btx_e003/instruments/codebooks_i5.json

{
 "instrument": "btx-e003 R0b I-5 codebooks (model-free recipe)",
 "when": "2026-09-06 00:47",
 "corpora": {
  "d5": {
   "n_states": 27,
   "dims_used": 27,
   "class_counts": {
    "word": 18,
    "Name": 9,
    "digit": 0
   },
   "count_1_dropped": 0,
   "top_states": [
    [
     "is",
     264000
    ],
    [
     "If",
     120000
    ],
    [
     "So",
     120000
    ],
    [
     "are",
     120000
    ],
    [
     "someone",
     120000
    ],
    [
     "then",
     120000
    ],
    [
     "they",
     120000
    ],
    [
     "tovi",
     32270
    ],
    [
     "marn",
     32175
    ],
    [
     "trell",
     32173
    ],
    [
     "dorn",
     32077
    ],
    [
     "vint",
     32048
    ]
   ],
   "singular_share_top4": 0.5727,
   "erank_feats": 11.9,
   "rsa_E0_vs_raw_feature_cosines": 0.0653,
   "rsa_convention_note": "the record gauge RSA(E_C, byte features) compares a TRAINED codebook against E0 (its init); this row is E0 against the raw centered feature cosines (low by construction: the SVD init equalizes directions)",
   "rsa_E0_vs_random": -0.015,
   "b_prior_range": [
    -4.188,
    -1.776
   ]
  },
  "sub3": {
   "n_states": 1009,
   "dims_used": 256,
   "class_counts": {
    "word": 6,
    "Name": 4,
    "digit": 999
   },
   "count_1_dropped": 0,
   "top_states": [
    [
     "is",
     40100
    ],
    [
     "Compute",
     20050
    ],
    [
     "It",
     20050
    ],
    [
     "Reply",
     20050
    ],
    [
     "What",
     20050
    ],
    [
     "minus",
     20050
    ],
    [
     "number",
     20050
    ],
    [
     "only",
     20050
    ],
    [
     "the",
     20050
    ],
    [
     "with",
     20050
    ],
    [
     "122",
     220
    ],
    [
     "135",
     206
    ]
   ],
   "singular_share_top4": 0.3015,
   "erank_feats": 50.5,
   "rsa_E0_vs_raw_feature_cosines": 0.096,
   "rsa_convention_note": "the record gauge RSA(E_C, byte features) compares a TRAINED codebook against E0 (its init); this row is E0 against the raw centered feature cosines (low by construction: the SVD init equalizes directions)",
   "rsa_E0_vs_random": -0.0009,
   "b_prior_range": [
    -9.156,
    -2.14
   ]
  }
 }
}