Divergence stability score (v1)

The formula

stability = clamp(
  (mention_count_30d / max_mention_count_30d) × agreement_ratio,
  0, 1
) × 100

What each term means

• mention_count_30d — the number of evidence rows for this pair where the article was processed in the last 30 days. Pulled from entity_pair_evidence.
• max_mention_count_30d — the highest such count across every pair connected to the question entity in the same window. Used to normalize to [0, 1].
• agreement_ratio — in a future version, this will be the share of evidence rows whose stance label (NLP-classified) matches the majority. In v1 it is fixed at 1.0 because we do not yet run a stance classifier in production. The "low-evidence" badge surfaces pairs with fewer than 2 evidence rows.

What this number is

A relative ranking of which connections to the question entity have the most cited support over the last 30 days. Higher = more cited.

What it is NOT

A measure of consensus, popularity, importance, factual accuracy, or peer-reviewed agreement. The literal label every card carries is "signal, not consensus". A pair with a stability of 100 has more evidence rows than its siblings — nothing more.

Known limits (v1)

• No stance classifier — every evidence row counts as "agreement"
• 30-day window is fixed; not configurable per query
• Co-mentioned edges are excluded (86% of the graph is co-mentioned noise)
• Auto-pick selects the entity with the most typed-edge evidence; user can override via ?question=<id>

When this will change

A v2 of this formula will land when a stance NLP classifier is in production. v1 scores will be preserved alongside v2 during a transition window so cards do not silently shift values. See docs/plans/kg-four-view-redesign.md ADR-03 for the full rollback / versioning policy.