[Atlas] Wiki-claim debates - extract claims from wiki pages and debate each done

Goal

The 18,447 wiki pages contain thousands of factual claims. Most are
unverified by structured debate. Build a pipeline that (a) extracts
top-level claims from a wiki page using a lightweight LLM call, (b) registers
each claim as an artifact_type='claim' artifact (or reuses artifact_type='hypothesis' with a claim subkind), and (c) auto-triggers
a 2-persona evidence_assessment debate on each claim against the citing
wiki page. Result: every wiki page acquires a tail of debated claims linked
back to it.

Acceptance Criteria

--max-claims-per-page 5 --dry-run. hypothesis with metadata.kind='claim'); links each claim to its
source wiki page via artifact_links link_type='derived_from'. personas [evidence_auditor, counterexample_finder], num_rounds=2). does_not_support, also writes an artifact_links contradicts row
pointing back at the wiki page (so the talk-page surfaces the dispute). processed in last 30 days). session and quality_score ≠ 0.5.

Approach

New module scidex/atlas/wiki_claim_extractor.py with the LLM extraction

prompt and the registration/queueing pipeline.

Add a claim value to ARTIFACT_TYPES in

scidex/atlas/artifact_registry.py (with metadata schema:
claim_text, source_page_slug, polarity).

Reuse queue_debate and the existing verdict parser.

Surface the debated claims in a "Claims under review" sidebar widget on

the wiki page (small, lazy-loaded).

Dependencies

• agr-ad-01-TARG — debate targeting.
• q-dsc-comments-on-wiki-pages — talk-page surfaces dispute UX.

Work Log