v2.7 (July 2026) fixes two mechanical issues in how the seat-level projection is reconciled to the national polls, while keeping the recency-weighted point-in-time polling nowcast that sets the targets. First, the reconciliation that returns the per-constituency projection to the national margins now uses a structure-preserving additive swingrather than a proportional rescale: the old rescale hit the national totals but flattened each party's geography — collapsing safe seats into knife-edge margins — whereas the swing preserves the ordering and spread of a party's strength across seats. Second, the turnout step no longer double-counts turnout: because the polling targets already price in likely-voter behaviour, re-applying a turnout model was shifting roughly 3 points from lower-turnout parties to higher-turnout ones; v2.7 removes that shift at source while keeping the per-seat geography. Each party still lands within 0.75 points of its poll average, on the central path and inside every Monte Carlo simulation.

Because Britain's vote is split five ways, modest share shifts move the central allocation materially: with roughly 40% of seats decided by margins under 5 points, small distributional shifts swing many seats at once. The honest reading is a knife-edge — but over the majority, not over who leads. Reform is very likely the largest party (~95%), yet whether it reaches an overall majority is a coin-flip.

We lead with these probabilities and ranges, not a single seat number, because in a five-party first-past-the-post system the point estimate is the wrong object: symmetric polling noise scatters a fragmented plurality leader's narrow marginals to the runners-up, so the seat total swings by tens of seats on sub-point share moves. The displayed bands carry a 3-point national polling-miss floor and are never tightened below it. We publish what the model says rather than nudging it toward the MRPs — the reconciliation targets the model's own polling margins, never the MRP seat counts.

ReadingSignal is built and maintained by Jim Moodie. Jim spent 7.5 years as COO of Prolific, the online research platform used worldwide by academic researchers and social scientists — work that involved many of the same methodological challenges pollsters face: building representative samples, controlling for response biases, and assessing data quality at scale.

He has a long-standing interest in forecasting and prediction markets. ReadingSignal is one of several prediction modelling projects he's working on. More background at jahm.me.