As of October 2026｜Across five reaction rounds, how far is the fastest from the middle?

In 21091 eligible five-round records, the median of each record's middle-round time minus its fastest-round time is 22 ms. The difference can already occur within one test, before comparing people or devices.

This article uses a fixed snapshot read on 17 October 2026 at 03:54:35 Beijing time. Only human_benchmark_records is counted: 248024 records, dated from 23 December 2025 to 17 October 2026, with the latest at 01:23:55 that day. The legacy human_scores table's 6507 rows are not added. These are site submissions, not a random population sample.

All 21850 reaction records have five round values. Their stored-score median is 256 ms. For this comparison, the stored score and every round must lie from 100 to 1000 ms inclusive. This is an editorial comparison window, not a rule declaring other records invalid. A total of 759 records, 3.47%, fall outside the final selection; 173 are excluded by the stored-score condition alone.

In all 21091 retained records, the stored score is within 1 ms of the rounded five-round mean. That supports an approximate mean interpretation for this subset, without proving identical scoring across all historical versions.

The 22 ms difference is calculated inside each record before taking a median. It need not equal 249 minus 223, which is 26 ms: subtracting two group medians is a different operation.

There are 2637 records with middle-minus-fastest at least 50 ms, or 12.50%. Another measure, slowest-minus-fastest, reaches at least 100 ms in 5199 records, or 24.65%. The data cannot separate attention, equipment and other environmental causes.

The stored-score window contains 21677 records and 7005 labels. First and last are ordered by timestamp then record ID; first means the earliest eligible record visible in this snapshot, not a person's first-ever attempt. Label identity has the limitations described above.

Records are grouped by account ID when present, otherwise by guest name, with the two namespaces kept separate. A label is not a verified person: people may change labels and labels may be shared. Counts from different tools overlap and cannot be added as a total population.

Best records are selected minima of stored scores, not fastest individual rounds. More attempts provide more opportunities to select a low value, so a better best score alone does not prove a training effect.

Keep the overall result, fastest round and range, with the same device and input method where possible. This describes the task more usefully than claiming a universal professional threshold.

Percentiles use linear interpolation. The best record is the lowest stored score within the window. Browser timing includes the device and input path and does not isolate neural conduction.