Purpose
A lightweight, non-invasive annotation layer for UI/interpretation. It converts an alignment score a_out ∈ (-1,+1) into confidence bands and simple event flags. It never alters the deterministic evaluator.
Score → confidence
- Map to
[0,1]:conf = (a_out + 1) / 2 - Bands (default):
LOWifconf < 0.30,MIDif0.30 ≤ conf ≤ 0.70,HIGHifconf > 0.70
Event flags (ASCII, per day)
Let v = |v_deg_per_day| (from central differences on the unwrapped series) anddist_to_cusp_deg = min_k | (wrap360(L_pred_deg) % 30) − 30*k |.
flag_station := ( v ≤ v_thr )flag_cusp := ( dist_to_cusp_deg ≤ d0 )- Optional UI label:
flag_lowconf := ( conf < 0.40 )
Advisory escalation (UI-only)escalate := flag_lowconf or (flag_station and flag_cusp)
Use for gentle emphasis in displays (e.g., “near-station near-cusp, proceed thoughtfully”).
Per-body thresholds
Use the baseline set from the previous page (3.2) — v_thr (deg/day) and d0 (deg) for each body. Keep these constants identical across train/test/runtime for stable labeling.
Implementation notes
- Bands/flags are metadata. They do not change
L_hat_deg. - Compute
vfrom the unwrapped model track; display angles use one wrap:wrap360(y). - Document any threshold changes with the time span and timestamp used.
- For very slow outers, flags will naturally fire less often for
flag_cuspand more forflag_station; that is expected.
Navigation
Back: SSM-JTK – Results to Date — Artifacts & manifest (concept, no downloads) (3.5)
Next: SSM-JTK – Results to Date — Benchmark cross-checks (observation-only; optional) (3.7)
SHUNYAYA 