YUREI RESEARCHCurrent-task feature-family ablation
Which observable behavior families carry the 120-second held-seed task-inference signal?
Labnote 003: Current-task feature-family ablation
Question
Which observable behavior families carry the 120-second held-seed task- inference signal: movement, coarse inventory flow, or confirmed block actions?
Method
This exploratory ablation reuses the 12 accepted pilot-v2 sessions and the same leave-one-seed-block-out nearest-prototype classifier as Labnote 002. The 120-second horizon was selected after the prefix analysis because it was the earliest horizon with peak observed accuracy; it is not an independently preregistered endpoint.
Seven feature projections were compared: all 25 features, each of the three families alone, and the full set with each family removed. Every variant used the same 10,000 within-block label permutations. These p-values compare each variant with chance; they do not test a causal difference between variants.
Results
| Feature projection | Features | Correct | Accuracy | Macro F1 | Permutation p |
|---|---|---|---|---|---|
| All | 25 | 10/12 | 0.833 | 0.833 | 0.0015 |
| Block actions only | 8 | 10/12 | 0.833 | 0.833 | 0.0015 |
| Movement only | 7 | 9/12 | 0.750 | 0.760 | 0.0040 |
| Inventory only | 10 | 7/12 | 0.583 | 0.530 | 0.0742 |
| Without block actions | 17 | 9/12 | 0.750 | 0.743 | 0.0090 |
| Without inventory | 15 | 10/12 | 0.833 | 0.833 | 0.0015 |
| Without movement | 18 | 8/12 | 0.667 | 0.676 | 0.0200 |
Block actions alone reproduced the full model's predictions. Movement alone retained a strong held-seed signal. Coarse inventory flow alone did not reliably beat chance, and removing inventory did not change the full result. Removing either block actions or movement reduced accuracy, although this small descriptive ablation cannot establish that either decrement is a stable causal effect.
Interpretation boundary
At two minutes, task recognition appears to depend primarily on how the participant changes the world, with route structure providing a second useful view. The result is not explained by coarse inventory possession or churn. Because the horizon and feature families were examined after collection, this is hypothesis-generating evidence for a future frozen replication—not an estimate of performance on new participants or arbitrary tasks.
Complete aggregate confusion matrices and method metadata are retained in experiments/labnote_001/task-ablation.json. Session-level predictions and raw streams are excluded from this labnote.