About Broadcast Prism

Broadcast Prism compares how the UK's main TV news channels cover the same stories. For each day it extracts the topics each channel covered, detects where one channel omitted a story that most others ran, and produces a cross-channel summary of each gap topic.

1. Topic extraction. Each channel's broadcast transcript is sent to an LLM with a versioned prompt that returns structured topics (name, category, relevance, estimated airtime).
2. Deduplication. Newly extracted topic names are matched against the existing canonical topic list (LLM-assisted) so the same story is linked across channels and days rather than duplicated.
3. Coverage gap detection (deterministic). A topic × channel presence matrix is built from the linked topics. A channel is flagged as having a gap for a topic only if it did not cover the topic, a majority of other channels did, and its total bulletin airtime for the day is not materially below the cross-channel median — i.e. it had spare capacity rather than simply a shorter bulletin. Severity is assigned from how many channels covered the topic and the spare-capacity margin. The gap decision is a fixed rule, not an LLM judgement.
4. Neutral summary. For each gap topic, transcripts from all covering channels are passed to an LLM with a versioned prompt that produces a cross-channel summary with key claims and source attribution.