at.kotoba.state

{
  "$type": "com.atproto.lexicon.schema",
  "defs": {
    "main": {
      "description": "Single overwritten record: per-glyph SRS state + the directed confusion map. Read once at startup; overwritten on flush. rkey is literally `self`.",
      "key": "literal:self",
      "record": {
        "properties": {
          "confusion": {
            "description": "Directed, single-kana only: shown glyph -> { answered-as romaji: count }. Counts are decayed (x0.9/session) so the app stores them as fixed-point ×10000 INTEGERS (no float type). `unknown` map; the app owns the value scaling at the record edge.",
            "type": "unknown"
          },
          "glyphs": {
            "description": "Map keyed by glyph -> { box:int, kanji:{box:int,seen:int,lastSeen?:str}, lastSeen?:str, modes:{ tap:{seen,missed,ewmaParseMs|null}, type:{seen,missed,ewmaParseMs|null}, meaning:{seen,missed,ewmaParseMs|null} } }. Per-mode success + latency; `box` is kana-reading recall and `kanji` is the separate ruby-bearing written-form recognition ladder (#171). `kanji.lastSeen` is that lane's own date clock for overdue scheduling (#134), distinct from the glyph's `lastSeen`, which every card stamps; older kanji lanes may omit it. `meaning` (word -> gloss) only ever accrues for glossed word glyphs. Counts, boxes and latencies are integers; dates are strings. A missing mode row reads as empty; a missing `kanji` lane reads cold; legacy flat records ({box,ewmaParseMs,seen,missed}) are migrated to tap on read. `unknown` because lexicons have no dynamic-key map type (see README).",
            "type": "unknown"
          },
          "updatedAt": {
            "format": "datetime",
            "type": "string"
          }
        },
        "required": [
          "updatedAt",
          "glyphs",
          "confusion"
        ],
        "type": "object"
      },
      "type": "record"
    }
  },
  "id": "at.kotoba.state",
  "lexicon": 1
}