Do the prices come true?

The x-axis is the price actually paid, not the band midpoint. A 0–5¢ band is mostly longshots trading at one and two cents, so its mean price is nowhere near 2.5¢. Measured against the midpoint, that band reads −0.64¢; against what was actually paid, −0.28¢. The file's own calib_error and significant columns use the midpoint and are deliberately not what is drawn here.

One contest is one observation. Thousands of prints on a single match share one result. Every interval is a CR1 cluster-robust standard error with the contest as the cluster; treating prints as independent understates it by 26× to 90×. Dots are sized by the effective contest count, which additionally accounts for how concentrated traded volume is across them.

Thin products remain in the all-products view. The to advance product carries 3.2 million trades across roughly 32 contests, so it can materially move a pooled curve despite resting on very few independent outcomes. Bands that fail the effective-contest threshold remain visible as descriptive context but have no uncertainty whisker and are excluded from significance claims. Use the table's measurable-bands column before treating a product-level result as durable.

Whose price this is. Each Polymarket US symbol is a single binary leg and the daily market report lists only that leg, so this is a leg-price curve. Time-and-sales carries no aggressor flag, so unlike Kalshi's series this is not taker-side, and small level differences between the two venues should not be over-read.

What is excluded. Symbols naming more than one contract cannot be attributed to an outcome and are dropped rather than guessed at; markets delisted before reaching their event are never counted as losses. In total 89.8% of trades and 89.4% of traded value reach a resolved, unambiguously attributable market.