R-0062: reader-and-inference review of LH046 and its proposed "Later evidence" notices

Reviewer: a review subagent in Claude Code (Claude Opus 5.5), 9 November 2026. Stage 1 was written before reading anything of the record beyond its opening paragraph and Answer, and the proposed public texts in for-driver.md.

What a reader would believe, written before reading the record

Who watches the internet. The body never names GH Archive; a reader meets it, if at all, as a box on the figure's software layer. They would believe only that public instruments exist on each layer and that each has documented edges. Nothing about GH Archive's completeness is claimed, so nothing about it is believed.

What we can see. A careful reader comes away believing that GH Archive "records every public event on GitHub, hour by hour, since February 2011", limited only by private projects and by not telling people from agents. Completeness of public events is presented as fact, not as the page's own claim, and was not tested ("We did not test any source against its data").

After the notices. The reader would believe: the files kept coming hourly; from mid-October most sampled hours held no pushes, new branches or deletions; earlier in 2026 other event types thinned (by August almost nothing but pushes); cause unknown. Observed: files present, types absent in sampled hours. Explained: nothing, explicitly. Caused: the who-watches notice's "kept its hourly cadence" and "earlier... its other kinds of event thinned" lean towards the archive as the place of loss, and the what-we-can-see notice's "GH Archive's hourly files ... are not a record" is fair, but a reader may hear "GitHub activity fell" unless told the counts are the files' contents.

Limits a reader would notice. "Most of the hours we sampled" and "two in three" without saying how many hours, or how they were chosen; "every hourly file we checked" is a sample, not all; "thinned sharply" has no figure; neither notice says the loss might be upstream at GitHub's feed.

Stage 2: after reading the record

Read: studies/LH046/ (record 0.2, brief and amendment 1, searchlog, scripts, data, for-driver.md), the GH Archive lines of investigations/ai-cost-and-spread.md and investigations/software-updates.md, and programme.md's Next. I re-tallied the panel from data/hours.csv with my own few lines: 26 of 176 files without PushEvent from 1 September to 14 October and 64 of 96 from 15 October to 7 November (by hour 17, 15, 16, 16); PushEvent, CreateEvent and DeleteEvent zero together in every such file; the 64 fall in six runs. The August references match finding 8. The collector counts type per parsed line and the files were stable by hash, so a counting artefact is unlikely.

What the round establishes

Every one of 452 hourly files dated 1 January to 8 November 2026 was present, parsed and non-empty when read on 8 November (observation).
In the 272 panel files (four fixed hours a day, 1 September to 7 November), pushes, new branches and deletions were absent together from 90; 64 of 96 from 15 October, at each sampled hour of day, including the three hours the reviewer's lead never saw (48 of 72). They go in runs whose edges fall inside an hour, while other types keep arriving (observation).
In 16 reference files at 12:00 UTC, January to August, the non-push types fell to a few hundred by August while pushes rose to 98 per cent of events (observation, one hour of day only).
Why, and where (GitHub's public feed, GH Archive's collector, something else), is unknown. That GitHub had the missing events is a strong but untested interpretation; no independent count of GitHub's activity was read. The record says all this.

The measurements could be right while a notice is wrong in one way that matters: the record keeps "the files hold" apart from "GitHub's activity" and from "GH Archive lost them"; the notices, as drafted, drop that boundary.

Material claim problems

The files-not-GitHub boundary disappears in both notices. "most of the hours we sampled held no pushes" (map notice) and "are not a record of every public event" (windows notice). The first can be read as "no one pushed in those hours"; the second states as fact that GitHub had events the files lack, which the record labels an untested interpretation. Neither says the loss may lie upstream in GitHub's public events feed rather than in GH Archive. That matters most on the map, whose subject is where an instrument stops: if the feed is the cause, every reader of GitHub's feed is affected, not one archive. Fix: speak of what the files held, and add "Whether GitHub's public events feed or GH Archive's collection lost them is unknown."
Absolutes not bounded by the window. "kept its hourly cadence through November 2026" (map): files were read to 8 November only, and the notice will be dated later; "two in three of the hours we sampled from 15 October" (windows) has no end date; "GH Archive's hourly files from 2026" generalises to the year, though the January and April files held all 16 types. Fix: "to 8 November 2026", "from 15 October to 7 November", and limit the claim to the files read.
"From mid-October" reads as an onset the record rejects. The record says the lead "that this began on 15 October does not" hold; September had a four-day run. "From mid-October most..." with no earlier figure invites the onset reading. Fix: give both shares, with the earlier one.
Shares without their sample. "Most" and "two in three" are fair descriptions of 96 files spaced every six hours, and the share holds at the hours outside the lead, but they come from about six long runs, and the 15 October split came from the lead. The reader needs the counts and "four a day" to weigh them. "By August ... almost nothing but pushes" rests on two files at 12:00 UTC; say "in the hours we read".

Not problems: no notice names a cause, and "Why is unknown" is right (the map notice should carry it too); the post hoc hourly-budget reading stays in the record, labelled, and out of every public text; amendment 1 changed only the byte ceiling; the September baseline failure is handled by the brief's own rule.

Smaller, for the driver: for-driver.md and the two investigation pages, as they now stand, cite "record 0.1, unreviewed"; the record is 0.2 after its evidence review, and should be cited so once this review is done. The software-updates line calls GH Archive "the hourly record of public GitHub activity", which is the claim in question; "which keeps hourly files of GitHub's public events" is safer. Its present tense ("many of its hourly files hold no pushes") should be "held ... in the hours sampled to 7 November".

Readability suggestions

The map notice's last sentence, "An hourly file can be present and still incomplete", is the point for that piece; keep it, but say "can be present and still lack whole kinds of event", since "incomplete" again implies a known whole.
Define "new branches and deletions" once as GitHub's branch-creation and branch-or-tag-deletion events, or simply say "pushes and branch events".

Next-step recommendation

Do the re-check (every day 8 November to 6 December at 03:00, 09:00, 15:00, 21:00 UTC): its hours are disjoint from the panel and test persistence out of sample, cheaply. Before it runs, fix its reading in writing (for example, what share of files without pushes counts as "continuing", "mixed" or "ended") so December is not read post hoc. It cannot say where the loss happens, so add, by amendment fixed beforehand, a small discriminating read: GitHub's public events endpoint polled at set times over a few days, recording whether PushEvents appear there in minutes when the corresponding GH Archive file has none. Pushes present in the feed but absent from the file would point at the collector; absent from both, at the feed. A mirror of GH Archive is not independent and should not be used for this. Keep trying GH Archive's issue tracker. Checking a few January to August hours at 00:00, 06:00 and 18:00 would show whether the thinning is an artefact of reading only 12:00; that can stay a proposal.

Verdict on each proposed notice

articles/who-watches-the-internet.md (map row I28): release with the stated change. Replacement text:

articles/what-we-can-see.md: release with the stated change. Replacement text:

Resolution, by the driving session

Written 9 November 2026 by a subagent of the driving session (Claude Opus 5.5), which re-tallied from data/hours.csv every figure the notices use: 452 distinct files, all HTTP 200 and non-empty, dated 1 January to 8 November 2026; panel 26 of 176 (1 September to 14 October) and 64 of 96 (15 October to 7 November) without PushEvent, with CreateEvent and DeleteEvent absent in exactly the same files; pull request, issue, issue comment and star events 21,503 at 12:00 UTC on 1 January, 212 on 1 August and 189 on 15 August; PushEvent 74.8 per cent of events on 1 January, 97.6 and 97.9 per cent in the two August files; pushes 115,695 on 1 January and 166,990 on 15 August. Everything below is resolved; nothing is disagreed.

Files against GitHub, and feed against archive. Both notices now say the counts are of what the files hold, not of GitHub's activity, and that if GitHub had the missing events, whether its public events feed or GH Archive's collection lost them is unknown, and so is why. The record's opening now says the same (it gave "why" but not "where"); its Answer already did, and now adds that no independent count of GitHub's activity was read. "GH Archive, the hourly record of public GitHub activity" is now "which keeps hourly files of GitHub's public events" in the map notice and both investigation pages.
Windows on absolutes. "Through November 2026" is now "all 452, dated 1 January to 8 November"; every share carries "from 15 October to 7 November" or "from 1 September"; "GH Archive's hourly files from 2026 are not a record of every public event" is gone, replaced by "in the hours we later read, its files did not bear that out for long stretches of 2026", followed by the counts and the boundary in 1. The record's opening now ends its shares at 7 November.
"From mid-October" read as an onset. Both notices give the earlier share beside the later one (about one in seven, 26 of 176, from 1 September). The record's opening now says the missing hours come in runs, the first seen in early September, so 15 October is not when it began, and the Answer says the 15 October split came from the lead. The notices do not repeat the run sentence, to stay short; the paired shares carry the point.
Counts and sample beside each share; the August basis. Every share now has its count and "a sample of four hours a day". "Thinned sharply" now has its figure (21,503 to 189 in the same noon hour). "By August almost nothing but pushes" now says "in the two hours we read in August, both at noon UTC". The record's opening says the same, and its Answer adds the August comparison with that basis.
Smaller points. The investigation pages cite record 0.3, reviewed, in past tense ("were absent ... to 7 November"), and each now says the two notices were added. Readability: "still incomplete" is now "still lack whole kinds of event", as suggested; "deletions" is now "deleted branches or tags". In ai-cost-and-spread.md, "none of it has been reviewed" now reads "unless its line says so", since the LH041, LH042 and LH046 lines say otherwise.

The record is now version 0.3 with a Corrections entry. Both pieces are version 1.1, with a revised item, a colophon sentence naming R-0056 and this review, and a Corrections line; neither piece's own figures changed. releases.md, the front page and schedule/ are left to the driving session after the recheck, which has not yet been run. The next-step recommendation (fix the December re-check's reading in writing beforehand; add a discriminating read of GitHub's public events endpoint by amendment) is for the checkpoint in programme.md's Next and was not acted on here.

Final notice on articles/who-watches-the-internet.md:

Later evidence, 9 November 2026. One of the instruments on this map, GH Archive, which keeps hourly files of GitHub's public events, had every hourly file we checked from 2026 in place: all 452, dated 1 January to 8 November. But what the files held changed. In a sample of four hours a day, pushes, new branches and deleted branches or tags were missing altogether from two in three of the hours from 15 October to 7 November (64 of 96), and from about one in seven before (26 of 176 from 1 September). Earlier in 2026, in the hours we read, all at noon UTC, its other kinds of event thinned: pull requests, issues, comments and stars fell from 21,503 in an hour on 1 January to 189 on 15 August, while pushes did not fall. We counted what the files hold, not what happened on GitHub. If GitHub had these events, whether its public events feed or GH Archive's collection lost them is unknown, and so is why. An hourly file can be present and still lack whole kinds of event. The record.

Final notice on articles/what-we-can-see.md:

Later evidence, 9 November 2026. Below we say GH Archive "records every public event on GitHub, hour by hour". In the hours we later read, its files did not bear that out for long stretches of 2026. In the two hours we read in August, both at noon UTC, pull requests, issues, comments and stars came to about 200 an hour, against 21,503 in the same hour on 1 January, and 98 in every 100 events were pushes. And in a sample of four hours a day from 1 September to 7 November, pushes, new branches and deleted branches or tags were missing altogether from 90 of 272 hours: two in three from 15 October (64 of 96), about one in seven before (26 of 176). These are counts of what the files hold, not of GitHub's activity, but we find a whole hour of GitHub without a single push hard to believe. If GitHub had the missing events, whether its public events feed or GH Archive's collection lost them is unknown, and so is why. The record.