Metrics & reporting
Reporting vs inference
CSCT distinguishes:
- Operational reporting criterion (binary pass/fail): e.g.,
withheld_sim ≥ 0.90. - Statistical analysis on the continuous metric (e.g.,
withheld_sim) to avoid threshold artifacts.
This separation is used consistently across EX6–EX9.
Common metrics
- withheld_sim: similarity score for withheld target (primitive or composite).
- mean_trained_sim: sanity check that trained patterns reconstruct well.
- reconstruction MSE: distortion of reconstructed signal.
- transition rate: how often the discrete routing state changes.
- maxP / entropy: confidence and dispersion of gate/code selections.
- cluster metrics (e.g., ARI/NMI): separability of trained vs withheld from internal features.
EX8/EX9 specific
- Hull modes (
IN_HULL,OUT_HULL,RANDOM) define the geometry of withheld targets relative to trained primitives. - Nonparametric tests (Kruskal–Wallis; post-hoc Mann–Whitney) are used when distributions are non-Gaussian or heavy-tailed.