# Does the measured b-value depend on time of day?

Japan box (30–46°N, 128–146°E), M ≥ 3.5, 2010–2025, split by local (JST) origin hour.

| Window | Mc | N | b | SE |
|---|---:|---:|---:|---:|
| Day (06–18) | 4.4 | 5,329 | 1.167 | 0.016 |
| Night (18–06) | 4.5 | 4,830 | 1.272 | 0.018 |

Difference (night − day): **+0.105 ± 0.024, z = 4.36.**

## The result is significant and the mechanism is not

The prediction was straightforward: cultural noise is lower at night, more small events are detected,
completeness improves, and apparent b rises. Half of that happened — b *is* higher at night.

But completeness moved the wrong way. Mc is **higher** at night (4.5 vs 4.4), not lower. If detection
were better at night, Mc should fall. So the simple detection-bias account does not fit these numbers,
and the difference cannot be attributed to it on this evidence.

Three candidate explanations remain, none tested here:

1. **Mc estimator instability.** Maximum curvature is a crude mode estimator; with ~5,000 events and
   0.1-magnitude binning, a one-bin shift is not informative on its own.
2. **Aftershock timing.** Sequences are not declustered. If large sequences fall disproportionately in
   one window, both b and Mc move together for reasons unrelated to noise.
3. **Genuine catalogue heterogeneity** in how the network processes events at different hours.

## Why it is still worth recording

A 0.1 shift in b is larger than most regional differences reported in the literature (Japan 1.30 vs
California 1.05 in the companion file came from the same estimator). If b varies by 0.1 with the clock,
then cross-region and cross-epoch comparisons inherit a bias of that order unless detection is controlled.

That is a caution, not a discovery. The observation is real; the explanation is open.

## Limits

- One region, one threshold, one estimator.
- No station-level noise data; local hour is a proxy.
- No declustering.
- USGS `time` is origin time, not detection time.
- The analysis was run once, not pre-registered.