Why bazi calculators disagree

The clearest case is a birth at 23:30. Some calculators treat 23:00 as the start of the next day and advance the day pillar; others hold it until midnight. Here is what came back for the same moment, 15 June 1990 at 23:30.

One birth, one moment, and the character your entire chart is read from comes out as Water or as Metal depending on which tool you happened to open. Among calculators that take a position the split is 2 to 5, and Jade Almanac is in the group of 2. We roll the day at 23:00, our reasoning is on the methodology page, and it is a choice rather than a standard. Most of the field goes the other way.

Two rows deserve a second look. prokerala.com makes the birth city a required field and then returns an identical chart for Kashgar and Beijing, two time zones apart; a reader who carefully typed their birthplace would reasonably assume it was doing something. And the top organic result for “bazi calculator” has no engine of its own at all. It embeds a white-label plugin, which is what we actually measured. A list of calculator websites is not a list of calculators.

Most people are unaffected. These conventions only bite near a boundary, and away from one every calculator here agrees. But if you are near one, it is worth knowing which:

• Born between 23:00 and midnight. Your chart is genuinely ambiguous. Cast it both ways, as a day-master Water and as a day-master Metal reading, and see which one describes a life you recognize. No calculator can settle this for you.
• Born in late January or early February. Check whether your birthday falls between Chinese New Year and Li Chun. If it does, your year pillar and your zodiac animal depend on which rule the tool used.
• Born far from your time zone’s meridian. Western China, western Spain, and the western edge of any wide zone can be an hour or more off solar time. That moves the hour pillar, which is a quarter of the chart.
• Born in summer, anywhere that runs daylight saving. Subtract the hour before you enter the time, or use a calculator that asks. This one is not a matter of interpretation.

Each probe is a birth moment built to isolate exactly one convention, paired with a control that must not move. If a control moves, the capture is wrong rather than the tool. Solar-term instants were checked against the National Astronomical Observatory of Japan’s published tables rather than against the library our own engine uses, so the probes are not graded by the thing they are testing.

Every value here is what another implementation returned, recorded and frozen. None of it is regenerated from our own engine, because a dataset that computed its own expected answers would only be measuring itself.

Three capture failures happened during the run, and all three produced output that looked completely reasonable. A dropdown whose values were not zero-padded silently fell back to “hour unknown” and returned a plausible midnight chart. A checkbox left ticked from an earlier probe quietly changed every later reading. One tool enforces a daily request cap, after which it serves a chart for today instead of the date submitted, and says nothing about it.

Hence the rules this capture ended up needing: read the result page instead of scraping the answer, since one calculator prints its entire correction in plain text and reading it settled two questions more firmly than any amount of probing; check every input against the page’s own echo of what it received; carry no state between probes; and re-run a control in a fresh session before calling any row finished.

Where a tool could not be measured the cell is empty and the reason is recorded. Two sites block automated access, and neither was worked around.

• This is a snapshot from 2026-08-01. Calculators change without announcing it, and several of these are actively developed.
• Some rows are partial. Where a tool’s four conventions were established but not every probe was run, the convention table is complete and the observation matrix is not.
• feng-shui-institute.org and bazi-calculator.com agreed on every probe where their inputs were comparable. One takes hour-granular input, so this run cannot establish whether they are independent implementations.
• Probes B1 and B2 fall on 1990-06-15, which sits inside China’s 1990 summer-time window. Against a bare clock reading, which is how every implementation here was probed, they isolate the day rollover cleanly. But a tool that removes daylight saving from a Chinese birthplace would pull 23:30 back to 22:30 and stop testing the rollover at all. A re-run that supplies birthplaces should move this probe off those months.
• The roster came from US search results for three queries. It samples what an English-speaking reader actually finds, not the whole of bazi software.
• Nothing here speaks to whether bazi predicts anything. It measures software behavior and calendar convention, and that is all it measures.