Dootdoot Asset Spec

The primary committed runtime asset is assets/dootdoot_asset_v1.doot. The .doot file is a Protocol Buffers message that carries every runtime byte needed for tokenization and the semantic lookup table; normal builds embed it with include_bytes!. Since VOICE_V12 a second, independent sidecar asset — assets/dootdoot_pos_v1.doot, the baked noun/verb class table — is committed and embedded alongside it (its layout is specified at the bottom of this document); the semantic asset itself remains at spec v1.

Protobuf Message

syntax = "proto3";

package dootdoot.asset.v1;

message DootAsset {
  uint32 spec_version = 1;
  uint32 token_count = 2;
  uint32 axis_count = 3;
  repeated float axis_scales = 4;
  float weight_scale = 5;
  SquashFunction squash_function = 6;
  repeated SquashAxisStats squash_stats = 7;
  bytes model_sha256 = 8;
  bytes tokenizer_sha256 = 9;
  bytes pca_sha256 = 10;
  bytes tokenizer_json = 11;
  bytes token_records = 12;
}

message SquashAxisStats {
  double mean = 1;
  double standard_deviation = 2;
}

enum SquashFunction {
  SQUASH_FUNCTION_UNSPECIFIED = 0;
  SQUASH_FUNCTION_TANH_Z_SCORE = 1;
}

Required Values

• spec_version is 1.
• axis_count is 4.
• axis_scales has four positive finite f32 values.
• weight_scale is positive and finite.
• squash_stats has four entries, each with a finite mean and positive finite standard deviation.
• model_sha256, tokenizer_sha256, and pca_sha256 are 32 bytes each.
• tokenizer_json is the complete HuggingFace tokenizer configuration used at runtime.
• token_records has exactly token_count * 10 bytes.

Per-token record

token_records is a compact fixed-width byte field inside the protobuf message. Each token record is 10 bytes, little-endian:

Offset	Size	Field	Type
0	8	projected PCA axes	4 * i16
8	2	pooling weight	i16

The compact record encoding keeps the semantic table small while protobuf provides the forward/backward-compatible envelope for the runtime asset.

Freeze status

VOICE_V1 is locked. The Phase 7 voice tuning accepted the final sample-affecting synthesis constants, T-52 squash finalized the tanh z-score mapping without regenerating the asset, and T-53 learnability spread validated the final semantic/audio cluster separation.

Any future change that alters one or more rendered output samples requires a new voice identifier plus regenerated golden fixtures. VOICE_V2 was the first post-V1 sample-affecting contract bump.

VOICE_V12 class-table sidecar (dootdoot_pos_v1.doot)

The noun/verb class table (FR-114) is a separate, deliberately simple little-endian binary — not protobuf — parsed by dootdoot_core::PosTable and serialized by xtask::serialize_pos_table from the same layout constants, so writer and reader cannot drift.

Header (78 bytes):

Offset	Size	Field	Value / type
0	8	magic	DOOTPOS1
8	2	spec version	u16 = 1
10	32	ranking-corpus SHA-256	pinned in [pos]
42	32	tagged-counts SHA-256	pinned in [pos]
74	4	entry count	u32

Entries follow, strictly sorted by surface bytes (lookups binary-search):

Offset	Size	Field	Type
0	1	class	u8: 0 = noun, 1 = verb
1	2	surface length	u16
3	n	surface	UTF-8 bytes

Parsing validates the magic, spec version, class bytes, UTF-8, strict sort order, and exact total length. Absent words classify Other. Regeneration is documented in asset_regeneration.md; the committed sidecar must reproduce byte-for-byte from the committed assets/pos/tagged_counts.tsv snapshot.