The posting windows on this page are historical patterns, not validated predictions, and they cover only the seasonal rows: recompete rows carry a recorded expiry date and no window at all. Backtested 2026-07-25 on 39,003 base awards: out of sample the month rule scored worse than assuming every month equally likely (Brier 1.49 vs 0.92; 95% CI on the gap excludes zero on both substrates). It named about 1.4 months of the year and the award landed in one of them 18.9% of the time. A second attempt was then pre-registered and measured: shrinking each cell's own month histogram toward the global month prior. It lost too, and on the contracts substrate that whole family is mathematically dominated, so no calibrated cell-level month probability is available from this corpus. That is why this page is not called a forecast.

Some month concentration is real, but far less than it looks. Cells that declare a peak are screened for multiplicity once same-day batch signings are counted as one buying decision instead of many, and then again for whether the peak recurs across more than one calendar year. Most do not survive: the rest are a batch of related awards, not a cycle. Rows we cannot support have been removed from this board: each remaining pattern row survived that screen and recurs across at least two years. Accuracy is unchanged by the removal; nothing was tightened or improved.

Recompete rows no longer carry a predicted window. They used to assume the successor posts 2 to 8 months before expiry; that assumption was measured and removed. Anchoring the window on a contract's own expiry found a solicitation in the same office and NAICS 17.4% of the time, against 15.5% for the identical window anchored at a placebo date drawn from the same expiry calendar: a lift of +0.019 with a 95% interval of [-0.008, +0.048] across 22,564 expired contracts in 2,394 buying cells. The interval includes zero, so the expiry date carries no timing signal we can demonstrate. What remains on these rows is the expiry date itself, the incumbent, and the contract value, all of which are recorded facts. Full method and figures: docs/model-card-forecast.md.