RQ-002State: Evidence gatheringConfidence: MediumWorld ModelsModel Evaluation
Which DAWM Phase 7 regressions are architectural rather than tuning noise?
Hypothesis
The recurring regressions likely come from state-abstraction pressure, not a single optimiser or dataset setting.
Evidence
- Phase 6/7/8 diagnosis logs
- Ablation windows
- Negative transfer runs
Experiments
- Phase-local rollback matrix
- Representation bottleneck probe
- Short-horizon control eval
Results so far
Regression clusters are stable across two unrelated parameter sweeps.
Current belief
Preserve the Phase 6 baselines as a control while Phase 7 is re-cut around smaller state transitions.