# Cohort baseline for the pre-removal download figures (v2)

> **Erratum (v2, 2026-07-04, same day).** v1 framed the captured 666 as "~2 days" of accumulation for v7.2. A
> platform timestamp obtained the same day (TikTok Post analysis: image posted Jun 15, 2026, 4:42 PM — the day of
> the v7.2 upload) supersedes the recollection that produced that framing, and the 666 is best read as the
> record's cumulative counter at that moment. The corrected anchors: **666 cumulative at 2026-06-15 16:42**
> (platform-timestamped) and **1,000+ at removal, 2026-06-19** (contemporaneous issue body, unrefuted) — a delta
> of **>=334 downloads in <=4 days**. The cohort table below is unchanged; the Reading and Conclusion are
> restated on the corrected anchors. v1 is preserved in repository history.

**Measured 2026-07-04, unauthenticated Zenodo REST API.** Raw sample: `zenodo-baseline-sample.json` (checkable; re-runnable from the query below).

**Question.** Is 666 downloads in ~2 days (captured, this set) / 1,000+ in ~4 days (asserted contemporaneously in zenodo/zenodo#2606, day of removal) "a lot by Zenodo's standards"?

**Method.** Cohort = all Zenodo records of type `dataset` with `publication_date:[2026-06-14 TO 2026-06-16]` — the same publication window as the registry's v7.2 (June 15). Cohort population: **1,791**. Sample: first 200 by `mostrecent` within the window (API page cap 25 unauthenticated; 8 pages). Stats read 2026-07-04, i.e. after **~19 days** of accumulation for the cohort, versus **~2 days** for the registry's captured figure — the comparison is therefore conservative by roughly a factor of ten in accumulation time, in the cohort's favor.

**Findings (n=200).**

| statistic | this-version downloads | all-versions downloads |
|---|---|---|
| median | 7 | 6 |
| mean | 43.0 | 21.1 |
| p90 | 53 | 43 |
| p95 | 116 | 77 |
| p99 | 804 | 397 |
| max | 3,087 | 561 |
| ≥ 666 after ~19 days | 2/200 (1.0%) | 0/200 (0.0%) |

**Reading (v2).** The four-day delta alone (>=334 downloads, Jun 15->19) is ~48x the cohort's nineteen-day median (7) and exceeds its nineteen-day p95 (116) at the version level. The cumulative 1,000+ at removal exceeds the cohort's version-level p99 (804) and the all-versions maximum observed (561). On any field mapping, the removed dataset's usage sat at the extreme top of its publication-week cohort.

**Limitations, on the record.** (1) The sample is the window's 200 most recent, not a random draw; the window is only three days wide, limiting ordering bias, but this is a convenience sample. (2) The Zenodo UI counter captured at 666 is not field-labeled in the interface; API naming (`downloads` = this-version, `version_downloads` = all-versions) means the captured figure maps to one of the two columns above — it clears the 99th percentile threshold in either mapping at the version level and exceeds the observed maximum at the all-versions level. (3) Bot and crawler traffic is unmodeled on both sides of the comparison. (4) Cohort stats were measured post-hoc on 2026-07-04 and will drift; the raw sample is preserved for exactly that reason.

**Conclusion, bounded (v2).** By the platform's own contemporaneous cohort, the removed dataset's usage was in the top ~1% of same-week datasets, and its final four days alone outpaced what 95% of that cohort accumulated in nineteen. "A lot by Zenodo's standards" is hereby a measurement, not an impression — and this note's own correction trail is part of the measurement's warrant.