Yes, this is about drill matching. When a student types their answer, each key-fact bubble
  lights up when checkMatchGroups fires. The bug is that 80% of the content library (288/361
  keyFacts) was generated with the wrong group structure — and the bubbles are lighting up on
  almost any legal-sounding answer.

  What the groups are supposed to look like (exemplar):
  "groups": [
    ["grab", "seiz", "snatch", "pull"],   ← 4 synonyms for the action
    ["binder", "folder", "document"]       ← 3 synonyms for the object
  ]
  2 groups, 3-4 synonyms each. Student must demonstrate BOTH the action AND the object.
  Runtime: both groups must hit (exactly 2 groups → hitCount === 2).

  What's actually in the library:
  "groups": [
    ["equal", "equals"],   ← stem + inflection, one word
    ["pai", "pay"],
    ["act", "acts"],
    ["claim", "claims"],
    ["rais", "raised"],
    ["first", "firsts"],
    ["action", "actions"]  ← 7 single-word groups
  ]
  7-13 groups, 2 entries each (word + stem). Runtime: ceil(7/2)=4 → student needs any 4 of 7
  words. A student who writes "the court's first action on the claim" matches {first, action,
  claim} + {act} = 4 hits → bubble fires. They never mentioned the Equal Pay Act.

  How widespread: 80% of keyFacts (288/361). It's the entire generated content library — it
  was built before the current prompt was tightened. The exemplar files are fine; everything
  else is broken.

  What the quality gate in PR #2127 catches: It requires ≥2 groups and semantic independence —
   but it does NOT cap groups at 4. The broken content has disjoint stems across groups so it
  passes independence. The gate misses this failure mode entirely.

  ---
  Three things need to happen:

  1. Add a group-count ceiling to the quality gate — max 4-5 groups per keyFact. This would
  have rejected all 288 broken keyFacts at generation time.
  2. Update the prompt to specify "2-4 groups, each with 3-5 synonyms covering one semantic
  concept" and add an explicit BAD example of word-by-word tokenization.
  3. Regenerate the content library — the 63 drill files need new match groups generated from
  the corrected prompt. The embeddings are fine and don't need to change; only the
  match.groups fields need to be replaced.