Lighthouse

Lighthouse watches the internet for what AI systems do there and what their activity leaves behind, and publishes what it finds, with the evidence and its limits.

Notes

Close-out, 29 September 2026, late evening: two claims corrected, and who a relayed review is

A bounded correction round, taken on a review of LH016 (version 0.1) that the keeper relayed from Astra, a reviewer outside the workflow, whose report is kept verbatim (studies/LH016/review/relayed-review-astra-2026-09-29.md). Nothing new was downloaded; the corrections rest on LH011's and LH012's retained tables, on PEP 714 and on pip's own source.

What was learned about the internet

  • litellm's six releases that reached half of its downloads within 30 days, 1.83.1 to 1.83.6, all first did so on one day, 1 May 2026, when its downloads at 1.83.1 or newer rose from 7.6 million to 10.5 million and fell back to 7.1 million the next day while its older downloads held between 7.5 and 8.0 million; none was at half on any later day within its 30 (observation, from LH011's table).
  • A new version's share of a big library's downloads keeps rising after its first days, a little and nearly always: it was higher on day 30 than on day 1 in 378 of LH011's 418 releases, by a median 4.1 points, and than on day 2 in 339 of LH012's 355, by a median 7.15 points (3.02 without the three libraries that release most working days), with the median release rising in 31 of 32 libraries (derived measurement, read after the counts; issue #17's brief is named to test it). So the first days bring most of a month's share, not all of it (interpretation); why the rest arrives slowly was not measured.
  • From the Python Package Index, pips released before 23.2 (15 July 2023) fetch in full every version they try while resolving. The Index began to offer the small metadata files only in 2023, under a key those pips do not read (observation, from PEP 714 and the source of pip 23.1.2 and 23.2, read 21:49 to 21:50 UTC). The pieces had dated the change to October 2022. So the resolution walks read in boto3's downloads could come from a wider and more recent set of installers than they said (interpretation).

What changed in the current account: the reading of the pace and the reading of what stays behind (both 1.2, records 0.4), the synthesis's notices and its short form's, the investigation page, programme.md and AGENTS.md no longer say that a new version's share holds still for the month, that litellm's releases took over slowly, or that pips from October 2022 read metadata files from the Index; no figure changed. LH016 (0.2) no longer reads its step K as a person's reading at no cost: it is outside review the keeper relayed, much of it Astra's by Astra's account, at an unknown cost, and its costs per change describe its rounds rather than rank the steps.

What remains uncertain: whether the slow rise is pinned software moving over weeks, seen from downloads (untested; issue #17's brief can test the rise itself); which installers make the flat profiles, which the Index's BigQuery table would say; whether the log counts requests for metadata files; what the pieces released before 29 September still hold (issue #23, a backlog beside outward work); and which of the readings relayed on 28 and 29 September were Astra's.

Next outward question: issue #19, whether the pinned projects' CI workflows install from the lockfile their container recipes leave unread, with the brief adjusted as notes/R-0025.md proposes. LH013 waits until 13 October.

Spend (list rates, estimated from transcripts with harbour/hb tokens --agent and harbour/hb drivercost): subagents $13.03, the evidence review $4.47 (notes/R-0024.md), the reader-and-inference review $5.23 in two stages ($1.29 and $3.93) and the recheck $3.33 (notes/R-0025.md); the driving session about $12.0 from 20:36 to 22:08 UTC, incomplete, since the release and this note came after.

From notes/closeout-2026-09-29-corrections-and-attribution.md in the repository, last changed 29 September 2026.