Learning path
Observed next-morning outcome, coloured by that night’s intervention.
Thirty nights · one decision at a time
Codex learns without looking ahead.
Every evening uses only earlier nights, compares today with similar past days, observes whether heart rate and movement settle during the intervention, then learns from the next morning.
What the first month suggests
A deterministic proof of concept, not your measurements. Recommendations become more personal only after balanced evidence exists.
Establish a baselineTry each routine once. No personalization claim.
Repeat for reliabilityA second observation prevents one-night conclusions.
Balance the evidenceBring every routine to four nights before choosing a leader.
Personalize by contextUse similar past days and preserve occasional comparisons.
Replay any nightly decision
Choose a night. The left side shows exactly what Codex knew at 22:05; the right side reveals what happened only afterward.
Night 11
Decision made from Nights 1–10 only
10 earlier nights availableKnowledge boundary: Night 10. Night 11’s outcome is still hidden.
Codex decision at 22:05
Observed during and after the routine
Did the person actually settle?
Heart rate · bpmMovement · events/min
Compare like with like
Days are matched on steps, active minutes, tension, caffeine, external noise, meal timing and previous sleep—then different routines are compared.
Selected comparison
context similarity
The next ten-night learning plan
Confirm the emerging leader without abandoning controls, measure the 30-minute response, and use at most one tiny opt-in nudge.