Decision Model
Overview
ATA evaluates each candidate action using three core metrics, then applies safety gates to produce a final decision.
Core Metrics
| Metric | Description | Range |
|---|---|---|
| Confidence | Policy network's confidence in the action | 0.0 - 1.0 |
| Coherence | Agreement between belief model and planner | 0.0 - 1.0 |
| Consistency | Logical consistency (1 - contradiction_rate) | 0.0 - 1.0 |
Scoring
Each action's score is computed as a weighted sum:
score = (confidence × weight_c) + (coherence × weight_h) + (consistency × weight_k)Default weights are 0.33 each (equal weighting).
Thresholds
Actions must meet minimum thresholds to be considered:
conf_min- Minimum confidence (default: 0.25)coh_min- Minimum coherence (default: 0.25)cons_min- Minimum consistency (default: 0.25)
Actions failing any threshold are excluded from selection.
Selection Process
- Filter actions by thresholds
- Compute weighted score for remaining actions
- Select highest-scoring action
- Apply safety gates
- Return final decision with audit trail