Newest first. Every entry states what changed and why in its own words; open one to read the evidence behind it. Nothing here is edited or removed after the fact, including the rules that were later withdrawn.
October 5, 2026Markets are now also chosen by a published rule, up to 3 a day.
Until now every market Crossbook called was chosen by hand, so how many calls the record held depended on whether anyone had time that week — one call in the week of 28 September — and a reader had to take on trust that the calls were not simply the ones somebody liked. From 5 October a fixed rule also chooses, as set out in the rules above. Each call it makes is labelled as chosen by rule, with the rule’s version (auto-v1); calls chosen by hand continue and are labelled as chosen by hand. The rule only picks. Everything it picks goes through the same gates and is scored the same way as any other call.
October 4, 2026The review is closed. 42 corrections were appended, and the record now agrees with every venue's own settlement.
The review opened on 27 September, when every resolved call was checked against the venue’s own record of how the market settled. On 4 October the corrections were appended as 42 new entries in the resolution log; nothing was edited or removed, and every original entry still sits beside its correction on the call’s own page. Seventeen outcomes were changed to what the venue settled, six of them calls that had been recorded VOID although the venue never voided them. One call was reopened because its market has not settled. Twenty-three entries had only their settlement link or settlement time corrected, and one held call was recorded as settled. A re-check against the venues then found no disagreement left. With the corrections the published calls score 0.2054 against the market’s 0.1597 over 107 resolved calls; before them it was 0.2004 against 0.1781 over 105. The record trails the market by more than it appeared to, and that is the number that stands.
September 29, 2026The published number is now set by a fixed rule in code, and no model sees a price.
Since 20 September a model shown the market price decided whether a blind estimate was kept, moved partway, or dropped, and it could keep a disagreement only by naming a fact the market had not priced. That put the judgement of what was priced in with the party that wanted to publish, and in practice it refused almost everything: one of the last eleven calls before today was published. From today no model sees a price, a betting line or odds, and the published number is the market moved 50% of the way toward Crossbook’s own estimate, by rule. Sports estimates are now made without the sportsbook line, so a game is no longer refused for lacking one. Whether a cited fact was already priced in is still checked — against the page’s own date and the venue’s price history, in code — and recorded on every call, but it does not block one: replayed over the twenty most recent attempts, a blocking check would have published one, mostly because the pages facts come from rarely carry a publication date. Re-checking a call against a second venue is no longer offered, because it needed a model to see both prices. Calls recorded before today stay as made and are scored under the rules in force when they were made.
September 28, 2026A call made on one venue can be re-checked when its market is paired with the other.
A call is often made on one venue before the same question has been found on the other. The second price is new information — a second reading of where the money is — and until today there was no way for a call to take it into account short of being regenerated, which the 24-hour rule rightly refuses. Now, once a call’s market is paired, the call may be re-checked once against both prices. The blind estimate, the evidence and the citations are the original call’s and are not re-fetched; only the reconcile step runs again, with both venues’ prices and the pair’s settlement classification in view. It may move the number toward either venue, only as far as a cited, recent, material, unpriced fact allows, and never outside the range the blind estimate and the two prices span. The result is a new call with its own id and a distinct prompt fingerprint; the original stays as made. The score is unchanged by any of this: the earliest scorable call on a market is the one that counts, so a pairing added later cannot buy a second attempt at it.
September 27, 2026The record is under review. Checked against the venues' own settlement records, about one resolution in seven disagreed.
Until today every resolution on this record was entered by hand: an outcome and a settlement link typed in for each call. On 27 September every resolved call was checked against the venue’s own record of how its market settled, read directly from Polymarket’s and Kalshi’s published results. Of 121 resolved calls, 18 carry an outcome that disagrees with the venue’s; 8 link to a market other than the one the call was made on, five of them a different game of the same baseball series; and 22 were recorded before the venue had settled at all, one of them 46 hours early. Three things caused it. Outcomes and links were typed per call instead of read from the market. VOID was used to mean “do not score this repeat call”, when it means the venue cancelled or refunded the market — and repeat calls on one market were already excluded from the score by rule, so it was never needed. And corrections were stored but shown nowhere. Ten of the eighteen move the score: seven were being scored against the wrong answer or on a market that has not settled, and three had been marked VOID, which kept them out of it altogether. Seven more were logged after their event had started and count toward nothing either way; the last is a repeat call on a market already scored. One of the ten is t-0179, described in the 13 September entry below as a published loss: its resolution says NO, so it has been scoring as a correct call, and Polymarket settled it YES. What changes: resolutions are read from the venue, the settlement link is the market’s own, and each correction is appended as a new dated entry beside the one it replaces, which stays visible. Nothing is edited or removed. Checked in full, the corrected record scores 0.2029 against the market’s 0.1583 over 104 resolved calls — further behind than the 0.1938 against 0.1740 over 101 it showed before this check. Those figures become the record’s when the corrections are applied, and the notice on the record comes down only after they have been re-checked against the venues.
September 24, 2026Moving toward the market now needs a reason as well as keeping distance from it.
The reconcile step introduced on 20 September could keep an estimate only by citing a fact the market had not priced, but it could move an estimate toward the market by any amount with no citation at all. Over its first thirteen attempts it never once chose to keep. Every prediction that got through was a move toward the market, and every such move stopped between 37% and 55% of the way — on average 44.5% of the original disagreement left standing, with no fact named for any of it. That is splitting the difference, not weighing evidence, and it let half a disagreement with nothing behind it clear the three-point floor. Six predictions were published that way, on 21 and 22 September. They stay on the record as made and are scored like any other; a rule change never reaches back. From today, any disagreement the reconcile step keeps, whole or partial, must cite a recent, material, unpriced fact, and the citation is recorded on the prediction. Predictions made under this rule carry a new prompt fingerprint.
September 21, 2026The authored-prediction rule is withdrawn. It was never used.
On 15 September a rule was published allowing predictions written by the operator against a price the software withheld until the number was committed. Three drafts were opened under it and no prediction was ever recorded — every call on this record, before and after that date, was written by the model. The rule is removed and the software behind it is gone: the authoring page, its endpoints and the table that held the drafts. The reason is not that it failed but that it did not belong. Crossbook exists to measure whether a model is calibrated against two markets, and a human-authored forecast is a different claim sitting in the same table — a second thing to caveat in every figure, for a prediction nobody asked this site to make. The earlier entry stays below, because a changelog that deletes its own entries is not a record of anything.
September 20, 2026The price gets a say after the estimate, the sports outside view is computed, and three holes close.
Of the eight calls held for review on 20 September, seven were held for the right reason and all seven had the same shape: an estimate formed from a base rate the model remembered, with nothing in the brief that the market did not also have, drifting twenty to forty points from the price. A blind estimate with nothing to push against drifts, and the further it drifted the worse it scored. What changes: a separate reconcile step now sees the blind estimate and the price together and may keep the estimate only by naming a cited, recent, material fact the market has plausibly not priced, move it toward the market, or abstain — never move it away. Both numbers are stored on the row. For a sports fixture, retrieval must bring back the sportsbook consensus line (the points, never the odds — the one exception to the rule against searching for what anything is pricing), and the outside view is computed from that line and the league’s margin distribution rather than remembered; the blind estimate may sit at most 8 points from it. Chamber-control markets must retrieve the generic ballot and presidential approval. Three holes close: a market is not regenerated within 24 hours of its last call (two Senate calls two minutes apart moved eight points on fewer facts); a held call cannot be released once its market has closed, whether or not a resolution has been recorded (one was, eleven hours after close); and the hold reason now states the edge and the excess over the ceiling as two numbers. Nothing already on the record moves. Calls made under the reconcile step carry a new prompt fingerprint.
September 15, 2026Predictions may be written by the operator, not only by the model.
Each one records its author, and the two are shown separately as well as pooled. An authored prediction is made against a hidden price: the market’s number is withheld until the probability is committed, and the row records when each happened. Every gate that applies to a model prediction applies unchanged — the 3-point floor, the citation requirement, the disagreement ceiling and the settlement-claim filter. Predictions recorded before this date were all written by the model. Why: the model path cannot make this claim. The price is in its prompt because it is asked to explain divergence from it, and whether it anchored is invisible from outside; for a person the ordering is a property of the software rather than an instruction in a prompt. What the software enforces, exactly: the authoring page lists markets without prices, the draft carries the brief and not the price, and the price is joined only after the number is committed, with both moments recorded on the row and in a separate draft record. What it cannot enforce: that the author did not look elsewhere. So the claim is stated as what was withheld and when, and a question whose price this site has already shown the author — through a recorded call, or through a draft that was committed or refused — cannot be drafted blind again. The headline Brier still pools both authors; the split is shown beside it.
September 13, 2026Four gates that were computed and recorded now refuse, a fifth is new, and the disagreement floor has a ceiling.
On 9 September Crossbook recorded t-0179, a 30% call on Virginia covering a 44.5-point spread against Norfolk State, from a brief of five schedule facts — kickoff time, two records, a score, and “the teams have never played” — against a book quoted 31 bid and 71 ask. Virginia won 59–3. Reading the pipeline afterwards found that the retrieval step had been told to prefer recent facts and leave background to the forecaster, that the brief was graded “sufficient” because the retriever admitted no gaps, that the citation check had correctly found nothing cited and was configured to record rather than refuse, and that the model’s own “no independent basis” flag read false on every one of the twenty-seven takes made with retrieval — because the response schema both providers enforced did not contain the field. The prompt asked for it; the schema forbade it. t-0179 stays on the record as a published loss. It is not edited, reclassified or removed. What changes: retrieval is now told that who the sides are is its job, not background, with required categories per market shape and a larger budget for spread, total and margin markets; coverage is graded from what the facts contain rather than from what the retriever admitted; the model states its reference class and base rate as structured fields before estimating, or abstains; a thin brief, a false basis flag, or a thesis citing nothing material each refuse the take; no take is recorded against a book wider than 10 points; and a call more than 10 points from the market on sports, or 20 elsewhere, is held for review rather than published. Retrieval still never searches for what the market is pricing; a historical cover rate for a reference class is not a price. Every one of these is a rule at the door. Nothing already on the record moves, and predictions made under the new rules carry a different prompt fingerprint so the two eras can be compared rather than pooled.
August 30, 2026Eleven of the forty-one scored predictions were written inside the 48-hour window, and they still count. Keeping them makes the score slightly better, not worse.
An audit of every entry against the window found twenty-two written with less than 48 hours’ notice. Eleven were caught by the August 22 correction and count toward nothing. The other eleven predate the window entirely, and the rule directly below this changelog says a prediction is scored under the rules in force when it was recorded — so they are scored, including one written six minutes after its event began. The August 22 entry did not say this and reads as though no such entry survives. It does, and a reader should not have had to check. They are not being removed, and the number is why the decision needs showing rather than asserting. Those eleven score 0.1677 between them, which is better than the record as a whole. Dropping them would move the Brier from 0.1795 over 41 to 0.1838 over 30, and the gap to the market from 0.0113 behind to 0.0125 behind. Keeping them therefore flatters Crossbook on both figures, which is the direction that obliges disclosure. Each is now labelled on its own row in the record, with how far ahead of the event it was written. Separately, the two dates a contract carries are now told apart everywhere. A market has an event date and a settlement date, and on Polymarket game markets they differ by up to a week — measured across 100 live markets, 33 of the 73 fixtures settle later than the event, baseball consistently by seven days. Crossbook stored only one of them and showed it unlabelled. Both are now recorded and named: the record counts down to Contract settles, because that is when a call can be graded, and the screener counts down to Event date, because that is the last moment a forecast means anything. The 48-hour window still uses the earlier of the two, which is unchanged and is what stops a finished game being forecast for the week its contract takes to settle.
August 22, 2026Eleven entries stopped being scored, and the score improved sharply as a result. Both halves of that sentence matter.
Crossbook refuses to record a prediction inside 48 hours of resolution, because a number written after late information arrives is not a forecast. That check reads the market’s closing date, and the closing date was wrong: Polymarket publishes a date that is the event plus a dispute buffer, up to 31 days of it, so finished games looked days away and were accepted. Eleven predictions were recorded after the event they were about had already begun. They are not forecasts and can no longer be scored as if they were. They stay in the record, visible, each labelled with why it counts toward nothing. Every remaining prediction now carries the real event time, taken from the same date the market’s own rules cite, and the original value is kept alongside it. Removing those eleven moved the Brier score from 0.2557 over 23 resolved predictions to 0.1565 over 14. It also cost Crossbook its lead over the market, and that is the number that matters. Against the market baseline on the same predictions, the record was ahead by 0.0220 before the removal and is behind by 0.0029 after it. The eleven scored 0.4100 between them, which is worse than guessing, but the market scored 0.4706 on those same eleven, so they were the record’s strongest entries relative to the book even while being its worst in absolute terms. Removing them was still right, because a number written after the event has started is not a prediction whatever it scores. But it made the headline look better and the actual claim look worse, and a reader is owed both halves of that. The generation path now refuses on the real event time and separately refuses anything already under way.
August 20, 2026, v1.4Initial publication.
On the same day, 22 stored entries that fell below the 3 point threshold were deleted, to bring storage in line with Rule 1. None of them had resolved, so no score changed: the Brier score read 0.1594 against a market baseline of 0.1506 over 12 resolved predictions both before and after. They were never predictions under this rule; they were noise an earlier version of the software kept.
