6.7 Deployment: batch, stream, embedded
Batch recognition.
phi_obs.csv -> ssm_clock_runner.py -> tau_hat_days, confidence
Streaming recognition.
obs_stream.csv (pairs) + dt_days -> ssm_clock_stream_v2.py -> tau_days per tick
Latency & memory (rules of thumb).
Memory ~ O( #channels * S ) # manifest ~ O(1 KB)
N_grid = T_search / step_d
Coarse CPU ~ O( S * #channels * N_grid )
- Speed-ups: raise K before shrinking grid_step_min; use coarse_stride=4 for large
T_search; stack S conservatively (5–7).
Embedded notes.
- Port the one-line kernel and a Brent 1-D minimizer to C; keep double precision.
- Wrap once (inside
angdiff). Avoid trig inside tight loops where possible.
6.8 Minimal API contracts
Single-shot API (function).
inputs:
manifest: channels [{label, period_days, b0_deg, w_deg_per_day}]
phases: dict label->phi_deg for one snapshot (or S snapshots if stacking)
knobs: grid_step_min, stack, stack_dt_days, multistart_k
outputs:
tau_hat_days: float
confidence: float in [0,1)
residuals: dict label->err_deg at tau_hat # optional
Streaming API (tick).
inputs:
manifest, dt_days, phases_row
outputs:
tau_days # printed per tick
conf, residuals # optional
Error reporting (always modulo horizon).
err_days = min( |tau_hat - tau_true| , T_search - |tau_hat - tau_true| )
6.9 Guardrails and privacy
- No wall-time dependency. Keep it strictly phase-in -> tau_hat-out.
- Confidence gates. Use
confthresholds for actions; log low-conf episodes. - ASCII artifacts. Log only CSV/JSON/TXT; version manifests; checksums optional.
- Local compute. Prefer on-device inference; share only derived features if needed (e.g.,
[sin, cos, confidence]).
Navigation
Back: SSM-Clock—Media, Robotics & JTK Interop (6.4–6.6)
Next: SSM-Clock—Determinism, Audit & Provenance (7.1–7.3)
SHUNYAYA 