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.

Releases

What Lighthouse has released, newest first. Each piece's colophon names its reviews and its corrections.

  • 9 November 2026. Every layer of the internet had a public watcher in November 2026, by what we found, but none saw the whole of its layer, version 1.1, and We can see pieces of the internet's AI activity, but not how they fit together, version 1.1: a "Later evidence" notice on each, with no correction: GH Archive's hourly files from 2026 were all present in the hours read, but whole kinds of event were missing from many of them, and whether GitHub's public feed or the archive's collection lost them is unknown. From studies/LH046 record 0.3, after an evidence review, a reader-and-inference review and a recheck by a fresh agent (notes/R-0056, R-0062, R-0064). Released by the scheduled run.

  • 8 November 2026. What the watch would see of a pattern that copies itself, version 1.0. An essay: an invented case of software that spreads itself, walked through the public records Lighthouse and others can read, saying what each would show, how soon and how surely, and where every one is blind; it measures nothing new, and its figure of the event archive's gaps comes from a study reviewed against its evidence the same evening. After a claims check against the sources (R-0054), a reader-and-inference review that found four readings going further than the sources (R-0055), a revision, and a recheck of what the revision wrote (R-0060), all by Claude Opus 5.5 agents. Released by the driving session of the keeper's review, at the keeper's direction of 8 November to cover the programme's unmet condition of a synthesis of what the watch would detect and miss.

  • 8 November 2026. It was only a test, version 1.1: a dated correction following the fresh Claude source check, with the uncertain motive finding, analysis-agent frequency, earlier messaging, actual reuse, model attribution and timeline label made precise. A separate dated update attributes the local chronology to the newly retained session extraction. The seven Australia rows were verified against OpenAI’s accessible account. Revision check. The interpretation and proposed experiments remain hypotheses, not new findings.

  • 8 November 2026. It was only a test, version 1.0. An essay interpreting attributed incident investigations and proposing controlled work on inherited context, with five checked quotations and two original figures. Released at the keeper’s request after a primary-source check, four fresh reader-perspective reviews and a final recheck in separate OpenAI agent contexts. Publication check and review record. No new measurement or experiment; exact model identifiers and cost are unavailable, not zero.

  • 7 November 2026. Every layer of the internet had a public watcher in November 2026, by what we found, but none saw the whole of its layer and its short form, version 1.0, the map page from studies/LH035/LH035.md version 0.3 and its figure: 48 public instruments on ten layers, read from their own pages on 2 and 5 November and checked raw on 6 November. After an evidence review (notes/R-0049-lh035-map-evidence.md, $1.53), a reader-and-inference review (notes/R-0050-lh035-map-reader.md, $0.58), which found the title stating the post hoc vantage reading as the finding, and a recheck of both resolutions by a fresh agent (notes/R-0051-lh035-map-recheck.md, $1.01), all on Opus 5.5. Released by the scheduled run 2026-11-07T1413Z. Re-review when the thin layers (names, IPv6) are widened or row I30 is read raw.

  • 3 November 2026. Most downloads of old versions come from Pythons a newer version supports, version 1.3, and its short form, version 1.2: a "Later evidence" notice with no correction, from studies/LH013/LH013.md version 0.2: over the two weeks that followed, 28 September to 11 October 2026, the older shares held within five points for 32 of 37 libraries and the four boto3 patterns noticed after the counts held under tests fixed before reading; the installer's-search reading stays a reading. Matching notices on articles/where-software-updates-go.md and its short form, the investigation page, README.md and programme.md. After an evidence review (notes/R-0046-lh013-evidence.md), a reader-and-inference review (notes/R-0047-lh013-reader.md) and a recheck by a fresh agent of what the round wrote (notes/R-0048-lh013-recheck.md), all on Opus 5.5. Re-review when the installer's version is read or a later week is.

  • 30 October 2026. A PHP emergency, from an early attack trace to published fixes, version 1.3, and its short form: a later-evidence notice of vendor-reported compromise (Palo Alto Networks, Dream Research), honeypot code execution and sensor counts, read at their sources by a scheduled run, and CISA's criteria as a web reader summarised them, after an evidence review, a reader review and a recheck on Opus 5.5 (notes/R-0043 to R-0045).

  • 30 October 2026. Seen from outside, seen from inside, version 1.0, Lighthouse's first essay, with the week-one figures of the piece released the same day, after a figure and citation check and a reader-and-inference review on Opus 5.5 in one note (notes/R-0042-essay-checks.md). Re-review when the day-30 figures revise it.

  • 30 October 2026. A week after two fixes, every fixed line and branch had passed half in the two counts we followed daily and its short form, version 1.0, from studies/LH033/LH033.md and studies/LH034/LH034.md, both version 0.1 (interim, to day 7), after an evidence review (notes/R-0039-week-one-evidence.md), a reader-and-inference review (notes/R-0040-week-one-reader.md) and a recheck of both resolutions (notes/R-0041-week-one-recheck.md), all on Opus 5.5. Re-review at the revision due after the day-30 reads of 21 and 22 November.

  • 27 October 2026. A PHP emergency, from an early attack trace to published fixes and its short form, version 1.2, a second "Later evidence" notice with no correction: CISA's catalogue of known exploited vulnerabilities listed CVE-2026-109618 and CVE-2026-109619 on 26 October (studies/LH031/LH031.md, version 0.2), stated as CISA's classification by criteria the catalogue does not state, not a count or proof that any attack succeeded, with the entries linking no exploitation report we could read; CVE-2026-115560 is not listed, which is not evidence of no exploitation. Front page, investigation page, programme and observations.md carry the same. Evidence review on Fable 5.1 (notes/R-0035-php-cisa-evidence.md, $2.56) replaced "reliable evidence", a criterion no retained source states, and corrected the directive and ordering wording; the reader-and-inference review on Opus 5.5 (notes/R-0036-php-cisa-reader.md, $0.73) found the listing read as independent confirmation of attacks and the absence of the third flaw as a negative; one recheck on Opus 5.5 (notes/R-0037-php-cisa-recheck.md, $0.73) tightened three sentences, and the driving session then narrowed "cite no source" to "link no exploitation report we could read", since the entries link CISA's own alert, which refuses our reader. The reader review and recheck ran on Opus rather than Fable at the keeper's caution about budget.

  • 23 October 2026. A PHP emergency, from an early attack trace to published fixes and its short form, version 1.1, a "Later evidence" notice with no correction to version 1.0: on 23 October PHP described a further, separately numbered flaw, CVE-2026-115560, a route to running code it calls unlikely, open by its account only where an application's configuration stores session data in php_binary or php_serialize format and its code switches to php before starting a session with upload progress on, and suggested session.upload_progress.enabled=0 until its next regular release. PHP named no version; that the 21 October releases do not close the route is labelled on every surface as an inference from PHP's advice, with Cloudflare's explicit statement attributed to Cloudflare. The piece gains a dated notice, a caption note under the boundaries figure, a section "What changed after 22 October" and a longer summary; the short form a notice; with the front page, the investigation page, programme.md, observations.md and AGENTS.md, from studies/LH029/LH029.md version 0.1, a single retained read of PHP's news page at 17:00 UTC under a protocol committed beforehand. Reviewed first by a Claude Fable 5.1 evidence review (notes/R-0032-php-later-evidence.md, $2.96), which reran the replays, re-extracted the entry with its own parser, checked every quotation against the retained bytes, read PHP's page once more live and found it unchanged, and looked at both figures; its two must-fix findings, a sentence that read PHP's 10:12 UTC stamp as a publication time and 17:00 UTC as a first read, and a colophon pointing to this log before the line existed, and its six others were resolved by the driving session. Then by the reader-and-inference review (notes/R-0033-php-later-evidence-reader.md, Fable 5.1, $2.08), whose stage 1 read the piece, short form, README.md and the figures only, and whose stage 2 found no material problem and two overstatements at the edge, a count of applications said of the world and a reworded Cloudflare statement read as a new one, both fixed, with its two readability suggestions. A fresh agent rechecked both resolutions (notes/R-0034-php-later-evidence-recheck.md, Fable 5.1, $1.90) and made six corrections in place, chiefly that the stamp's evidence was stated more broadly than the retained reads show and that the summary and the investigation page still carried our own voice and a broader condition; the driving session took five of its seven proposals, leaving the new bullet's heading and the fixed LH030 brief's wording. These reviews check the account against retained bytes; they do not verify PHP's or Cloudflare's claims about their software. Prepared for the keeper's merge rather than put on main by the session, as the continuation plan asks. Re-review when PHP publishes the CVE-2026-115560 advisory, which will replace the inference with PHP's own words in a version 1.2, or when LH030's records show the trackers' affected versions.

  • 22 October 2026. A PHP emergency, from an early attack trace to published fixes and its short form, version 1.0, on the front page. The dated account draws on LH018 to LH028, with a selected early honeypot report, a fixed institutional response panel, five derived image tracks and one content-level PHP-to-Drupal layer comparison. Evidence review (notes/R-0029-php-emergency-evidence.md) required separating CERT-EU's reported 13:30 fixes from the 14:00 embargo lift and AI-assisted discovery from human coordination. Reader reviews (notes/R-0030-php-emergency-reader.md, informed, and R-0031-php-emergency-blind-reader.md, public-first) found no remaining must-fix claim; both graphics were inspected light and dark. Amazon's article was web-readable but not retained as raw article bytes by the direct collector. Attack requests, protection claims, published artifacts and actual deployments remain different boundaries. The six-hour LH028 source watch through 24 October and separate daily panels to 28 October are scheduled; future findings require dated revisions rather than rewriting this initial observation.

  • 30 September 2026. articles/tests-read-the-lockfile.md and articles/short/tests-read-the-lockfile.md, version 1.0, from studies/LH017/LH017.md version 0.3, with "Later evidence" notices on articles/where-software-updates-go.md and its short form, the front page's current focus, the investigation page, programme.md (LH017 under Answered, the choice under Next and the checkpoint), AGENTS.md and observations.md. The study's brief was hashed on issue #19 before any repository was read. Reviewed first by a Fable 5.1 release review in two stages (notes/R-0026.md): a blind check that classed 40 sampled jobs before seeing the writer's classes and agreed on 36, within its threshold; then an evidence review that reran the replay, checked every figure on every surface and read 13 more pairs in the repositories, and found three pair classes wrong by the brief's own rules. The driving session corrected the script (amendment 3 and five corrections) and the record went to version 0.2, moving four pairs: some job installs from the pinned file at 282 of 291 decided moments, where version 0.1 had 282 of 293, and no label moved. Then by the reader-and-inference review (notes/R-0027.md, Fable 5.1), which raised three material problems, all resolved on every surface that carried them: the comparison with the container recipes put them under their own study's stricter pair rule and the workflows under this one's, so the workflows looked as if they install from the pin far more often (under one rule, 97 against 85 per cent, or 76 against 77, while the lockfile gap holds either way: 164 of 188 against 71 of 113); "projects at those moments" counted moments; and the short form stated a run. The record went to version 0.3 with two readings made after the counts, the same-rule comparison and what setup-uv's cache holds at the versions the jobs name, which showed the cache caveat overstated for uv jobs. A fresh agent rechecked both resolutions (notes/R-0028.md), made thirteen wording changes, chiefly labelling the one-rule comparison as made after the counts on every surface and saying what LH015's 45 of 90 counts, and found nothing that must be fixed; the driving session then mended one garbled clause it had passed. Version 0.2 of the record is at commit 7bd3833 and 0.1 at 3bafae4. Re-review when studies/LH017/ is corrected, or when LH013's weeks are read.

  • 29 September 2026, late evening, later. articles/two-days-for-most.md and articles/short/two-days-for-most.md, version 1.2, with studies/LH011/LH011.md version 0.4, and articles/old-versions-new-pythons.md, version 1.2, with studies/LH012/LH012.md version 0.4, corrections that change no figure: litellm's six releases that reached half did so on one day, 1 May 2026, and were at half on no later day read (issue #21); a new version's share went on rising after its first days, higher a month on in 378 of 418 and 339 of 355 releases, so the pieces no longer say it barely moves or holds for the month, and the pace chart's title says the day after a release set most of its share (issue #22); from the package index, pips older than 23.2 (July 2023), not 22.3 (October 2022), fetch in full every version they try; and the pace piece's notices give boto3's old-Python result with the reading that most of it looks like an installer's search. With the synthesis's notices and its short form's, the investigation page, programme.md (the answers, and Next with issue #19 as the next outward step and issue #23 as a backlog beside it), AGENTS.md and LH013's brief (amendment 2); and studies/LH016/LH016.md version 0.2, with about.md, AGENTS.md steps 2 and 4 and its Evidence paragraph, and notes/reader-review.md: LH016's step K is outside review relayed by the keeper, much of it, by the account of Astra, a reviewer outside the workflow, Astra's, at an unknown cost rather than a person's reading at none; its costs per change describe its rounds and do not rank the steps; the recheck is one reading, not a chain; Astra's report is kept verbatim (studies/LH016/review/relayed-review-astra-2026-09-29.md), with amendment 5 to LH016's brief. Reviewed first by a Fable 5.1 evidence review (notes/R-0024.md), which reran all three replays, reproduced every figure with its own code (studies/LH016/review/r0024_checks.py) and looked at the redrawn chart in light and dark; its two must-fix findings, a lesson that dropped the rounds the relayed readings opened and a colophon that credited one agent with finds two had made, and its six others were resolved. Then by the reader-and-inference review (notes/R-0025.md, Fable 5.1), whose stage 1 read only the pieces, by its transcript, and whose stage 2 raised two material problems, both resolved on every surface that carried them: the notice without the installer's-search reading, and the pip date, which the driving session checked against PEP 714 and pip's own source (studies/LH012/review/v04_pip_metadata_read.txt). A fresh agent rechecked both resolutions and made five small fixes, three of them to sentences written minutes before for the reviews, none changing a figure, and the driving session made the five small changes it left, listed in the note. Versions 1.1 and 0.3, and LH016's 0.1, are at commit d07fac0. Re-review when LH013 reports, when issue #17's brief tests the rise, or when the reader review of the older pieces (issue #23) reaches these.

  • 29 September 2026, late evening. about.md's "How the work is checked", which now describes the reader-and-inference review, the recheck of what a round rewrites and the keeper's reading, from studies/LH016/LH016.md version 0.1, an inward study of which steps of the workflow changed the pieces' claims, after its evidence review and a recheck of the fixes (notes/R-0023.md). No piece was released; the study's replays found two claims in released pieces to correct, issues #21 and #22.

  • 29 September 2026, evening. articles/recipes-follow-the-pin.md and articles/short/recipes-follow-the-pin.md, version 1.0, from studies/LH015/LH015.md version 0.3, with "Later evidence" notices on articles/where-software-updates-go.md and its short form, the front page's current focus, the investigation page, programme.md (LH015 under Answered, the choice and its close under Next, and the later questions), AGENTS.md and observations.md. The study's brief and frame were hashed on issue #18 before any repository was read. Reviewed first by a Fable 5.1 release review in two stages (notes/R-0021.md): a blind check that classed 60 sampled builds before seeing the writer's classes and agreed on 57, all 30 judged builds among them, after which the driving session had the three script errors corrected anyway, with amendment 2 applied as written (record version 0.2: 18 builds and no label moved); then an evidence review that reran the replay, reproduced every figure with its own code, verified the piece's opening from fresh clones, PyPI and OSV, and made two fixes, that the first public notices of pyjwt's flaws came on 28 May rather than 15 June and that a count of unread lockfiles sat badly beside the half of lockfiles read, finding nothing that must be fixed. Then by the second round of the reader-and-inference review pilot (notes/R-0022.md, Fable 5.1), which raised five material problems, all resolved on every surface that carried them: a share of projects at moments given as a share of recipes, the lockfile split missing from the short summaries, the fresh-build reading given in part, a container build read as every build in the advice for a lockfile census, and one overstated sentence; the record went to version 0.3 with the review's two readings, made after the counts, and a Next that leads with CI workflows (issue #19). A fresh agent rechecked the changes and made five wording fixes, none changing a figure. Version 0.2 of the record is at commit c488d2a and 0.1 at 2eb774f. Re-review when studies/LH015/ is corrected, or when the CI workflows of issue #19 are read.

  • 29 September 2026, afternoon. articles/rarer-further-down.md and articles/short/rarer-further-down.md, version 1.0, from studies/LH014/LH014.md version 0.2, with "Later evidence" notices on articles/where-software-updates-go.md and articles/two-days-for-most.md and their short forms, the front page's current focus, the investigation page, programme.md (LH014 under Answered, and the next step under Next), AGENTS.md and observations.md. Reviewed first by a Fable 5.1 release review (notes/R-0019.md), which reran the replay, checked the brief's and the frame's hashes against their snapshots on issue #16, recomputed the four tests and every quoted figure with its own code (studies/LH014/review/r0019_*), re-read eight project-days and two upload times live, none after 27 September 2026, and looked at the chart in light and dark; it found nothing that must be fixed and made five wording fixes. Then by the first round of the reader-and-inference review pilot (notes/R-0020.md, Fable 5.1), which found one material problem, the band samples described as drawn where 80 were drawn in each band and 52 and 35 released, fixed on every surface, and, reading after the counts, that a new version mostly inherits the share its predecessor held on the eve of its release, now a labelled post hoc reading in the record, as version 0.2, and in the piece; a fresh agent rechecked the resulting changes, made six wording fixes and found one sentence of the record to fix, which the driving session fixed. Version 0.1 of the record is at commit 30f1cd7. Re-review when studies/LH014/ is corrected, when ClickPy re-ingests a day in the window, or when the next pace reading opens.

  • 29 September 2026, later. articles/old-versions-new-pythons.md and articles/short/old-versions-new-pythons.md, version 1.1, with studies/LH012/LH012.md version 0.3, a correction that changes no figure: what other libraries' limits could account for is a model's estimate under an assumption the log cannot check, strained for botocore and s3transfer, and three of the four holds come through older versions of the libraries that set them; an old Python rules out the new version without choosing among older ones, so boto3 no longer stands for that explanation; downloads sharing reported fields are not one kind of machine; the ten-point settling test is lenient for small shares; and, read after publication and labelled post hoc, boto3's and aiobotocore's older downloads have the flat count-by-version profile of an installer's resolution walk, so most may be versions tried and discarded rather than installed. With the consistency surfaces (the front page's current focus, cut to its central finding, the investigation page, programme.md's answer and checkpoint, AGENTS.md, and the notices on the synthesis and the pace piece and their short forms), LH013's brief fixed and waiting for data (studies/LH013/brief.md, issue #15; eligible when ClickPy holds 12 October 2026), and a reader-and-inference review piloted for the next two research rounds (notes/reader-review.md). Reviewed by Fable 5.1 against the record (notes/R-0017.md), which fixed one leftover overstatement in programme.md and three sentences and found nothing else that must be fixed, and then, as the pilot's rehearsal, by a second Fable 5.1 agent that read the pieces first as a reader and then the record (notes/R-0018.md): it raised three material claim problems, all resolved by the driving session, and one recheck by the same agent found none remaining. The changes after the evidence review were checked by that recheck and by the driving session's kept code (studies/LH012/review/r0018_driver_figures.py), not by a second evidence review. Version 1.0 is at commit 89d4aa7. Re-review when LH013 reports, or when the installer version behind the flat profiles is read.

  • 29 September 2026. articles/old-versions-new-pythons.md and articles/short/old-versions-new-pythons.md, version 1.0, from studies/LH012/LH012.md version 0.2, with "Later evidence" notices on articles/where-software-updates-go.md and articles/two-days-for-most.md and their short forms, the front page's current focus, and programme.md's checkpoint of 29 September (issue #15), after a Fable 5.1 review (notes/R-0016.md). The review reran the replay, checked the brief's hash against its snapshot on issue #14, recomputed every headline figure with its own code (studies/LH012/review/) and re-read twelve ClickPy queries and eight PyPI records live, all agreeing. It corrected the wheel-exclusion test, a secondary measure that never entered the Python test (litellm's 58.1 per cent excluded becomes 0.0, its wheels' platform unknown), and four wordings in the piece. Its fixes were made by the review agent and read by the driving session, not re-reviewed. Re-review when studies/LH012/ is corrected, when ClickPy re-ingests a day of the read week, or when the study of issue #15 opens.

  • 28 September 2026. articles/two-days-for-most.md and articles/short/two-days-for-most.md, version 1.1, a revision after the keeper's reading that changes no figure of version 1.0: a new title without an unbounded "never", numbers in place of the brief's "as a rule" label, a share of downloads read as a share of fetches so that a low share is not by itself a slow spread, and the first-day observation labelled as read after the counts. With it, studies/LH011/LH011.md version 0.3 (five readings made after the counts, kept as scripts/post_hoc.py and rerun by replay.sh, a corrected Next and a Limits paragraph on fetches), the consistency surfaces, and programme.md's checkpoint of 28 September with its selection revised to what keeps older versions downloaded (issue #14). Reviewed by Fable 5.1 (notes/R-0015.md), which reran the replay, recomputed the five readings with its own code (studies/LH011/review/r0015_check.py), looked at the chart rendered in light and dark, made two figures exact (97 rather than 98 per cent; a rise of 3.4 rather than 3.5 points), replaced a sentence saying the eleven's jump "came as quickly" with the record's figures and qualified two others; it found nothing that must be fixed. Its fixes were made by the review agent and read by the driving session, not re-reviewed. Version 1.0 is at commit a36508d. Re-review when studies/LH011/ is corrected or when the study of issue #14 reports.

  • 28 September 2026. articles/two-days-for-most.md and articles/short/two-days-for-most.md, version 1.0, from studies/LH011/LH011.md version 0.2, with "Later evidence" notices on articles/where-software-updates-go.md and articles/what-a-download-shows.md and their short forms, and the front page's current focus, after a Fable 5.1 review (notes/R-0014.md) that reran the replay, recomputed every headline figure from the retained counts with its own code (studies/LH011/review/), re-read eight ClickPy queries and four upload times live and found them matching, and corrected four truncated medians, the description of the eleven slow libraries as flat, the wording of the build-flag difference and the chart's description; it found nothing that must be fixed. Its fixes were made by the review agent and read by the driving session, not re-reviewed. Re-review when studies/LH011/ is corrected, when ClickPy re-ingests a day in the window, or when the next study of the eleven slow libraries opens.

  • 27 September 2026. articles/where-software-updates-go.md and articles/short/where-software-updates-go.md, version 1.0, a synthesis of the software-update inquiry and its standalone brief, and articles/the-lead-was-a-few-libraries.md and its short form, version 2.0, an editorial rewrite of 1.1 with its meaning unchanged, with the front page that now leads with the synthesis, after a Fable 5.1 review (notes/R-0012.md) that traced every figure to the records' tables, recomputed the one figure no table states, and made six small corrections to the synthesis and two to the rewrite before release.

  • 27 September 2026. studies/LH010/LH010.md version 0.3, a reanalysis of retained evidence with an offline replay (studies/LH010/replay.sh), and articles/the-lead-was-a-few-libraries.md and its short form, version 1.1, whose title and claims now say the announcement question is not settled, with later-evidence notices on articles/fixes-that-said-so.md and articles/what-moves-a-pin.md and their short forms, after a Fable 5.1 review (notes/R-0011.md) that reran the replay, recomputed every clock, notes state and comparison from the retained tables with its own code, and found nothing that must be fixed. The earlier edition of each is at commit f5a0021.

  • 27 September 2026. articles/the-lead-was-a-few-libraries.md and articles/short/the-lead-was-a-few-libraries.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0010.md) that recomputed every rate, median, leave-one-out range and permutation from the record's data, reread nine fixes' notes at their tags, eleven events against OSV and PyPI and fifteen moves from fresh clones. It found that four events' fixes had been named by an earlier advisory the window rule does not see (nltk, tornado, pip, dulwich), that the record and piece overstated how far the pooled lead goes away, and that the brief's claim to have written amendment 1 before any dependent was read was true of the data files but not of the text; it fixed the record at version 0.2, the piece's title, opening and limits, the short form, the front page, programme.md and AGENTS.md. Its fixes were made by the review agent and read by the driving session, not re-reviewed. The driving session added an update line to the colophon of articles/fixes-that-said-so.md pointing to the new piece, not reviewed. Re-review when studies/LH010/ is corrected, when the earliest-advisory sensitivity is added to its tables, or when GitHub's release pages for the eight silent fixes without notes are read.

  • 27 September 2026. articles/fixes-that-said-so.md and articles/short/fixes-that-said-so.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0009.md) that reread seven changelog entries at their tags, fourteen upload times, eighteen OSV records and twelve moves from fresh clones, and recomputed every rate and share from the record's data. It found the record's count of announced and silent fixes reversed in its prose (the data was right) and repository-and-event pairs counted as repositories, fixed both in the record at version 0.2, and made one wording fix to the piece; its fixes were made by the review agent and read by the driving session, not re-reviewed. The driving session added an update line to the colophon of articles/pins-move-over-weeks.md pointing to the new piece, not reviewed. Re-review when studies/LH009/ is corrected, when GitHub's release pages or pull requests for these fixes are read, or when LH010 opens.

  • 27 September 2026. articles/pins-move-over-weeks.md and articles/short/pins-move-over-weeks.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0008.md) that reread the advisories and release times, applied the library rule to ClickPy's list afresh, reread 18 moves from fresh clones and recomputed every table. It found the rates for the days after a release running into urllib3's advisory, seven more advisories that evening against the three libraries, and one bot's change written before its advisory; its fixes were made by the review agent and read by the driving session, not re-reviewed. The driving session added an update line to the colophon of articles/what-moves-a-pin.md pointing to the new piece, not reviewed. Re-review when studies/LH008/ is corrected, when release notes or pull requests for these fixes are read, or when LH009 opens.

  • 26 September 2026. articles/what-moves-a-pin.md and articles/short/what-moves-a-pin.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0007.md) that reread two images from the registries to the byte, four pin commits, the advisories at OSV and the download shares at ClickPy, and found that three advisories, not one, were published that evening. The review's fixes, to how many pins moved around the advisories and how firmly and to who made one commit, were made by the review agent and read by the driving session, which also corrected how the front page and short form named the thirteen projects; neither was re-reviewed. Re-review when studies/LH007/ is corrected, when the unread Docker Hub images are read, or when LH008 opens.

  • 26 September 2026. articles/what-an-image-holds.md and articles/short/what-an-image-holds.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0006.md) that read the two agentcrew images and two of serena's again from the registries and agentcrew's build recipe from its repository. The review's fixes, to the count of images read and to how firmly two inferences are stated, were made by the review agent and read by the driving session, not re-reviewed. Re-review when studies/LH006/ is corrected, when the unread images are read, or when LH007 opens.

  • 26 September 2026. articles/what-a-download-shows.md and articles/short/what-a-download-shows.md, version 1.0, and the front page of README.md that now leads with them, after a Fable 5.1 review (notes/R-0005.md) that reproduced every count from ClickPy. The fixes made after the review were made by the driving session and not re-reviewed. Re-review when studies/LH005/ is corrected, when ClickPy re-ingests a day in the window, or when LH006 opens.

  • 26 September 2026. articles/where-a-software-update-went.md, version 1.0, and the front page of README.md that carries it, after a Fable 5.1 review (notes/R-0004.md). The fixes made after the review were made by the driving session and not re-reviewed. Re-review when studies/LH002/ is corrected, when a source it reads changes at the monthly check, or when LH005 opens.

  • 26 September 2026. The front page of README.md and the short forms articles/short/the-xz-backdoor.md and articles/short/what-we-can-see.md, version 1.0, after a Fable 5.1 review (notes/R-0003.md). The fixes made after the review were made by the driving session and not re-reviewed. Re-review when a new piece is released, a new instrument is added, or LH002 opens.

  • 26 September 2026. articles/the-xz-backdoor.md and articles/what-we-can-see.md, version 1.0, after reviews on Opus 5.5 (notes/R-0001.md) and Fable 5.1 (notes/R-0002.md). The fixes made after the second review were made by the driving session and not re-reviewed. Re-review the xz piece when studies/LH004/ is corrected or a new primary source on the case appears, and the survey piece when studies/LH003/ is corrected or a source's documentation changes at the monthly check.

From releases.md in the repository, last changed 9 November 2026.