SSM-Clock—Formats, Determinism & Performance (2.7–2.9)

2.7 Formats and telemetry

Summary CSV columns

periods, noise_deg, seeds, grid_min, pass_count, pass_rate_pct, mean_err_min, p95_err_min

Runner print (console)

tau_hat_days=...  confidence=...

Optional residual dump (debug)

  • Per-channel err_i(tau_hat) in degrees.
  • Emit as CSV under notes/ for audit and post-hoc plots.

Pairs vs wide inputs (reminder)

# Pairs header (robust to label order, handles missing labels)
label,phi,label,phi,label,phi,label,phi

# Wide header (fast; must align with manifest order)
day,week,lunar,solar

2.8 Determinism and reproducibility checklist

  • No randomness at inference. The runner never samples; synthetic bench seeds are logged separately.
  • One wrap rule. Wrap only inside angdiff; never re-wrap intermediates.
  • Manifest frozen. Do not alter period_days, b0_deg, w_deg_per_day after acceptance.
  • Coarse -> refine. Record grid_step_min, multistart_k, stack, stack_dt_days, refine_steps, bracket_mult; keep constant for comparable runs.
  • ASCII artifacts. Keep CSV and TXT under notes/ for audit; avoid binary formats.
  • Horizon parity. Report time errors modulo T_search consistently across runs.

2.9 Performance knobs (safe order)

Recommended order of tuning

  1. multistart_k : 7 -> 9 (robustness up; small cost)
  2. stack : 5 -> 7 with same stack_dt_days (alias suppression up; cost x S)
  3. grid_step_min : 2 -> 1 (precision up; coarse cost ~2x)
  4. coarse_stride : 1 -> 4 (faster coarse by skipping points; tiny risk at very high noise)
  5. refine_steps : 80 -> 120 (marginal improvement; minor cost)

Guideline

  • Prefer raising K (multistart) and S (stack) before shrinking the coarse grid.
  • Widening the refine bracket is last; use only if bowls are unusually broad.

Navigation

Back: SSM-Clock—Acceptance & Longbench (2.5–2.6)
Next: SSM-Clock—CSV Modes & Packaging (2.10–2.11)