Claim Extractor

claim-extractor

Reads a fetched source document and extracts the falsifiable claims — statements with truth-value, anchored to a verbatim excerpt in the source. Downstream skills (verify-claim, memory-curator) consume the standard claim schema; without this layer, research outputs become a soup of "the article says…" assertions with no falsifiability and no anchor.

When to use

• Inside any Category 1 research skill after fetching a source via source-fetcher — extract claims before deciding which to verify and persist.
• Inside weekly-intelligence-digest / voices-watcher / any Category 2 agent when an item is interesting enough to mine.
• Inside memory-curator when staging a multi-claim document into _inbox/ — each claim becomes its own candidate.

When NOT to use

• Single-fact inputs (the user typed "GitHub Copilot data handling is X" — already a claim, no extraction needed).
• Code snippets, schemas, structured data — those aren't claims.
• The body of a digest the toolkit just wrote — that's our own output, not external sourced material.

Input

{
  "content_md": "<markdown body, post prompt-injection-guard>",
  "source_url": "https://example.com/article",
  "source_tier": 1 | 2 | 3,
  "fetched_at": "2026-06-20T15:00:00Z"
}

The content_md is whatever source-fetcher returned. Quarantined blocks (from prompt-injection-guard) should be excluded from claim mining — they are not trustworthy source material.

What counts as falsifiable

A claim is falsifiable if a competent reader could in principle confirm or refute it from independent evidence. Concretely:

• ✅ Falsifiable: "Copilot Enterprise audit log now exports to S3 with KMS encryption." "FFIEC's 2026 IT Handbook adds Section 7.4 on AI/ML risk." "Anthropic released Claude Opus 4.7 on May 14, 2026."
• ❌ Not falsifiable (opinion / marketing / vague): "Copilot is a game-changer." "Most banks struggle with X." "This is the most important AI update of the year."
• ❌ Forward-looking without commitment (skip unless tied to a dated commitment): "We plan to support…" "Coming soon…" "Expected in Q4." (Keep if tied to an official roadmap entry; drop if vague intent.)

What to exclude

• Content inside quarantine blockquotes (per prompt-injection-guard).
• Boilerplate (author bio, footer disclaimers, "subscribe" CTAs).
• Headings without a corresponding declarative sentence (they're navigation, not claims).
• Direct quotes attributed to a person (those are claims-about-a-person, not the source's claim; capture as a separate claim_type: attribution).

Output schema

[
  {
    "claim": "Copilot Enterprise audit log now exports to S3 with KMS encryption.",
    "quoted_anchor": "Audit log now exports to S3 with KMS encryption (previously CSV download only).",
    "source_url": "https://example.com/article",
    "source_tier": 1,
    "claim_type": "fact | event | quantitative | attribution | comparative | roadmap-commitment",
    "entities": ["copilot", "github"],
    "tags_suggested": ["copilot", "ghas"],
    "line_range": [12, 12],
    "confidence_hint": 2
  }
]

Field guidance:

• claim: cleanly stated, self-contained — "another AI with zero prior context should understand it" (per the writing standard).
• quoted_anchor: a verbatim excerpt from content_md that supports the claim. Required; if you can't find an anchor, the claim isn't extractable.
• source_tier: copy from input.
• claim_type: one of the enum values. Most claims are fact (typed claim about an entity) or event (something happened on a date). quantitative for numeric claims with values. attribution for "X said Y." comparative for "X is better/larger/cheaper than Y." roadmap-commitment for dated forward-looking statements with vendor backing.
• entities: lowercase-kebab entity names (e.g., copilot, ffiec, johncutlefish). Used downstream to route facts to facts/{entity}/{predicate}.md.
• tags_suggested: from the controlled vocabulary in vault/_meta/tags.md. Used downstream to set frontmatter tags:.
• line_range: approximate [start, end] line numbers in content_md for the quoted anchor — helps a human reviewer locate context.
• confidence_hint: a starting confidence (1-3) based on source_tier + claim specificity. The actual confidence gets set after verify-claim.

Default caps

• Max 25 claims per source to keep downstream verification tractable (per the built-in deep-research pattern). If the source has more, return the top 25 by load-bearing-ness (specific facts > comparatives > attributions > marketing-adjacent).
• If 0 falsifiable claims found, return an empty list with a warnings field explaining why ("source is opinion piece", "all content was quarantined", "source is a CTA / signup page"). Per the stop-and-report guardrail: caller must know zero was a deliberate result, not silent failure.

Composes with

• source-fetcher — input is its content_md.
• prompt-injection-guard — already applied; quarantined blocks excluded from mining.
• verify-claim — downstream; takes the extracted claims and runs adversarial verification.
• memory-curator — uses extracted claims to decide promote/patch/drop per claim.

Acceptance test (for step 5 done-criteria)

Run on tests/sample-article.md (in this skill folder). Expected:

• ≥3 falsifiable claims extracted.
• Each claim has a non-empty quoted_anchor that exists verbatim in the source.
• Vague / opinion lines are not extracted.
• A claim from inside a quarantine block (if the fixture has one) is not extracted.
• Output includes entities and tags_suggested from the controlled vocabulary.