Skip to content

Calibration

Jump to bottom
Chuyue Wang edited this page May 19, 2026 · 3 revisions

Calibration

Purpose

Calibration establishes personal baseline distributions used by state scoring and trigger personalization. Cortex 0.2.0 keeps the original seed capture and adds richer baseline statistics for robust z-score/percentile logic.

What Changed in 0.2.0

  • Baseline artifacts now include per-metric distribution stats:
    • mu, sigma, p10, p90
  • Baseline loader includes migration logic so legacy baseline JSON files still load.
  • Blink threshold personalization support is available from EAR sample percentiles.
  • Config includes rolling re-baseline and decay knobs for longitudinal adaptation.

Running Baseline Capture

cortex-calibrate --duration 120

Optional simulation mode:

cortex-calibrate --simulate

Output Schema (Key Fields)

UserBaselines now persists both legacy scalar baselines and additive distribution metadata:

{
  "hr_baseline": 72.0,
  "hrv_baseline": 50.0,
  "blink_rate_baseline": 17.0,
  "mouse_velocity_baseline": 500.0,
  "metric_distributions": {
    "hr": {"mu": 72.0, "sigma": 4.2, "p10": 66.0, "p90": 78.0},
    "hrv": {"mu": 50.0, "sigma": 9.5, "p10": 38.0, "p90": 63.0}
  }
}

The full UserBaselines schema (including hr_std, mouse_variance_baseline, shoulder_neutral_y, resp_baseline, circadian_hr_cosinor, and rolling/decay knobs) lives in cortex/libs/schemas/state.py.

How Baselines Are Used

  • Pulse/HRV sub-scores use personalized z-score style logic when distribution stats exist.
  • FLOW/HYPER rules reference personalized baseline bands before fallback heuristics.
  • Stress integral normalization can use baseline variance (sigma) rather than absolute ms deficit.
  • Trigger threshold adaptation and dismissal modeling use feedback tied to personal history.

Compatibility

Legacy files without metric_distributions are auto-migrated in runtime load paths. No manual data migration is required.

Clone this wiki locally