Release review of LH014: the piece "Further down Python's download rankings, fewer libraries' new versions take half the downloads in two days" at version 1.0, its short form, the record at version 0.1, and the consistency surfaces changed for them

Findings

This is the release gate for LH014, which drew 160 projects at random from two bands of ClickPy's August 2026 ranking (positions 51 to 500 and 501 to 5,000) and applied LH011's two-day measure and LH012's settling test to their April to August 2026 releases, reading no download after 27 September. The reviewer read AGENTS.md, the piece and the short form whole, the record at 0.1, the brief, amendment 1, sources.md and the replay; ran the replay; recounted every tested and quoted figure from the retained reads with its own code (studies/LH014/review/r0019_check.py, output r0019_check.txt), reading analyse.py only for the order of its random draws so that the intervals could be compared exactly; re-read eight project-days live from ClickPy and two upload times from PyPI (review/r0019_live.py, r0019_live.txt); rendered the figure in light and dark; checked every absolute in the titles and summaries, the labelling of what was read after the counts, and every surface the round changed; and ran the site's tests and build. Every figure in the piece, the short form, the record and the surfaces is what the retained files give, and the retained files are what ClickPy and PyPI serve today for the sample read. Nothing had to be fixed for the piece to be released; four wordings in the piece, one in the short form and one on a surface were tightened. The verdict is that the piece and its short form can be released at version 1.0 with the record at 0.1.

What was checked against what

Claim	Checked against	Result
The brief and amendment 1 govern the steps they precede	`sh studies/LH014/replay.sh`	brief.md matches snapshot_sha256.txt (31c19166...) and amendment 1 matches ccb7a346...; the frame, all eight analysis tables and uptake.svg regenerate byte for byte; 293 ClickPy queries of 800, 1,371,500,991 rows of 60 billion, largest query 101,818,368; 160 PyPI reads of 400; no query naming a day after 27 September; no failed query
The four tests: band A primary 25 of 52 (48.1, 34.6 to 61.5), band B 9 of 35 (25.7, 11.4 to 40.0); settling A 39 of 50 (78.0, 66.0 to 88.0), B 23 of 35 (65.7, 51.4 to 80.0)	r0019_check.py, from counts/, releases.csv and pypi_versions.csv, with the brief's gap rule, cut, PEP 440 at-or-newer set and percentile intervals over 10,000 project resamples from `random.Random(20260929)` in the brief's order	identical to the decimal, conventions included ("not as a rule" twice; "holds as a rule", "holds for some projects")
Descriptors the piece quotes: 21 every and 22 none in A, 7 and 24 in B; median day-1 shares 48.1 and 22.7 (top 57.6); 149 and 81 releases at half by day 30, median day 1 and day 4; 13 and 16 projects with none; pooled 114 of 295 and 38 of 196	the same recount; LH011's summary.csv and record for the top fifty's 57.6, 26 of 37, 54.1 to 83.8 and "median day was day 1"	all agree; the record's settling descriptors (290 and 195 known, 219 and 87 settled; project-median Δ 4.4 and 6.1, ratios 0.88 and 0.81, Δ from day 1 3.9 and 8.2) and its two lists of projects that did not settle also agree
Opening figures: pillow 12.3.0 at 11:53 UTC on 1 July, "twenty million" downloads the next day and 53 in every hundred; sentry-sdk 2.64.0 the day before, 31 on its first full day; 64 and 31 a month later counting later versions	PyPI JSON live (2026-07-01T11:53:27Z; 2026-06-30T08:13:45Z); the recount (20,043,783; 52.7 own and at-or-newer, pillow having no later version; 30.8; 63.6 and 31.2 on day 30, sentry-sdk's counting 2.65.0, 2.66.0 and 2.66.1); ClickPy live	all agree; the piece says the two were chosen after the counts
Mirrors: three of pypistats's four in the log; a median 148 and 62 downloads on days 0 and 1; well under one per cent; no library's result changed	the recount from mirror/ (medians of project medians 148, quartiles 88 and 270; 62, 46 and 84; shares 0.01 and 0.23 per cent; no release's and no project's outcome changes without the mirror class)	agree
Settling paragraph: "the second band did grow more after the first days"; typical four days to half against one; 16 of 35 with none	ratios 0.81 against 0.88; median days to half 4 against 1 and 1; 16 projects	agree
Post hoc reading: five of six with a release; bands would read 45 and 22 per cent	the recount without hatch, editables, identify, httpcore2, nab-index and pydantic-ai-shields: 22 of 49 (44.9) and 7 of 32 (21.9); pydantic-ai-shields the one whose releases mostly did not reach half	agree; labelled post hoc in the record (2 comparisons) and as unplanned and untested in the piece
Record figures the piece does not carry: CI 14.2 and 7.1 points below (intervals 18.3 to 12.3 and 13.3 to 0.0); 50 of 52 and 23 of 35 below zero; 26.7 against 43.6 and 40.0 against 51.3; without the overlap projects 22 of 47 (46.8, 31.9 to 61.7), 34 of 45 (75.6, 62.2 to 86.7), CI 14.3 (20.5 to 11.7)	the recount from ci/ with amendment 1's gap rule, continuing the same generator	agree; the reviewer's "flag not known" shares (2.3 and 2.8 per cent) are over the subsample's pip, uv and other downloads on both sides and differ in denominator from the record's 3.0 and 4.3, which were not recomputed on the record's definition
ClickPy's two-table checks: 8,704 project-days, 8,389 identical, at most 1.0 per cent; 1,244 release days, 1,180 identical	the recount: 8,704, 8,389, 1.01 per cent; on release days the reviewer compared 1,473 days (1,394 identical, 1.02 per cent) because it did not stop at the cut or at days with no by-version row the same way	agree where the same days are compared; the inconclusive condition is far from met either way
Retained counts against the source today	ClickPy live at 13:28 and 13:30 UTC: pillow 12.3.0 on days 1 and 30, sentry-sdk 2.64.0 on days 1 and 30, astropy 7.2.1 on day 1, djlint 1.37.0 on day 2 (by-version rows), pillow's per-day total on 2 July, and the mirror classes' cells for pillow 12.3.0 on 1 and 2 July	every row identical (a first run folded 12.2.0 differently from the study and differed only by that fold, with sums equal; the script was aligned and rerun)
The figure, rendered light and dark	Chromium headless shell, 720 by 560	both themes legible, with the three intervals, the one-half and three-quarters lines and the two marker kinds as the piece's description says; the description's ranges checked against project_outcomes.csv and LH011's: A dots 46.4 to 94.4, squares to 53.7 with three above 45 (50.7, 53.3, 53.7); B dots 33.3 to 86.1, squares all but one (51.0) below 45; top dots 50.7 to 76.7, squares 2.0 to 34.0 (the description said "below 34"; now "at 34 or below")
Absolutes in titles and summaries	both titles and summaries	no never, always, all or none; both summaries carry April to August 2026; the piece's "most of the fifty" was 26 of 50 by a denominator the study did not measure (fixed, below)
Observation, derived measurement and interpretation kept apart; what was read after the counts labelled	the piece whole	the opening pair and the tools-and-young-projects reading are labelled as chosen or looked at after the counts; "Rank is not a reason" and the closing paragraphs are signposted; one interpretation opened its section as fact (fixed, below)
Consistency	programme.md's LH014 answer and its Next entry, README's opening, investigations/software-updates.md (three passages), AGENTS.md's three lines, observations.md's paragraph and table row, the four "Later evidence" notices	all give 25 of 52 and 9 of 35 against 26 of 37, 39 of 50 and 23 of 35, the smaller first-day share with no reason measured, mirrors making almost none, and no download after 27 September; observations.md's 491 versions is 295 plus 196; one AGENTS.md clause could be read as the bands where the pace was not the rule (fixed, below)
Checks	`npm test` and `npm run build` in site/	15 tests pass; the build writes 125 files, and the only unresolved links are the two pre-existing ones in notes/R-0003.md
House rules	grep for non-ASCII in the piece, the short form and the record; the prose before the colophon; the short form's length	only the bylines' middle dot and the record's ellipsis and Δ; no document name, study number, commit or role before the colophon; the short form is 178 words after the byline and about 250 after the title

Must fix, fixed

None.

Should fix, fixed

Both summaries. "For most of the fifty most downloaded Python libraries, most new versions ... were half" counts 26 against 50 where the study's denominator is the 37 that released, which every other surface gives. Now "For 26 of the 37 most downloaded Python libraries that put out new versions between April and August 2026, most of those versions were half of the library's downloads within two days."
Piece, "A smaller share, set early". "The difference lies more in how large a share a new version took than in how fast" is the piece's reading of the descriptors, stated as fact before its own qualification; now "The difference, we think, lies more in ...".
Piece, the figure's description. "11 squares below 34": s3transfer's median is 34.0 rounded; now "at 34 or below".
Piece, colophon. The study directory "follows [the latest reading]" linked the reading of what stays behind under a name this piece now carries on the surfaces; now "the reading of what stays behind".
AGENTS.md, "Current understanding". "below the top fifty it was not the rule, in 25 of 52 libraries ... and 9 of 35" could be read as the count of libraries where it was not the rule; now "it was not the rule, holding for 25 of 52 libraries ... and 9 of 35".
Record, Method. One sentence added naming review/ as the place of the release review's recount and live re-reads, as AGENTS.md's Evidence asks; no figure, claim, table or script changed, so the version stays 0.1.

Notes, kept

The piece's opening says that further down, "For most of them, a new version did not reach half so soon": 53 of the 87 drawn libraries that released, and the body gives both bands and band A's interval crossing one half. Kept.
The record's "flag not known" shares (3.0 and 4.3 per cent) were not recomputed on the record's own denominator; the reviewer's differently defined shares are of the same order and the piece does not quote them.
The piece carries no CI finding, though the record and programme.md describe one; a piece need not carry everything, and the notices do not claim it. Kept.
Both colophons already name notes/R-0020.md and a release "after those reviews", and the record's header names no commit for this edition; these are the driving session's release steps, as in R-0017, along with the line in releases.md and the front page.

What the review could not check

That the driving session and the research agent ran on Opus 5.5 rests on their own statements.
The brief's and amendment 1's hashes were checked against snapshot_sha256.txt and the hash the dispatch gave; the issue #16 comments that carry them were not read, since the review posts nothing on GitHub and GitHub's pages were not needed.
The ranking itself was not re-read: ClickPy's August table has been read four times by earlier studies and once here, identically each time by the record's comparison, and reading it again returns download counts the review did not need.
The CI counts (pypi.pypi) were not re-read live; each such query scans one row per download and the piece quotes no CI figure.
The reader-and-inference review (notes/R-0020.md) is not this review's and was not run.
No count of any day after 27 September 2026 was read, by design.
The review's own cost: priced by the driving session from the transcript.

Nothing remains that must be fixed before release; the colophons' lines on the reader's review and the release, the line in releases.md, the front page and the commit that carries the edition are the driving session's to complete.