Russian missile and drone strikes kill three in Kyiv and leave about 600,000 without electricity

A poll is shown only with pollster, race and result. Field dates and sample size are required before it counts toward an average; until then each poll is a single Reported claim.

Averages are analysis by the named aggregator, not a Manifest finding. Shown with attribution; never in a headline.

1. Three signals per race. Fundamentals: the state's 2024 presidential margin, plus half of the national swing implied by the generic ballot, plus an incumbency edge (±3; appointed ±1). Ratings: the Cook rating as a margin (Toss-up 0, Lean ±4, Likely ±9, Solid ±18). Polls: qualifying toplines from the last 21 days, weighted by pollster quality and recency; subgroup polls excluded.
2. Combined by precision. Each signal carries its own uncertainty (σ 7 fundamentals, σ 5 ratings, σ 4.2/√weight polls); the model's margin is the precision-weighted average, so a race with two good polls listens to the polls and a race with none leans on fundamentals and ratings.
3. Errors that move together. Each simulation draws one national error (σ grows with days to the election), one error per region (Northeast, Midwest, South, West), and a race-level error. A 2016- or 2020-style miss shifts every race at once; the Midwest can miss on its own.
4. Seat math. Senate: 34 Democratic-caucus seats not up + simulated wins; 51 controls, 50 is Republican control via the Vice President. House: the national vote is modeled from the generic ballot (shrunk 25% toward even this far out), then the 435 Cook ratings are shifted by the gap between that vote and the environment the ratings already assume. Governors: ratings with a reduced national effect.
5. What it doesn't know. Turnout, late shifts, scandals, ballot design, court rulings, third-party dynamics. It is uncalibrated until backtested. Treat 60/40 as a coin with a thumb on it.