Broadcast Prism turns TV news transcripts into a map of attention — what each channel covered, for how long, and where agendas lined up or diverged. This page explains every metric on the site: what it measures, how it's calculated, where it can mislead, and the minimum sample size before it's worth trusting. These are measures of attention, not verdicts on truth.
Story & claim clustering
What is measured. The grouping of similar reports across different TV channels into a single canonical story, and the consolidation of individual factual claims within those stories so the same claim made across channels lines up.
How it's calculated. Reports are first grouped by shared entities and semantic similarity within a single topic. A language-model deduplication pass maps provisional story hooks to active stories or creates new ones. A second, global cross-topic merge pass (default 14 days lookback) collapses stories describing the same real-world development across different topics. Finally, a cross-topic claim merge identifies and merges duplicate factual claims across the unified story.
Known failure modes. Two genuinely different stories might be merged if they share similar entities (like generic names or places), or a single story might be split if channels report it using completely different details or vocabulary. Claim merging may fail to group statements that differ slightly in nuance.
Minimum sample size. Grouping has no meaning for a single channel's isolated report; it requires coverage from at least two channels to trace alignment and overlap.
Airtime
What is measured. The amount of time, in seconds, a channel spent on a story or topic in a broadcast.
How it's calculated. Airtime is summed per channel from the durations of identified segments. Segment boundaries and durations are extracted from the raw transcripts based on the start and end timecodes of the speech segments associated with a story.
Known failure modes. Mis-segmented transcript boundaries can cut off sections or include unrelated content; ads and handovers may be misidentified as story content; a single topic split across non-adjacent segments may have its airtime under-counted or double-counted depending on segment alignment.
Minimum sample size. A single broadcast's airtime is noisy and susceptible to format variation (e.g. short summaries vs long packages); stable comparison of channel priorities requires a sample of several days.
Coverage gaps & blind spots
What is measured. Stories some channels covered while others omitted, flagged with HIGH, MEDIUM, or LOW severity, along with "persistent blind spots" that track these omissions over time.
How it's calculated. Omissions are detected by comparing the presence and absence of stories across the channel-by-channel matrix. Severity is scored based on the overall duration of the story and the coverage capacity of the channel. Blind spots aggregate these omissions into all-time channel-specific profiles.
Known failure modes. Shorter broadcasts naturally cover fewer stories, which the algorithms try to account for but can still misidentify as gaps. A channel covering a story under a totally different frame or topic title may be incorrectly flagged as having omitted the story.
Minimum sample size. Individual daily gap flags are indicators of difference, not bias; identifying a systemic blind spot requires tracking coverage over a sustained period of at least two to three weeks.
Editorial fingerprints
What is measured. The unique focus areas of a channel—specifically which topics it covers significantly more or less than its peers, and which topics it covers exclusively.
How it's calculated. For each channel, we calculate the proportion of its daily airtime spent on a topic and compare it to the average share that topic received across all channels. A channel "over-indexes" on a topic if its proportion exceeds the average peer emphasis. Topics covered by only one channel are flagged as distinctive. The channels page condenses this into a distinctiveness band: we take the root-mean-square of a channel's category-share deltas from the network average (in percentage points) and label it Mainstream below 3, Independent from 3 to 6, and Contrarian at 6 or above.
Known failure modes. A single massive breaking story can dominate a channel's daily airtime, throwing off fingerprints. Ratios can also be highly volatile for topics with very low absolute airtime (e.g., under 10 seconds). The band thresholds are editorial calibrations against the live dataset, not statistical significance levels.
Minimum sample size. A single day's fingerprint can be heavily skewed by one broadcast's running order; identifying a channel's true editorial signature requires analysing several days of coverage.
Agenda alignment
What is measured. The degree of similarity between the daily news agendas of any two channels—showing which outlets align in their story selection and which diverge.
How it's calculated. We represent each channel's daily agenda as a vector of its topic airtime shares, then calculate the cosine similarity between these vectors. A score of 100% indicates identical relative emphasis across all topics, while 0% indicates no overlapping topics.
Known failure modes. Agenda alignment measures focus, not viewpoint. Two channels that spend half their broadcasts covering the same political debate will score as highly aligned, even if their reporting frames the debate in opposite ways. Low-volume days also make similarity scores highly sensitive.
Minimum sample size. Unreliable for broadcasts covering fewer than three topics, as sparse vectors artificially inflate similarity scores; requires a normal multi-story news day to be meaningful.
Topic co-occurrence & editorial network
What is measured. The network of topics that tend to be covered together in the same broadcasts, and the resulting layout of broadcaster relationships based on shared coverage patterns.
How it's calculated. Jaccard similarity is computed over the sets of days topics appear together. For the broadcaster network graph, we construct undirected links between channels whose all-time cosine similarity is at least 20%.
Known failure modes. Co-occurrence shows correlation rather than direct editorial intent. High-volume general categories (such as Politics) will naturally co-occur with many minor stories, creating dense clusters that obscure more subtle editorial relationships.
Minimum sample size. Constructing a stable co-occurrence network requires a large historical dataset of at least 30 distinct days of multi-channel coverage.
Narrative synthesis
What is measured. The neutral, cross-channel summary and channel-by-channel breakdown generated for major news stories covered by multiple broadcasters.
How it's calculated. We gather all covering channels' transcripts for a given story, clamp them to the story's time window (up to a character limit), and pass them to a large language model configured per deployment (e.g. via OLLAMA_MODEL or OPENROUTER_MODEL). The model is prompted as an expert news analyst and editor to write a balanced summary and map individual channel angles.
Known failure modes. The model can smooth over genuine editorial disagreements into a false consensus, inherit errors or misattributions from raw transcripts, or exhibit hallucination, especially on complex or fast-moving developments. It is not manually fact-checked.
Minimum sample size. Summaries are only generated for stories covered by at least two distinct channels; single-channel reports do not qualify for cross-channel synthesis.
Story lifecycle badges
What is measured. The active status of a topic or story in the news cycle, classified into one of five states: Emerging, Peaking, Fading, Recurring, or Dormant.
How it's calculated. Status is classified deterministically by evaluating the topic's relative airtime history. It is Fading when latest airtime falls below 0.5 × its all-time max; Peaking when latest is at least 0.8 × max and greater than its previous appearance; Recurring when it has appeared on at least three distinct dates without peaking or fading; Emerging when it is a new topic or growing in airtime; and Dormant when it does not appear on the latest archive date.
Known failure modes. A sudden one-day spike can prematurely label a topic as Peaking, even if it immediately disappears the next day. The boundaries are relative to each topic's own history, meaning low-volume topics can trigger state changes as easily as major stories.
Minimum sample size. A topic must appear on at least three distinct dates to qualify for Recurring status; prior to that, it will default to Emerging or Dormant.
Attention volatility
What is measured. How much a channel's daily agenda swings from one day to the next, indicating whether it maintains a stable focus or reactively shifts its priorities.
How it's calculated. We compute the cosine distance (1 minus cosine similarity) between a channel's category airtime shares on consecutive days. The average of these day-on-day distances defines its volatility score: Stable (under 25%), Moderate (25% to 45%), or Reactive (over 45%).
Known failure modes. A genuinely fast-moving, high-impact news week will cause even normally consistent channels to score as highly volatile; the score does not distinguish between chasing sensations and covering major daily developments.
Minimum sample size. Requires at least two days of coverage for a channel to compute a single day-on-day shift, and a minimum of 7 consecutive days to establish a reliable average.
Contested & charged quotes
What is measured. Direct quotations that are flagged as contested (challenged by other sources or fact-checkers) or loaded/emotionally charged, mapped directly to their surrounding transcript excerpt.
How it's calculated. A language model extracts candidate quotes from transcripts based on loaded language or claims under dispute. A verification pass matches these candidates back to the exact transcript segments to establish evidence anchors, timecodes, and still images, showing the highlighted span in its exact context. Trend graphs normalise the quote rate by dividing counts by the channel's analysed hours on that day. The channels page also shows an intensity chip from the raw flagged-quote count over the window: none at zero, moderate below 10, high at 10 or more.
Known failure modes. Sarcasm, hypothetical statements, or paraphrases can be misidentified as loaded language; matching can fail if the transcript contains minor phonetic errors. Normalisation can also produce high rates on days with very short broadcasts if a single quote is flagged, and the intensity chip uses raw counts, so high-volume channels reach it more easily.
Minimum sample size. Single flagged quotes are illustrative examples of framing, not statistical proof; drawing conclusions about a channel's overall quote profile requires tracking quote rates over at least a week.
Category classification
What is measured. The thematic category (such as Politics, Health, Economy, Ukraine, or Royal) a story is filed under.
How it's calculated. Assigned by the language model when topics are first extracted from transcripts. The model selects from a defined set of categories, falling back to uncategorised if none match.
Known failure modes. Multi-category stories (e.g. a political scandal about water companies) must be forced into a single primary category, losing secondary context; niche or edge topics can be misclassified.
Minimum sample size. A single story category is a simple classification; category shares and channel agenda breakdowns require a full broadcast or a week of broadcasts to represent real emphasis.
Broadcast stills
What is measured. Representative video frames captured from news broadcasts to illustrate stories, show visual framing, and verify transcript contents.
How it's calculated. Video frames are sampled upstream at a fixed interval of every 60 seconds (as declared on the copyright page). We select the frame nearest to the midpoint of the story's coverage window to represent it, using the images strictly under the fair dealing exception for criticism and review.
Known failure modes. Sampling at a fixed interval can miss brief but significant visual moments, or accidentally capture commercials, channel graphics, or presenters during transition segments.
Minimum sample size. Not a metric; stills are purely illustrative and have no statistical threshold or sample size requirements.
News density
What is measured. The proportion of a broadcast slot or a channel's total airtime spent on identified news stories rather than commercials, presenter handovers, or transitions.
How it's calculated. For a single broadcast, it is calculated as the duration of the union of all story time-windows divided by the total broadcast duration. For a channel, it is the sum of all its broadcasts' news seconds divided by the sum of their total durations, preventing short broadcasts from dominating.
Known failure modes. Stories the pipeline fails to identify read as "not news"; overlapping story windows are combined (unioned) rather than double-counted, but mis-timed story windows will skew the ratio.
Minimum sample size. A single broadcast's density is highly dependent on formatting; channel comparisons require a large sample size of at least 10 broadcasts to represent the true news-to-filler ratio.
Signature vocabulary
What is measured. The distinctive words and phrases a channel uses unusually often compared to all other channels in the archive.
How it's calculated. Calculated deterministically (without an LLM) using a log-odds ratio with an informative Dirichlet prior (alpha0 = 500). Frequencies of unigrams and bigrams are compared against the all-channel baseline. Terms must appear at least 5 times, have a statistically significant z-score of at least 1.96, and are filtered through a stopword list and a two-layer blocklist. Surfaced terms link directly to the transcript segments as receipts.
Known failure modes. Presenter names, recurring programme titles, and transcription errors can pass the filters and score highly; channels with smaller transcript volumes can have highly inflated scores.
Minimum sample size. To prevent noise, a channel must have accumulated at least 5,000 total tokens in its transcripts to qualify, and terms with fewer than 5 occurrences are omitted.
Weekly overview & superlatives
What is measured. Weekly awards (Most Divergent, Closest Pair, Biggest Blind Spot, Most Contested Quote, and Vanished Fastest) and glance stats that summarize broadcaster focus over a completed week.
How it's calculated. Rendered only when the week is complete (every captured day status is final, latest data on/after the week's Sunday, and at least three data days are present). Divergence and closeness are computed via cosine similarity over weekly topic-airtime-share vectors. Blind spots require at least three channels covering a topic for at least 10 minutes (600 seconds) total. Vanished fastest requires a story peak of at least 5 minutes (300 seconds) between Monday and Friday followed by silence on at least two subsequent data days. Accused or winning channels must have broadcasted on at least three days that week.
Known failure modes. A channel that broadcasts just under the three-day qualification threshold will be omitted from awards; gaps in data capture will shrink the evidence base, though the site displays missing days as inactive chips.
Minimum sample size. The minimum qualifications (at least three days of data, completed week status, and channel eligibility thresholds) serve as the minimum sample gate.
Web article context & source spectrum
What is measured. External web press articles attached to major stories to provide context, along with a per-source editorial spectrum label indicating the outlet's political positioning.
How it's calculated. An entity-first search query built from the story title is sent to Meilisearch. Results are filtered to keep articles at or above a relevance floor of 0.45, or rescued if they score above 0.15 and mention a key entity in the title. Political spectrum labels (left, centre-left, centre, centre-right, right) are editorial assignments mapped per source in the database.
Known failure modes. Relevance filtering can occasionally exclude valid articles or admit irrelevant ones with similar titles; political spectrum labels are shorthand tags for the outlet itself rather than a direct measurement of the specific article.
Minimum sample size. Not a metric; articles are context and have no statistical threshold or minimum sample size.
Trending & new story rails
What is measured. The stories surfaced on the homepage rails, highlighting trending coverage and newly detected developments.
How it's calculated. Trending scores are computed by summing a story's daily airtime with an exponential decay of 0.7 per day. A story must have aired within 24 hours of the newest data to qualify as trending. A "NEW" badge is assigned to stories first detected within 48 hours of the newest data. All windows are anchored to the newest broadcast date in the archive, not the wall clock.
Known failure modes. With sparse or missing data, the decay calculation can let a single long broadcast dominate the trending rankings; "new" status refers to the first time the pipeline processed the story, not its first broadcast globally.
Minimum sample size. Rankings are only as fresh as the last pipeline run and depend on having at least one recent broadcast within the qualification window.
Lead story comparison
What is measured. The story each channel led with — and where channels agreed or split — shown on date pages and the homepage as “What led the news”.
How it's calculated. When a channel has several analysed broadcasts in a day, the longest one counts as its flagship. Each story's time window is intersected with the flagship's opening ten minutes (600 seconds), and the story with the greatest overlap is the lead; if nothing overlaps the opening window, the earliest story is used instead and labelled “earliest story”. The consensus lead is the story that led on the most channels.
Known failure modes. Headline round-ups or teaser segments inside the opening window can outweigh the true lead; mis-timed story windows shift the overlap; when a broadcast's end time is missing, its duration is approximated from summed story airtime, which can occasionally pick the wrong flagship.
Minimum sample size. One day's lead is a single editorial decision — patterns need a run of days. The comparison renders only when at least two channels have an analysable flagship broadcast.
Broadcast Prism summaries are cross-channel consensus, not an objective account. Topics, airtime, and gaps come from automated transcript analysis and may contain errors. Treat every metric here as a map of where attention went, not a verdict on what happened.