dormant_blocks.json

{
  "schema": "v-tokenizers-dormant-blocks-1",
  "source": "v12/dormant_block_map.yaml (C7)",
  "masking_policy": {
    "implementation": "active-slice projection: hidden @ active_embedding.T over the contiguous active range",
    "dormant_rows_in_matmul": false,
    "dormant_rows_in_optimizer_state": false,
    "dormant_strings_in_clean_115M_stream": false,
    "fingerprint_init_preserved_until_activation": true
  },
  "layout_order": [
    {
      "block": "T-core natural language",
      "category": 0,
      "active_at_115M": true,
      "active_at_1B": null,
      "merge_opaque": null,
      "token_count": null,
      "tokens_frozen": null
    },
    {
      "block": "active specials",
      "category": 2,
      "active_at_115M": true,
      "active_at_1B": null,
      "merge_opaque": true,
      "token_count": null,
      "tokens_frozen": null
    },
    {
      "block": "dormant: T-cell",
      "category": 3,
      "active_at_115M": false,
      "active_at_1B": true,
      "merge_opaque": true,
      "token_count": 0,
      "tokens_frozen": false
    },
    {
      "block": "dormant: T-scale",
      "category": 4,
      "active_at_115M": false,
      "active_at_1B": true,
      "merge_opaque": false,
      "token_count": 0,
      "tokens_frozen": false
    },
    {
      "block": "padding",
      "category": 1,
      "active_at_115M": true,
      "active_at_1B": null,
      "merge_opaque": null,
      "token_count": null,
      "tokens_frozen": null
    }
  ],
  "note": "Dormant ids are present in the tokenizer's vocabulary but are excluded from the projection matmul and optimizer state at 115M. Size an active embedding slice from the active blocks, NOT from len(tokenizer). Where token_count is 0 the block is schema-pinned but not yet frozen (filled at TOK-5 once the winning vocab layout is known)."
}