{
  "_note": "Human/LLM-read labels for a 28-feature sample of the sentence-level k8 SONAR-SAE (sentence-sae_h16384_k8.pt), read from each feature's top-activating sentences in feat_dict_sentence-sae_h16384_k8.json. 'mono' = clearly monosemantic (single theme). Illustrates that low-k sentence-level features are often crisp; types include single-entity, clean-topic, and stylistic-template features.",
  "features": {
    "0":     {"label": "dated crime/incident report entries", "mono": true,  "type": "topic"},
    "585":   {"label": "philanthropy & social-impact investing partnerships", "mono": true, "type": "topic"},
    "1170":  {"label": "mixed arts/culture figures", "mono": false},
    "1755":  {"label": "immigration driving Canadian economic growth", "mono": true, "type": "topic"},
    "2341":  {"label": "scattered (residents / hip-hop / celebrity)", "mono": false},
    "2926":  {"label": "loosely fatherhood, but mixed", "mono": false},
    "3511":  {"label": "cultural heritage / art venues (loose)", "mono": false},
    "4096":  {"label": "film directors & their influence (one outlier)", "mono": false},
    "4681":  {"label": "memorial 'their legacy will live on' tributes", "mono": true, "type": "template"},
    "5266":  {"label": "regulatory 'Section N:' clauses (mixed)", "mono": false},
    "5851":  {"label": "scattered news vignettes", "mono": false},
    "6436":  {"label": "scattered (biotech / fiction / TV)", "mono": false},
    "7021":  {"label": "Ryukishi07's storytelling craft", "mono": true, "type": "entity"},
    "7606":  {"label": "'for more information, contact [X]' boilerplate", "mono": true, "type": "template"},
    "8191":  {"label": "scattered biographies", "mono": false},
    "8776":  {"label": "Surrey community kitchen programs", "mono": true, "type": "topic"},
    "9361":  {"label": "scattered arts/heritage projects", "mono": false},
    "9947":  {"label": "narrative vignettes (loose)", "mono": false},
    "10532": {"label": "governors declaring states of emergency", "mono": true, "type": "topic"},
    "11117": {"label": "scattered (geography / people)", "mono": false},
    "11702": {"label": "game/music releases (mixed)", "mono": false},
    "12287": {"label": "Gameboy Advance SP hardware variants", "mono": true, "type": "entity"},
    "12872": {"label": "profile of a person named 'Dawson'", "mono": true, "type": "entity"},
    "13457": {"label": "'as the dust settles on [event]' transitional phrasing", "mono": true, "type": "template"},
    "14042": {"label": "'* Work: [org] uses earth-science expertise' templated career blurbs", "mono": true, "type": "template"},
    "14627": {"label": "scattered biographies/film", "mono": false},
    "15212": {"label": "seasons / time of year (loose)", "mono": false},
    "15797": {"label": "geopolitical agreements & intelligence reports", "mono": true, "type": "topic"}
  },
  "summary": {
    "sampled": 28,
    "clearly_monosemantic": 12,
    "fraction": 0.43,
    "monosemantic_types_observed": ["single-entity", "clean-topic", "stylistic-template"],
    "takeaway": "Low-k sentence-level SAE features are frequently crisp; the monosemantic ones split into single-entity, clean-topic, and stylistic-template features. Full auto-labeling is a mechanical batch LLM job over feat_dict_sentence-sae_h16384_k8.json."
  }
}