How Sunlight works

Sunlight measures how readers perceive political slant — not bias as an objective fact. It is a mirror of the crowd, and we say so plainly.

What we measure

Readers place a news article on a five-point left–right scale — far left, mid-left, center, mid-right, far right. Those one-tap placements roll up into a placement for each article, journalist, and outlet. The number reflects perception by the people who choose to rate, not a verdict on what is true.

Any reward model encodes a value judgment. Sunlight rewards agreement — reporting a broad cross-section of readers agrees reads as unslanted — not mere proximity to the numeric center. That is a definition, openly stated, not a claim of objective truth.

The three numbers

Every entity can carry up to three placements, and the gaps between them are part of the story:

Raw — a straight one-person-one-vote mean. This is the headline.
Lean-balanced — the five self-reported cohorts re-weighted to equal size, so the platform's own user-base skew doesn't tilt the result.
Inferred-balanced — the same balancing, but by each rater's inferred lean. Shown only once enough raters carry a confident inference, so it isn't noise dressed up as precision.

Consensus vs. contested

A mean alone can mislead: an article half its raters call far-left and half call far-right averages to “center.” So disagreement gates the placement. When raters are too divided, the entity is marked Contested and we show how the room split — never a misleading lone center dot.

Reporting vs. opinion

Opinion pieces are typed separately (a human confirms the call) and are kept out of the article rubric — they feed an outlet's opinion-balance instead, so an op-ed is never counted against a journalist's reporting record.

Your lean stays yours

Telling Sunlight how you lean is optional and consent-gated. It is used only to balance cohorts and — later — to reflect your own patterns privately back to you. It never makes your individual ratings count more or less.

Known limits

It measures perception by a self-selected rater pool. Lean-balancing corrects cohort size, not who chooses to rate in the first place.
Individual-journalist placements are held to a higher bar and a legal review before they go public.
Early numbers are thin; placements appear only past a minimum volume of ratings.