story and the market may concern different developments about the same entity. An event-level gate is the named next tightening.
  • Time to resolution: We parse each market's deadline from its own question. A case whose reprice window ends within 7 days of the deadline at an extreme price (below 5% or above 95%) is excluded: markets drift toward 0% or 100% near expiry, and we do not present expiry mechanics as foresight. Each case discloses days-to-resolution measured from the flag date.
  • Event shocks: The move must be visible across the window, not delivered by a single terminal jump. A market that sat still and resolved on one surprise announcement is not evidence the world repriced toward our flag.
  • One per event: At most one case per entity and one per underlying event family. Markets about the same event are correlated, and we do not present correlated questions as independent confirmations.
  • No novelty markets: Handshake lengths, net-worth trackers, phrasing bets and similar entertainment markets are excluded even when they qualify mechanically.

Known limitations

  • Cases are selected from qualifying rows, so this is a curated exhibit, not a random sample. The population-level question — do flagged markets move more often than comparable unflagged markets? — is the standing measurement behind the scenes, and cases are illustrations of it, not proof by themselves.
  • Surviving cases are still correlated with one another (the same world events drive many markets); do not treat the case count as N independent trials.
  • The dataset is pinned to a stated as-of date and refreshed monthly, so the newest possible case always lags by the length of its own reprice window.
  • Prediction-market prices are the crowd's belief, cited as a public sensor — not our forecast, and never advice.

Reproducibility

The retirement is machine-readable: GET /api/record-data returns an explicit retirement object — why the cases were withdrawn, the empty former dataset, and the surviving regression result (estimate, 95% CI, cohort sizes, strata, window, as-of). For the population-level scoring — calibration curve, Brier score, log loss — see the calibration page.

Informational only — not financial, legal, or investment advice. Prediction-market prices are shown as a signal of what the crowd believes.