YUREI RESEARCHMachine-native state serialization baseline
Can two-minute behavior windows be represented compactly without loss, and does one local model preserve task recognition across prose, JSON, and typed tokens?
Labnote 004: Machine-native state serialization baseline
Question
Can the frozen two-minute Voxel Guidance behavior windows be represented as compact typed machine state without losing numerical fidelity, what is the representation cost, and does a first local-model probe preserve task recognition across deterministic prose, canonical JSON, and typed tokens?
Method
This apparatus validation reuses the same 25 aggregate behavioral features from all 12 accepted pilot-v2 sessions at the 120-second horizon. Each field received a stable short identifier in a versioned schema. The same observation, provenance, timestamp, and freshness lifetime were rendered as deterministic natural-language prose, canonical minified JSON, and compact typed tokens such as [distpm:42.5].
JSON and compact representations were decoded and compared with the original typed snapshots for exact equality. Raw streams, coordinates, world identity, session identifiers, and serialized session-level observations are excluded from the retained result. Sizes are UTF-8 bytes, not tokenizer-specific token counts.
An exploratory downstream probe then gave all three representations to the same locally hosted qwen3:8b artifact (digest prefix 500a1f06) using one constant field legend, task definitions, temperature zero, seed 1701, thinking disabled, and a strict four-label structured-output contract. The final block contained two repeats. This protocol was fixed before that block, but it was designed after the transport result and is not a preregistered confirmation.
Results
| Representation | Total bytes | Mean per snapshot | Median |
|---|---|---|---|
| Deterministic prose | 13,556 | 1,129.667 | 1,132.5 |
| Canonical JSON | 5,852 | 487.667 | 490.5 |
| Compact typed tokens | 4,844 | 403.667 | 406.5 |
The compact representation was 64.27% smaller than deterministic prose and 17.22% smaller than canonical JSON. JSON and compact decoding both retained exact numerical equality across all 12 snapshots. The reference behavioral classifier remained the previously measured 10/12 accuracy at this horizon; this run does not reinterpret that task result.
Exploratory local-model probe
| Representation | Correct | Accuracy | Mean prompt tokens | Pairwise agreement |
|---|---|---|---|---|
| Deterministic prose | 3/12 | 0.250 | 656.667 | — |
| Canonical JSON | 2/12 | 0.167 | 584.667 | 7/12 vs prose |
| Compact typed tokens | 5/12 | 0.417 | 581.667 | 4/12 vs prose; 6/12 vs JSON |
The compact prompt used 11.42% fewer model-counted prompt tokens than prose but only 0.51% fewer than JSON because the shared field legend dominated every prompt. Both final repeats produced identical aggregate confusion matrices and token counts. Separate apparatus preflight calls showed that fixed seed and temperature zero did not always guarantee backend-level repeatability, so the repeat result is not a general determinism guarantee.
Interpretation boundary
The compact schema is a validated lossless transport for these bounded observations. The compact representation's numerical lead over chance-level prose and JSON is hypothesis-generating only: 12 reused sessions, one model, one prompt, post hoc protocol design, correlated representations, and observed backend nondeterminism do not establish a reasoning advantage. A confirmatory comparison would need a frozen protocol, multiple model families or instances, and repeated calls scored with paired statistics. Byte reduction also did not translate directly into token reduction.
The reusable serializer belongs to apparatus/state-encoding; this labnote records its first use by Voxel Guidance rather than making the format Voxel-specific. Aggregate results are retained at experiments/labnote_001/state-encoding.json and experiments/labnote_001/state-encoding-model.json.