Perspectives — how outlets differ

For stories covered by several sources, perspectives shows where outlets agree, where they disagree, and how each one frames the story — strictly grounded in the sources' own text. There are no truth ratings and no political labels for outlets.

How it works

1. Per-source digest — each member article gets a digest: {claims: [{text, quote} ≤3], stance, kind, key_parties, framing} where every quote is at most 15 words and must appear verbatim in the article text (whitespace and curly quotes normalized). Numbers must appear in the text. Invalid output is retried with the specific error and then dropped; the digest row is stored anyway with status invalid. Headline-only members get a deterministic title digest (stance: unclear) without an LLM call. Digests live in the article_digests table, keyed by content hash — a digest is computed once per (article text, story claim, prompt version, model) and reused.

2. Synthesis — once at least perspectives.min_sources (default 4) digests exist, one synthesis is generated: {agreement: [{text, citations}], disagreements: [{topic, positions: [{label, summary, citations, quote}]}], different_angles: [...]} and stored as a story_summaries row with kind="perspectives".

A disagreement is kept only when at least two positions each cite a different source, each with a verified quote, and the cited stances differ; otherwise it is omitted silently — balance is never invented. Position labels must describe the view (e.g. "says the tax will raise revenue"): an outlet's name or a political label is rejected. The synthesis runs the same number check and verbatim guard as briefs.

1. Counter-narrative retrieval — deep stories get extra context search angles: "criticism of", "defense of", "fact-check", affected parties' "reaction" and expert "analysis". Fact-check pages are stored as context_kind="fact_check" and labeled in the UI; opinion links stay labeled and capped by context.max_opinion.

2. Scheduling — the summary worker runs a perspectives pass after the follow-ups pass, once per story version (the input hash covers the member set, each member's text hash and the central claim). It is rate limited by perspectives.max_per_hour, backs off on transport failures like late regenerations, and never blocks briefs.

UI and API

The story page gets a collapsible How outlets differ section (agreement, disagreements with source chips and the quoted evidence, different angles) and each source in the Coverage panel gets a stance badge (hidden when unclear). API v1 adds story.perspectives and source.stance. A profile can turn the section off from the pipeline settings page (perspectives_enabled).

Configuration

Per profile (perspectives:) — see configuration reference:

perspectives:
  enabled: true
  min_sources: 4        # synthesize only with at least this many digests
  max_per_hour: 6       # rate limit, like late regenerations
  max_articles: 12      # at most this many members are digested
  digest_max_words: 1400  # article words sent to the digest prompt
  max_age_hours: 168    # stories are eligible while published within this

The work uses the profile's llm.perspectives role, defaulting to llm.summarizer — point it at a stronger endpoint if you have one.

Tools

python -m app.perspectives.diagnose --story-id N      # digests, quotes, errors, synthesis
python -m app.perspectives.evaluate --stories 20 --markdown   # review runs for false balance

evaluate is a dry run by default (writes nothing); pass --store to persist syntheses, --json for machine output.