A Unix timestamp is an offset from an epoch, expressed in a unit. The bare number does not say whether the producer meant seconds or milliseconds. A count-the-digits heuristic happens to work for recent dates, but a heuristic is not a data contract. The Explorer never applies one.
Start from a known instant
Input: 1780496820 · Unit: seconds · Timezone: UTC
- Expected ISO: 2026-06-03T14:27:00.000Z
- Expected Unix milliseconds: 1780496820000
- Expected weekday: Wednesday
Run the explicit-seconds example →
Now switch the unit to milliseconds without touching the number. You have described a different instant, one close to January 1970. The tool follows the unit you chose. It does not silently override your selection because the resulting year looks surprising — the surprising year is the diagnosis.
The normalized model is milliseconds
Every supported input normalizes to integer Unix epoch milliseconds. Seconds accept up to three decimal places; milliseconds must be integers. One normalized value drives every output, so the ISO string and the numeric outputs cannot drift apart through different rounding paths.
- Input
-0.125seconds → milliseconds-125, ISO 1969-12-31T23:59:59.875Z - Input
1780496820123ms → seconds1780496820.123, ISO 2026-06-03T14:27:00.123Z
Excess precision is rejected, not rounded. Scientific notation, Infinity and NaN are refused outright. The supported range matches JavaScript’s Date model at ±8,640,000,000,000,000 ms, and leap seconds sit outside the model, as they do there too.
UTC names an instant. A local calendar time names a wall-clock reading that may occur twice or never. The tool keeps those two ideas in separate fields so you never have to guess which one a producer meant.
Where the heuristic breaks first
Digit-counting fails on purpose in the ranges where precision matters most. Ten-digit seconds and thirteen-digit milliseconds look distinct today, but sub-second timestamps in seconds (three decimal places) and legacy systems emitting odd widths blur the boundary immediately. The failure is not hypothetical: ingest pipelines that guessed the unit have produced duplicate events an entire epoch apart.
The durable fix costs one field. Producers add unit or emit an ISO 8601 string, consumers stop guessing, and the Explorer stays useful as the verification step that both sides can run locally during development.