Release review of LH011 (record version 0.1, its brief with amendment 1, its retained evidence and scripts), the piece "two days for most", its short form and figure, and the consistency surfaces changed for it

Findings

This is the release gate for LH011 and the piece it earned. The reviewer read AGENTS.md, the brief with its amendment, the record, its sources, the piece, its short form and figure script, and the surface changes in 1bc6e80; reran the replay; recomputed every headline figure from the retained frame and count files with its own code; re-read a few counts and upload times from the live sources; checked the brief's rules against the commit and read-log times; and traced every figure in the piece and short form to a table. Every figure in the record's Answer stands. What had to be fixed was wording: four truncated medians, one "stayed flat" that the day-30 medians do not bear out, a "below zero" whose summary needed naming, a missing limit on what was kept, a wrong writing window in the amendment, a figure alt text that was wrong for three libraries, and a sentence in the programme's checkpoint that described two papers as saying what their abstracts do not. The verdict is that the piece can be released.

The reviewer's calculations are kept in studies/LH011/review/: recompute.py and recompute.txt (the headline figures from data/releases.csv, data/pypi_versions.csv, data/counts/ and data/ci/, without analyse.py), piece_checks.py and piece_checks.txt (the piece's and record's smaller figures from the analysis tables), figure_checks.py and figure_checks.txt (the alt text's ranges), and requery.py and requery.txt (the live re-reads).

What was checked against what

Claim	Checked against	Result
Replay	`sh studies/LH011/replay.sh`	passed: releases.csv, project_order.csv, the ten analysis tables and uptake.svg regenerate byte for byte; 152 ClickPy queries, 5,848,644,945 rows, 54 PyPI, 4 sdists, 5 pypistats
26 of 37 projects, 70.3 per cent; 21 all, 11 none, 5 mixed	recompute.py from counts/ and the frame: at-or-newer share from PEP 440 comparison of each project's non-pre-release versions against the release, day 1 or 2 at or above 0.5, "most" as more than half	26 of 37, 0.7027; 21, 11, 5
Interval 54.1 to 83.8	10,000 resamples of the 37 projects with replacement, `random.Random(20260927)`, percentile by nearest rank	0.5405 to 0.8378. The construction is the brief's, not a different one: a fresh generator per bootstrap and alphabetical project order where analyse.py uses one generator for both bootstraps in SHA-256 order, so the draws differ and the ends agree because the statistic is discrete (20/37 and 31/37)
Median of project medians 57.6, quartiles 29.5 and 69.0	recompute.py	0.5760; 0.2951 and 0.6902 (inclusive quartiles, as analyse.py; exclusive gives 0.2942 and 0.6925)
Pooled 127 of 424, 30.0 per cent	recompute.py	127 of 424, 0.2995
CI: 104 releases, 96 below; 37 project medians below zero; median of medians -6.6, interval -11.4 to -5.0; medians of medians 32.2 against 40.7; pooled 33.9 against 40.0	recompute.py from ci/, pip and uv rows only, days 0 to 2, flag 'true' over all	104, 96; 37; -6.6049; -11.4284 to -4.9639 (same construction as above, own draws); 0.3222 against 0.4071; 0.3393 against 0.3999
139 reached half by day 30, median day 1, 23 on day 0; ten projects with none; litellm 6 of 46 at median day 25; 4.5 hours left in day 0; 324/95/5 kinds; 310 followed within seven days; three yanked; six single-release projects all reached	piece_checks.py from release_days.csv and releases.csv	agrees in every figure
Piece figure's "34 of 37 within ten points of day 1 by day 30"	recompute.py from counts/ (independently of draw_piece_figure.py)	34 of 37; the three are aiohttp (55.2 to 65.6), pyjwt (57.6 to 72.3) and litellm (9.5 to 21.0)
The opening: pandas 3.0.3 at 18:52 and requests 2.34.0 at 19:29 UTC on 11 May, 37 minutes apart; next day 35 of 64 million, 55 in a hundred; pandas 26 million, 28; a month later 61 (with two patch releases) and 32	releases.csv (upload times; requests' later releases 2.34.1 and 2.34.2), release_days.csv (day 1: 35,233,968 of 64,346,148, 0.5476; 7,134,370 of 25,742,566, 0.2771; day 30: 0.6067 with own 303,650, 0.3232); PyPI JSON re-read for both upload times; ClickPy re-read for both day-1 counts	agrees
Live counts	ClickPy `pypi_downloads_per_day_by_version` on day 1 for numpy 2.4.5, boto3 1.42.81, requests 2.34.0, pandas 3.0.3, certifi 2026.4.22; `pypi_downloads_per_day` for four project-days; `pypi.pypi` for pathspec 1.1.0 against 1.0.4 and aiobotocore 3.4.0 against 3.3.0, days 0 to 2 (eight queries, 20:28:38 to 20:28:40 UTC, texts in requery.txt)	every named version's count, every "~other" fold, every daily total and every CI cell equal to the retained file; the CI shares 0.3947 against 0.6926 (pathspec) and 0.1502 against 0.2670 (aiobotocore) as ci_releases.csv
Upload times	PyPI JSON for requests 2.34.0, pandas 3.0.3, numpy 2.4.5, boto3 1.42.81 (20:28:41 UTC)	earliest file's `upload_time_iso_8601` equal to releases.csv in all four; none yanked
The CI-flag rule	data/installer_source_excerpts.txt: pip `looks_like_ci()` is `any(name in os.environ for name in CI_ENVIRONMENT_VARIABLES)` with BUILD_BUILDID, BUILD_ID, CI, PIP_IS_CI, and `data["ci"] = True if looks_like_ci() else None` with the "so as not to give the impression" comment, in both versions; uv `find_map(	&var_name
Frame fixed before counts	f833a4b committed 19:58:37 UTC; the first daily total (G01a) read 19:59:25 and the first per-version read (V02) 19:59:33 in read_log.csv; everything before is the catalogue, the monthly list, coverage, PyPI JSON and the sdists	the rule was kept, by 48 seconds. The amendment's own "written between 19:57 and 20:10" was wrong and is corrected (below)
Reporting convention, missing-evidence rules, project as unit	brief against the record: 70.3 read as "for some projects"; no gap day, so no unknown outcome; other installers' 1.7 per cent left out as "flag not known"; no 'unknown' value; per-project figures first and pooled beside them	followed
Cross-check and the two ClickPy tables	crosscheck_pypistats.csv (15 days, 9 identical, largest 1.6 per cent, pygments on 18 to 20 August), summary.csv (4,249 project-days, 151 differing, at most 1.0 per cent)	agrees
The figure's alt text: "every line is low on the day of release"; blue lines "between about 50 and 83 per cent from the first day on"; orange "between about 2 and 37 all month"	figure_checks.py: day-0 medians of the 26 run from 7.6 to 68.2 (pathspec 68.2, typing-extensions 52.9, pygments 51.1); over days 1 to 30 the blue project medians run from 48.9 (pydantic, day 14) to 84.9 (certifi, day 11), the orange from 1.9 to 37.4; every project's day-0 median is below its day-1 median	two wordings wrong; fixed (below)
Surfaces	AGENTS.md, README.md, programme.md's LH011 answer and checkpoint, investigations/software-updates.md, observations.md, the four "Later evidence" notices	all say "for most, not as a rule; for eleven none; the pace is usual, not the only one", at the record's certainty, and link to the piece rather than restating beyond the headline; the notices on the two earlier pieces say the earlier figures stand
The checkpoint	three options compared, a selection with reasons, five searches with URLs; the URLs resolve except GitHub's two, refused (403) from this container, as GitHub has been throughout	one sentence corrected (below)
House rules	no em or en dash or other non-ASCII character but the byline's middle dot in any of the fourteen files; no study number, note, commit or role in the piece's prose before the colophon; British spellings; the title states the finding; opening with the two releases, then the reason to care, then the byline; interpretation marked ("we think", "these remain guesses"); limits in their own section; the colophon with who wrote, version and date, Methods, Sources with read times, Corrections, Full record, and the one-line description; the short form 242 words with a question, one observation, the consequence and the main limit	pass

Must fix, fixed

Truncated medians (record, Findings). aiobotocore 1.9, boto3 7.4, litellm 9.4 and s3transfer 33.9 where project_outcomes.csv holds 1.98, 7.46, 9.50 and 33.97: now 2.0, 7.5, 9.5 and 34.0. The other seven were right. The piece's "between 2 and 34" and the figure reading were already right.
"stayed flat at between 2 and 34 per cent" (record, Answer; programme.md's LH011 answer). The eleven's day-30 medians run from 2.5 to 37.0, and litellm's rises from 9.5 to 21.0; now "between 2 and 34 per cent on the day after release, moved little over the month, to between 2 and 37 per cent by day 30, litellm's rising most", with the day-30 range added to Findings.
"Every project's median difference was below zero" (record, Answer and Findings). True as the median of each project's per-release differences, which is the brief's measure; but for boto3, botocore and litellm the median new-release share is above the median replaced share (by 0.9, 1.4 and 2.3 points), so a reader comparing the two columns of ci_projects.csv would see the opposite sign. Both places now name the summary and say so. The piece's "for every one of the 37 libraries the typical difference went the same way" is the median of differences and stands.
What was kept (record, Limits). data/counts/ folds every version older than a project's first qualifying release into one "~other" row a day (confirmed against the live table: the fold equals the sum of the other 137 numpy versions and 2,005 boto3 versions on the days re-read), and data/ci/ keeps date, version, installer and flag only, no Python version. A new Limits paragraph says so, as the programme's checkpoint already relied on it.
Amendment 1's writing window (brief). "Between 19:57 and 20:10 UTC" where its commit is at 19:58:37; now 19:57 to 19:58, with a bracketed dated note saying what changed and why, since an amendment is not silently edited.
The figure's alt text (articles/two-days-for-most.md and the <desc> written by scripts/draw_piece_figure.py). "Every line is low on the day of release" is false for pathspec (68 per cent on day 0) and two more above half; "between about 50 and 83 per cent" is 49 to 85 over the month. Now "every line starts lower on the day of release than the day after" and "about 49 and 85". The script was rerun; the SVG's lines and the printed reading (34 of 37) are unchanged.
The checkpoint's two papers (programme.md). PyPitfall (2507.18075) and LibEvoBench (2606.25402) were said to "read dependency metadata and requirements files and warn that download counts mix builds, mirrors and transitive installs". Their abstracts, read now, say nothing of download counts: PyPitfall reads the declared dependencies of 378,573 PyPI packages for vulnerable versions, and LibEvoBench is a benchmark of whether code-generating models keep track of library versions. The sentence now says that, and that the search summaries had suggested more. The selection does not rest on either.

Should fix, fixed

The record goes to version 0.2 with a Corrections entry naming version 0.1's commit (e1bab8e), the header's grounded_at names it, and the piece's and short form's Methods lines cite version 0.2.

Notes, kept

The piece's title, "and of some it never was": ten of the eleven had no release reach half in the 30 days read, and litellm had six of 46 after about 25 days. The body says "for ten of those, none was even a month later" and "for litellm, six of its 46 releases eventually did". Kept: "never" reads within the month the piece is about, and the body qualifies it once, where it matters.
The Answer's "gap of 38 points, larger than any project's median here (at most 30)" sets LH005's all-installer, three-week figure beside a three-day, two-installer one and says they are not the same measure. Kept, as marked.
The record's interpretation of why the eleven plateau is labelled as general knowledge with no dependent's requirements read; the piece says "these remain guesses". Kept.
programme.md's LH011 answer is long for a summary but each figure in it is the record's. Kept.

What the review could not check

ClickPy's ingestion step from pypi_raw to pypi (null to 'false') rests on LH005's reading of ClickHouse's blog, not re-read here; the record says so.
The installers' source archives were not re-fetched; the rule was checked in the retained excerpts and the retained hashes, which the collection script says matched PyPI's.
The two GitHub issue URLs in the checkpoint (aiobotocore 840 and 829) could not be opened from this container (403), so what they say about the botocore bound is the writer's search summary. SPEC 0 and NEP 29 resolve; their content was not read.
The re-reads cover eight of 152 ClickPy queries and four of 424 upload times. ClickPy re-ingests days it finds wrong, so agreement now does not rule out a later difference; none was seen.
The figure was checked through its script and its printed reading, and the alt text against the tables; the drawing was not rendered to an image here (no SVG renderer in the container).
The reviewer's resampler follows the brief's construction (10,000 draws, seed 20260927, percentile ends), so the intervals' agreement is a check of the arithmetic, not an independent estimate; a different construction was not tried.
releases.md, status: released and the colophon's "Reviewed by" and "Released by" lines are the driver's, as the dispatch said.
The review's own cost: priced by the driver from the transcript.

Nothing remains that must be fixed before release.