# Deviations from PREREG.md

Copied verbatim from research/model-lifespans/README.md (written after the results, as PREREG.md said it would be). PREREG.md itself was committed at 1c7a4b74c8, 2026-09-26 08:32 UTC, before the extractor ran; the repository is private, so the commit is named here rather than linked.


1. **The Archive half was not done as planned.** The plan used Internet Archive captures for the
   period before 2024-10-05 and to cross-check the daily record. Four 2023 captures were fetched
   before web.archive.org began resetting every connection from the build container (repeated
   `curl: (35) Recv failure: Connection reset by peer` with backoff up to two minutes). The
   cross-check was instead done against a second daily scraper (jvrck/openrouterlist), which is a
   stronger check (daily, independent, same window) and is reported as such. The pre-2024 era is
   reduced to one descriptive fact per capture.
2. **Q5's "kept" window** was specified as "gone within 7 days after that date"; the code also
   accepts a death one day before the date (a snapshot taken the morning after a midnight removal is
   dated the next day; so is one taken before a removal that same day). Early, late, overdue and
   pending are all reported separately so nothing is hidden by the window.
3. **Q7** was specified as descriptive. Measuring it from each id's birth turned out to be biased
   (immortal time), so it is measured from the day of picking, against every free id listed on the
   same days. Both numbers are reported.
4. **The cohort analysis** was not in the pre-registration and is labelled exploratory everywhere
   it appears.