Food-to-Sleep Pilot — September 2026

Interim summary

The first ten outcome nights produced a mean Garmin Sleep Score of 67.8, a median of 69.5, and a range of 46 to 81. The historical Sleepgenic baseline score is 67.5, so the pilot average is close to the earlier long-run reference despite large night-to-night variation.

The strongest early signal is a natural alcohol perturbation. The single sake-exposure night produced the lowest score in the pilot and simultaneous deterioration across several recovery markers. The following alcohol-free night showed a broad rebound. This is a strong within-person observation, but one exposure does not establish a causal effect size.

A second signal is methodological: sleep architecture and autonomic recovery should not be collapsed into one number. Several nights showed strong HRV/resting-heart-rate recovery with weak Garmin score or stage architecture, while another night showed strong deep/REM sleep with weaker autonomic recovery.

Observed nights

Garmin sleep stages are consumer-wearable estimates. Missing values are shown as missing and are not imputed.

Signals so far

1. The alcohol night is the clearest perturbation

Compared with the nine non-alcohol outcome nights, the sake night had a Sleep Score of 46 versus a non-alcohol mean of 70.2. HRV was 25 ms versus a non-alcohol mean of 43.9 ms among nights with HRV recorded. Resting heart rate was 60 bpm versus 55.1 bpm. Awake time was 87 minutes versus 44.9 minutes. Body Battery gain was +34 versus +60.7 among nights with that metric available.

Those signals moved in the same direction at the same time. That convergence is more informative than the headline score alone. It remains a single natural experiment, not a causal estimate.

2. Meal timing is not testable yet

Most recorded final caloric cutoffs cluster tightly between 6:30 PM and 7:15 PM. That is useful experimental control, but it leaves too little timing variation to estimate whether an earlier cutoff improves sleep. A September 16 late mooncake exposure at 9:00 PM is recorded as a pending perturbation and is excluded from this interim outcome analysis.

3. Food-specific conclusions are premature

Rice, quinoa, fish, chicken, vegetables, fruit, noodles, pork, beef, tofu, restaurant food, and mixed meals all appear in the pilot, but portion size, activity, sodium, fat, meal complexity, and timing change together. No individual food has enough repeated, controlled exposures to support a defensible ranking.

Alcohol perturbation: one night, many signals

The sake exposure was approximately 300 mL at 14% ABV, about 2.4 U.S. standard drinks. The following Garmin night showed Score 46, HRV 25 ms, resting HR 60 bpm, 87 minutes awake, 14 minutes REM, and Body Battery +34.

The next alcohol-free night rebounded to Score 77, HRV 40 ms, resting HR 54 bpm, 11 minutes awake, and Body Battery +62. The immediate reversal strengthens the interpretation that the alcohol night was unusual within this short run, while still leaving open other explanations such as unmeasured context and normal night-to-night variability.

Sleep quality and recovery quality are different outputs

The September 11 morning result scored only 54, with just 5:16 total sleep and 10 minutes of Garmin-estimated REM. Yet HRV was 46 ms and resting heart rate was 51 bpm—the strongest autonomic combination in the pilot to that point.

By contrast, the September 16 morning result had 1:48 deep sleep and 1:25 REM, but HRV was 35 ms and resting heart rate was 60 bpm. The architecture looked strong while autonomic recovery looked less impressive.

This supports a two-output analytical model for future Sleepgenic work: Sleep Quality (duration, stages, fragmentation) and Recovery Quality (HRV, resting/overnight heart rate, Body Battery and related physiology). Garmin's composite score remains useful, but it should not replace those separate layers.

What we cannot say yet

• We cannot rank foods or declare a sleep-promoting dinner from this sample.
• We cannot isolate meal timing because most cutoffs are tightly clustered.
• We cannot separate exercise modality from training dose, heat, meal composition, and other context on the current nights.
• We cannot treat wearable sleep stages as clinical measurements.
• We cannot generalize an n=1 result to other people.

Next phase

The protocol remains intentionally simple: photograph the meal, record the actual portion and any later food or alcohol, record the final caloric cutoff, capture previous-day activity, and pair the next morning's Garmin metrics to that exposure. The master log preserves missingness and marks photo-verified, self-reported, and retrospective entries separately.

The next analytical threshold is repetition: enough observations of similar meal types, timing windows, activity modalities, and naturally occurring perturbations to test whether the early patterns recur. Sleepgenic will continue to publish signals first and conclusions only when the data earns them.