Corrections

Every run of /api/health tests the outbound link route to confirm the referral is still attached, and each of those hits was logged in a way indistinguishable from a real click. Six of the seven clicks recorded that day were our own monitoring. The check now flags itself and is excluded.

On the next Israeli prime minister, Israel Katz appeared at 11c on Kalshi against 0.2c on Polymarket. That is not disagreement, it is a stale or mismatched quote. It cleared the outlier test on a technicality: the median gap on that market was 3.6 and three times 3.6 is 10.8, and the gap was 10.8 exactly. There is now an absolute ceiling of eight cents and a ratio test alongside the relative one.

It said "the only free board that prices prediction markets and sportsbooks together." True when written, and not true now: ArbBets, PredictionHunt, Converge, Matchr and Verso all compare across venues. The claim is now about the markup, which we can still defend.

Nothing was archived. The line has been replaced with the reason it matters: the only fair way to grade a forecast is against a price recorded before settlement, which is exactly the history we do not yet have.

On "Will OpenAI or Anthropic IPO first?" the page showed OpenAI at 17c on Kalshi against 85c on Polymarket. The 85c was Anthropic's price. Kalshi lists two outcomes there and Polymarket one, and the matcher only checked that Polymarket was a single market before stamping its one price across both rows.

"Who will be the next Supreme Court justice?" was matched against Polymarket's "Who will the Trump admin next nominate as SCOTUS Justice?" The judges' names are the same list. Nomination and confirmation are not the same question. Matches now have to agree on price, not just on names.

The Republican nomination market was published at 98.1%. An under-round is free money and therefore a measurement error rather than a finding: the sum was taken over only the contracts quoting on both sides. Books are now summed across the whole field or withheld.

Three seconds apart, because the run button was pressed twice and nothing remembered the first press. The poster now checks its own timeline before posting.

It graded settled markets on their last traded price. In a multi-outcome event each leg stops trading at a different moment and holds a stale quote, so those prices are not a coherent distribution: one ten-way contest had three legs at 37c, 49c and 49c. The section reported that things priced at 46% happened 1.9% of the time. Three attempted fixes failed. Restricting to clean standalone questions works and yields fourteen usable questions out of five thousand events. The feature was deleted rather than patched, and this site does not score anyone's forecasting until we have recorded our own prices before settlement.

The headline reported a gap on a game already under way, where one venue was trading live and the other had not caught up. Games in progress are now excluded.

It showed two venue prices and a saving measured across three, so the two numbers on the card did not account for the figure beside them. Later the same card rounded prices to whole cents while printing the saving to one decimal. Both fixed: the card shows the cheapest and the dearest, and the saving is derived from the two prices actually displayed.

The biggest daily swings were coming from obscure contracts with no volume, where a single stale quote lurching is not news. Three gates now apply: real open interest, real turnover today, and a price still genuinely in doubt.