Releases, corrections, and readings
Dated, sourced, newest first. A release ships, a correction is disclosed, or we read something and say what we think — in the open, and it keeps happening.
-
The retired D-series export is removed, B3 names workflows that ignore a failing test, and scans that stopped now write a report; v17 is unchanged where git answers
0.6.12 removes the
@cejel/cejel/d-seriessubpath, deprecated since 0.6.2, which is a breaking change for code that imports it. Under every rubric, B3 now names a counted CI workflow in which a failing test command is ignored, without changing any score. Scans that stopped without a report, on more than 16 scan limitations or a path over 700 characters, now write one, and a GitLab metrics report shows score changes on a merge request.Under the default v17 rubric no score changes where git answers its history-depth check; where it fails, A2 and B2 now read the repository’s own shallow file. Every rubric’s behaviour fingerprint moved because its corpus changed; the corpus change alone moved them, and it changed no rubric’s behaviour. The prospective v27 rubric names more of what it cannot read; it stays opt-in and carries no precision, recall or calibration claim. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.12 · v0.6.12 release · Leaderboard
-
Cejel is a GitLab CI/CD Catalog component, and GitLab Code Quality fingerprints hold across score changes; v17 is unchanged
0.6.11 publishes Cejel in the GitLab CI/CD Catalog, after its GitLab template ran on a GitLab.com runner, and keeps a finding’s GitLab Code Quality fingerprint when only its score or count changes. Every located finding gets a new fingerprint once, so the first merge request after upgrading shows each existing finding as resolved plus new, one time. The GitHub Action no longer reads or changes the calling job’s pnpm, and
cejel issuecan issue a certificate from an earlier year.Under the default v17 rubric no score changes. The prospective v27 rubric reads more of what CI runs; it stays opt-in and carries no precision, recall or calibration claim. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.11 · v0.6.11 release · Leaderboard
-
A scan that cannot read its git history now gives no score; certificates open with a plain-language page
0.6.10 makes a shallow clone, a checkout whose history depth cannot be read, and a directory with no
.gitunrated under every rubric, the default v17 included, with the instruction to fetch the full history. A shallow clone could score higher than the full clone of the same commit. GitHub Action users get a score again by checking out withfetch-depth: 0. Under v17, a complete clone is scored exactly as before. Certificates now open with a plain-language first page: what holds, what is wrong, what is not known, and what to check next.The release also carries the changes first tagged in 0.6.9, which reached GitHub Releases only. The published leaderboard, scored from full-history clones, is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.10 · v0.6.10 release · Leaderboard
-
History reads on large repositories, and a rubric pin for the GitHub Action; no public v17 score moved
0.6.9 makes A2 read all of a large history in pages and batches, within a time budget, and say when it could not read the history in full, under every rubric. The GitHub Action gains an optional
rubric-pininput. Under the prospective v27 rubric, a shallow clone gets no score. 0.6.9 was published as a GitHub Release only: its npm publish stopped at a timing guard, and 0.6.10 carries its changes on every route.In two preregistered confirmations on the 23 public corpus repositories, no public v17 score, status, placement or fingerprint moved. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.9 · v0.6.9 release · Confirmation result (#567) · Confirmation result (#586) · Leaderboard
-
Certificates now say what they set aside or did not check; no public v17 score moved
0.6.8 fixes thirteen places where a long repository line could stall a scan, present in 0.6.7 and earlier. Under every rubric, certificates now name the credential-shaped values A2 set aside as placeholders, label each CI workflow by its own trigger, and count and label Git LFS pointer files. Ingested evidence is now bound to the repository and commit it declares, and
--strict-ingestrefuses a mismatch. A prospective v27 rubric is opt-in; the default stays v17.In the preregistered confirmation, no public v17 score, status, placement or fingerprint moved across the 23 public corpus repositories, and no v27 score was higher than its v17 score. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.8 · v0.6.8 release · Confirmation result (0.6.8) · Leaderboard
-
A security fix for where a scan writes its output; no public report moved
0.6.7 fixes one security issue: a scanned checkout could commit its output directory, or a file in it, as a symbolic link and make a scan overwrite a file outside the checkout (#508). The CLI now refuses such links before writing anything. Under the opt-in v26 rubric only, A3’s build or type-check finding fires only on the evidence that settles it (#509).
In the preregistered confirmation, all 23 public corpus reports were byte-identical under v17 and v26. No public repository carried the finding v26 item 8 narrows, so the run does not test item 8. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.7 · v0.6.7 release · Confirmation result (0.6.7) · Leaderboard
-
A security release: three fixes, one narrow scoring change, no public report moved
0.6.6 is a security release. A repository could stall a scan with a regular expression or a workspace glob (#502); an HTTP-server upload could make Git read another repository on the host (#500); and control characters from ingested findings reached the terminal (#500). The one scoring change: a repository whose only route to its coverage script is a
pnpm run /…/selector loses that coverage credit.In the preregistered confirmation, all 23 public corpus reports were byte-identical under v17 and v26. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.6 · v0.6.6 release · Confirmation result (0.6.6) · Leaderboard
-
The release reads evidence four checks missed; five public rows moved, none down
0.6.5 makes four v17 recognition fixes under the existing rubric identifiers: B3 reads declared workspace packages (#479), A1 recognises more coverage configuration (#484) and more frameworks’ assertions (#487), and A2 flags a digest compare computed with a keyed HMAC (#489). In the preregistered public before-and-after every expected value held: five public repositories moved at scoring level, all up or unchanged, none down.
v26 is added as an explicit opt-in prospective rubric; it carries no precision or recall claim, and the public default stays v17. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.5 · v0.6.5 release · Before-and-after result (0.6.5) · Leaderboard
-
The release changes three scoring checks; the public before-and-after moved no row
0.6.4 changes three v17 scoring checks under the existing rubric identifiers: two Express error-handler spellings A3 missed (#459), AVA assertions in inline callbacks for A1 (#427), and a timing side-channel A2 asserted from a bare digest name (#430). In the preregistered public before-and-after, no score or placement moved on any public row.
The certificate now says how many workflow and package files it read beside the ones it cites. The published leaderboard is not re-run for this release; it declares a scorer version pin at 0.6.3.
Sources: Changelog: v0.6.4 · v0.6.4 release · Before-and-after result (#459, #427, #430) · Leaderboard
-
The board is regenerated on full history, and tied rows now share a rank
Earlier boards scored each public row from a single-commit clone, so the history-reading parts of A2 and B2 saw one commit. The leaderboard is now regenerated on 0.6.3 from full-history clones. Rows with the same comparable score to one decimal share a rank instead of being ordered by name. The rubric is unchanged and no verdict changed. Overall moved by 0.1 on seven rows and by 0.2 on one; four rows share first place, and fastapi enters the ranking.
Cejel’s own row is now marked low confidence. Under the #433 fix in 0.6.3, B6 is not applicable to it, which leaves 2 of 6 process criteria measured. Clone depth is not the cause. The known-issue note on the board is retired.
Sources: Leaderboard · Rubric changelog, 2026-10-11 entry · Measured result (#455) · Board regeneration (#431)
-
The release stops crediting text that only describes a control; the board is not yet regenerated
0.6.3 removes credits Cejel gave to text that only describes a control: detector source, tests, review notes, changelogs and committed certificates that mention a review gate or an error handler without being one. In the preregistered before-and-after, two public rows moved and no ranked placement changed.
The published leaderboard is not regenerated by this release: board rows still show the 0.4.9-era scoring until the full-history regeneration (#431). Update, 2026-10-11: the board is now regenerated; see the note above.
Sources: Changelog: v0.6.3 · v0.6.3 release · Rubric changelog · Board regeneration (#431)
-
The release adds a Claude Code plugin and discloses shallow clones on the certificate
0.6.2 ships a Claude Code plugin: the offline
scantool, started through a pinnednpxlaunch, and a skill on reading a certificate as evidence rather than as an audit. When the scanned repository is a shallow clone, the certificate now says which criteria read truncated history; no score moves.The changelog states who each change affects. The distribution table links the verified release surfaces.
Sources: Changelog: v0.6.2 · v0.6.2 release · Verified distribution table
-
The release attests what each binary and the image bundle, and makes the certificate read correctly
0.6.1 generates each release SBOM from the build, so it lists the packages each standalone binary and the Docker image actually bundle, and attests it to the artifact it describes. A release with an empty SBOM now fails before anything is published. The certificate’s wording and layout are clearer for a first-time reader; no score, band or verdict moves.
The changelog states who each change affects. The distribution table links the verified release surfaces.
Sources: Changelog: v0.6.1 · v0.6.1 release · Verified distribution table
-
The release adds a GitLab Code Quality export and an includable GitLab CI job
0.6.0 adds
cejel export gitlab-codequality, which writes a report’s located findings as a GitLab Code Quality report, and a GitLab CI template that runs the scan, publishes that report and can fail below a minimum score. No score, band or report field changes. The template has structural tests but has not yet been run on a GitLab instance by us.The changelog states who each change affects. The distribution table links the verified release surfaces.
Sources: Changelog: v0.6.0 · v0.6.0 release · Verified distribution table
-
The release tightens the CI score gate and makes the HTTP MCP server stateless
0.5.0 makes
--min-scorefail when fewer than half of the applicable dimensions were measured, even if the score clears the minimum; scores themselves do not change. The HTTP MCP server now returns certificates and badges with the scan that produced them, and reports move to format 1.4, recording the weight share each metric contributed.The changelog states who each change affects. The distribution table links the verified release surfaces.
Sources: Changelog: v0.5.0 · v0.5.0 release · Verified distribution table
-
The release fixes missed secrets and error handlers, and makes withheld evidence explicit
0.4.11 detects populated PostgreSQL passwords in committed environment files and recognizes more registered Express error-handler idioms. Every certificate now states whether content was withheld and which signals would have read it; the machine report carries the same disclosure.
The changelog records the fixes, paired measurements, and remaining limits. The distribution table links the verified release surfaces.
Sources: Changelog: v0.4.11 · v0.4.11 release · Verified distribution table
-
Jev judge follow-up: the operator post
The operator’s LinkedIn follow-up on the Jev judge test is linked here as its own dated post. The experiment page carries the preregistrations, results, and limits of the primary test and subsequent composition and anchoring arms.
Sources: Read the LinkedIn post · Experiment and follow-up record
-
Judge follow-ups: ask about one claim, then compose the verdict in code
Three preregistered follow-ups tested composition and anchoring on the same frozen corpus and pinned judge. They support a narrower use: ask atomic questions about a claim in front of its evidence, read the label, and compose the verdict in code.
The integrated probability still did not establish a calibrated report-verification decision. The published record includes contradicted predictions, remaining date-comparison weakness, and an erratum caught by the harness before the first call.
-
A preregistered test of a commercial judge model
The published Jev judge experiment compares constructed reports against frozen evidence. Three of five preregistered predictions were contradicted; the page publishes those outcomes alongside the predictions that held, the corpus, harness, and stated limits.
Sources: Primary experiment record
-
0.4.10 repairs npm provenance prospectively
The published 0.4.10 package carries npm attestations from GitHub Actions. The release record verifies identity across the distribution routes; the immutable 0.4.9 package remains disclosed as published without npm provenance.
Sources: Changelog: v0.4.10 · v0.4.10 release
-
The v22 detection-recall measurement: the operator post
The operator’s LinkedIn post on the v22 detection-recall measurement now has its own dated entry. The methodology page provides the measurement’s scope, interval, instrument, and distinction from rubric agreement and fixture-scoped results.
Sources: Read the LinkedIn post · Measurement and scope
-
0.4.9 states the boundary of repository evidence
Certificates describe the repository tree at its pinned revision. Evidence outside that tree is neither seen nor claimed to be absent. The changelog also records paired scoring changes under the same rubric and the historical npm provenance gap.
Sources: Changelog: v0.4.9 · v0.4.9 release
-
Public language, private belief, and what closes the gap
Jacob Coxon, a pretraining researcher, resigned from Anthropic on 2026-09-09 with a public statement that both frontier labs are racing toward self-improving systems.
Our reading, as posted: the argument is about frontier pretraining, a layer we do not work at; the pattern underneath it is one we see everywhere — public language and private belief diverging, and the only thing that has ever closed that gap is evidence a third party can re-run rather than a statement taken on trust. True of a lab’s safety claims, true of a vendor’s software.
Sources: Jacob Coxon’s public statement · Later operator reflection on Coxon’s resignation (LinkedIn)
-
Public language and private belief: the operator post
The operator’s LinkedIn reflection on Jacob Coxon’s resignation is listed separately from the following day’s reading. That reading connects the gap between public claims and private belief to evidence a third party can re-run.
Sources: Read the LinkedIn post · The accompanying reading
-
cejel.dev’s release record is verified across every surface
The site’s release record was completed after the release passed its own currency verifier on all fourteen distribution surfaces: npm with provenance attestations, the GitHub release with five signed binaries, the OCI image, the MCP Registry (now serving 0.4.8), the Homebrew tap, the GitHub Action floating tag, and this site.
The public leaderboard is still scored by 0.4.5 and says so; it is re-scored when the regeneration can be verified end to end.
Sources: Distribution table on /for-engineers/ · cejel verify-release-currency run 34365914100
-
A limit we imposed on ourselves was scored as evidence lost
A design partner ran two versions of Cejel on the same commit and the scores were far apart. Two compounding defects, both ours, both present since 0.4.6: a file over our own 512 KB read limit was mapped by the shape of its filename to criteria that had readable evidence elsewhere, and the resulting abstention was recorded like a file we could not read at all, so it stayed in the composite at zero.
0.4.8 fixes both, words the certificate so the two cases read differently, and records the producing tool version in every report. Disclosed the same day, fixed the same day, credited in the changelog without a name.
Sources: Changelog: v0.4.8 · v0.4.8 release · Later operator post: v22 detection-recall measurement (LinkedIn)
-
0.4.7: findings first
The certificate now shows critical and warning findings above the prose, and every criterion shows the weight actually applied. Presentation only; the machine artifacts are unchanged.
Sources: v0.4.7 release
-
AIR raises $50M for agent add-on vetting
AIR came out of stealth on 2026-09-01 with $10M from Sequoia and $40M from Greenoaks to discover agents inside enterprises, vet the skills and add-ons they load, and enforce a maintained allow-list.
A neighbour, not a competitor: AIR gates what an agent may run at runtime; Cejel produces re-runnable evidence about a codebase at the moment someone has to accept it. Gates consume evidence; the evidence has to come from somewhere the gate does not control.
Sources: TechCrunch, 2026-09-01
-
We withdrew every public score for a day
On 2026-08-18 we found our own leaderboard’s reproducibility claim was false: the published scores had been computed by an internal engine copy that had drifted from the published
@cejel/cejelpackage, not by running the package itself as claimed. Every score, rank, and verdict was withdrawn the same day; corpus membership and methodology were not — the design was never in question, one generation run’s provenance was.Scores returned the next day, once every row was regenerated by shelling out to the published package end to end, with a reproducibility guard now checking exactly that.
Sources: /leaderboard/ withdrawal record
-
A determinism claim about our own tool was wrong
The 0.4.1 release fixed a determinism defect: the checkout path leaked into
report.json, so identical repository content scanned from different directories produced different reports. The claim of byte-identical reproducibility was withdrawn, the fix shipped, and the claim was reinstated with a test guarding it.Sources: Changelog: v0.4.1 · Later operator post: Jev test follow-up (LinkedIn)