Cuba will release 51 people from prison, an unexpected move amid pressure from Trump

How coverage is measured: Prism extracts atomic claims ("nuggets") from every article covering this event using an LLM, then deduplicates them into a consensus claims pool — the set of all claims reported by at least one source.

Each article is then scored by an LLM judge on how many of these consensus claims it includes. Claims can be supported (clearly stated), partially supported (implied or incomplete), or not supported (omitted).

Weighted score: Partial support counts as 50%. "Vital" claims (confirmed by 3+ sources) are weighted 2× more heavily, so omitting a widely-reported claim penalizes the score more than omitting a claim only one source mentioned.

A score of 100% means the article covered every consensus claim. Lower scores indicate more claims were omitted.

How objectivity is measured: Each sentence in the article is run through a subjectivity classifier (GroNLP/mdebertav3-subjectivity-english), a DeBERTa model trained to distinguish objective from subjective language.

Rather than a simple yes/no, each sentence gets a continuous subjectivity probability from 0% (fully objective) to 100% (fully subjective). The article's score is the average across all its sentences.

Labels: Highly Objective (<10%) · Mostly Objective (10-15%) · Mixed (15-22%) · Mostly Subjective (22-30%) · Highly Subjective (>30%)

Lower scores indicate more neutral, information-focused reporting. Higher scores indicate more opinionated or narrative-driven language.

How framing is measured: Each sentence in the article is run through a political leaning classifier (matous-volf/political-leaning-politics), a model trained to detect left, center, and right-leaning language patterns.

Each sentence produces three probability scores — left, center, and right — which are averaged across all sentences to give the article's overall framing profile.

The stacked bar shows the proportion of each leaning. The dominant leaning is whichever scored highest.

This measures language patterns, not accuracy. An article can use left-leaning framing while reporting the same claims as other sources, or vice versa.