YUREI RESEARCHSemantic namespace token ablation
Can a semantic namespace token replace the full field dictionary when a local model interprets compact behavioral state?
Labnote 005: Semantic namespace token ablation
Question
Can a human-readable namespace such as [ns:voxel-guidance.behavior-window.v1] give a local model enough conceptual context to interpret abbreviated machine-state fields without the full field dictionary?
Method
This exploratory ablation reused the 12 accepted pilot-v2 behavior windows at 120 seconds and the qwen3:8b artifact from Labnote 004. The model, digest, task definitions, structured four-label output, temperature zero, seed 1701, and thinking-disabled setting remained fixed. Four compact representations were compared:
- the complete field-ID dictionary and no semantic namespace;
- a semantic namespace and no dictionary;
- a semantic namespace plus a shorter feature-group grammar; and
- an opaque schema digest with neither namespace nor dictionary.
The final block ran every variant twice. Only aggregate confusion matrices, prompt-token counts, timings, and repeat checks were retained. The variants were designed after Labnote 004 and are not confirmatory tests.
Results
| Variant | Correct | Accuracy | Mean prompt tokens |
|---|---|---|---|
| Full field dictionary | 7/12 | 0.583 | 579.667 |
| Semantic namespace only | 3/12 | 0.250 | 378.667 |
| Namespace + group grammar | 1/12 | 0.083 | 432.667 |
| Opaque schema only | 3/12 | 0.250 | 366.667 |
Namespace-only and opaque-schema control produced exactly the same 12 predictions. The semantic namespace therefore supplied no measurable zero-shot benefit in this run while adding about 12 prompt tokens. The short group grammar did not preserve the full dictionary's semantics and performed worse. Removing the dictionary reduced the prompt by 34.68% relative to the full-dictionary condition, but also removed its numerical task-recognition advantage.
Both repeats retained the same accuracy for every variant. The three reduced- context variants retained identical aggregate confusion matrices; the full dictionary changed its confusion matrix while remaining 7/12, reinforcing the known backend nondeterminism boundary.
Interpretation boundary
The proposed namespace token did not replace explicit field semantics for this model and task. A namespace may still be operationally valuable for schema routing, validation, caching, or systems that have learned its meaning through training or repeated context. This result only rejects the stronger zero-shot idea that a readable namespace header by itself would let this model recover the abbreviations' meaning.
The full dictionary's 7/12 result is exploratory rather than a validated model score: the sample is small, reused, single-model, and post hoc, and its detailed confusion pattern was not perfectly repeatable. Aggregate results are retained at experiments/labnote_001/namespace-ablation.json.