Methodology

Newscast AI reads public news feeds, groups articles that describe the same event, and compares how each outlet covered it. Three kinds of number appear in the interface: measurements taken from the article set, judgements made by a language model, and classifications maintained by hand. They are not equally reliable, and the interface says which is which.

Claim confidence tiers

A claim is a single checkable assertion extracted from an article — “the strike killed at least thirteen people”, not “the situation is deteriorating”. The same claim usually appears in several articles with different wording; those phrasings are grouped, and the group is graded by how many independent reporting chains carried it.

Confirmed

Three or more independent reporting chains assert this.

Corroborated

Two independent reporting chains assert this.

Reported

One reporting chain. Not yet independently corroborated.

Disputed

Outlets disagree — opposite polarity or conflicting figures for the same quantity.

Unverified

No reporting chain could be attributed.

The numeric confidence beside a tier is a smoothing function, not a probability: confidence = min(0.99, 0.35 x independent reporting chains + 0.25 x ln(attesting articles + 1)). Independence dominates deliberately: extra syndicated copies of one dispatch raise reach, not confidence.

Two limits worth knowing. Claim grouping is lexical — it matches content words, entities and numbers, and it will occasionally split two phrasings of the same fact, which understates support. And a claim can be carried by three independent newsrooms and still be wrong if all three trusted the same briefing; independence is a check on syndication, not on truth.

Source independence

Independent reporting chains, not outlet count. Articles are grouped into chains using byline agency attribution (a Reuters byline on another outlet is a republished dispatch), whether the outlet is itself a wire service, and near-identical body text across outlets. Ten papers running one AP dispatch counts as one chain.

Ten outlets running one agency dispatch is one piece of evidence, not ten. So every trust signal counts reporting chains rather than logos. A chain is identified three ways, in descending order of reliability: an explicit agency credit in the byline; the outlet being a wire service itself; and near-duplicate body text across outlets, which catches syndication where the feed has stripped the byline.

The honest caveat: many feeds publish no author field at all. Where provenance is unstated we label the article Byline not stated and treat it as its own chain, which can overstate independence. Wherever that happens, the evidence badge says so rather than quietly rounding up.

Heat

Heat measures how hard a story is being covered right now. It is a measurement, not an opinion — breadth of coverage and volume of articles, decayed by age:

(12 × outlets + 4 × articles) × recency

Recency falls linearly from 1 to 0.2 across a 48-hour window, then holds at the floor so an older story with heavy coverage never scores zero. Story age is clamped to a minimum of 30 minutes, because a cluster minutes old would otherwise produce an implausible rate. Clicking any heat figure shows the itemised arithmetic for that story, including what the score was when it was last computed and what it has decayed to since.

Velocity

Velocity is articles per hour — how fast outlets are filing on this story. Two figures are shown where both are available: the lifetime average since the first article, and the trailing 24-hour rate, which is the one that tells you whether the story is still moving. Velocity says nothing about importance; a celebrity arrest can out-file a famine.

Importance and sentiment

Editorial importance, 0-100, judged by the analysis model from the article set — scale of impact, how many people are affected, geopolitical consequence and durability. It is a model judgement, not a computed formula, so it is shown with the factors the model cited.

Aggregate tone of the coverage, 0-100, where 50 is neutral. This measures how outlets are framing the event, not whether the event is good or bad, and not the analyst's opinion of it.

These two are model judgements, and the interface marks them as such. They are useful for sorting and comparison and should not be read as measurements. Two runs on the same articles can differ by a few points.

Editorial lean

The lean chip on a source card is a fixed, editor-maintained classification of the outlet, stored next to its feed URL in this repository. It is not inferred by a model, it is not computed from the article, and it is not a quality score. Its only purpose is to give you context when comparing emphasis and word choice across outlets. For accuracy, read the claim tiers.

Labels are coarse by design and reasonable people dispute them. The current assignments:

Coverage comparison

Publication order comes from each article's feed timestamp, so “first to report” means first to appear in the feed, which can lag a website by a few minutes. Unique-claim counts compare the claims extracted from one article against every other article in the same story: an outlet credited with adding three claims published three checkable assertions nobody else in the set had. Coverage gaps are drawn from the same comparison plus the model's outlet-by-outlet reading, and are phrased as gaps rather than accusations — a gap is often just a house style or a wire-length constraint.

Publish gate

An episode is only marked ready when its script passes every gate: no claim in the script contradicted by the evidence layer, no unsupported figure, an accuracy score above the threshold, and audio that matches the script. Failures are listed individually with the specific fix, because “61%” on its own tells an editor nothing. A blocked episode can still be played and inspected — it simply is not marked publishable.

Audio and visuals

Speech is synthesised locally with Kokoro-82M. Per-line timings shown in the transcript are measured from the generated audio files, not estimated from word counts, which is what makes the highlight during playback line up with what you hear. Illustrations are generated by an image model unless a card explicitly credits a source photograph; every visual carries its own provenance label, and nothing generated is presented as documentary evidence.