How the NYT “vibes” component works

The vibes layer is an experimental adjustment on top of the base polling average. It tries to capture momentum and scandal effects that polls pick up only with a lag:

1. We pull each candidate’s recent New York Times coverage (Archive API) and extract every sentence that mentions them.
2. A sentiment model scores each mention, producing a five-point tone bucket from overwhelmingly negative (−2) to overwhelmingly positive (+2), plus a 0–1 scandal-severity score from curated triggers (indictment, ethics investigation, resignation, …).
3. Each candidate’s effect on the margin is 0.4 × tone − 2.5 × scandal points; the race adjustment is the Democrat’s effect minus the Republican’s, capped at ±3 points so vibes can nudge — never overturn — what the polls say.

The coefficients are conservative priors, not fitted values; they will be backtested as labelled cycles accumulate. When coverage data hasn’t been fetched yet, the toggle is disabled and the base average is shown.