Published November 4, 2026 | Version v5.0.0

A template/ approach to Reproducible Generative Research: Architecture and Ergonomics from Configuration through Publication

Authors/Creators

Description

5.0.0 — Research Project Template

Released from tag v5.0.0 (annotated tag object 9829f410cf1d4edddcefb5374363f11ef366044d) at commit e932a5d2a4f5c75f354352b43e54039eac9f63cf — the merge of pull request #223, whose parents are ee8d6a825aa9047b5179939542ce5b5d49886d74 and 4f5578ce6441aaeae300f99e7a478050f383a371.

Correction to the in-tree note

docs/releases/v5.0.0.md inside this tag states that "the tagged commit is ... c5af32a18". That is wrong. c5af32a18 is the merge of #215 — a genuine intermediate release-train commit, but 44 commits before the tag. Verify for yourself:

git rev-parse v5.0.0^{commit}          # e932a5d2a4f5c75f354352b43e54039eac9f63cf
git rev-list --count c5af32a18..v5.0.0 # 44

The note was authored in commit c65ee252c, whose sentence placed the intermediate merge in apposition to "the tagged merge" — a true fact about a real commit, presented as the tag. No gate caught it: check_release_versions asserts the note exists and that its first heading matches the version, and nothing verified the prose provenance. The check is being added, with a negative control that fails against this tag before it passes against its successor.

This Release body is the correct record. The tag's own file is superseded by v5.0.1, which carries the corrected note. v5.0.0 was deliberately not re-cut a second time, so that its identity stays immutable; a permanently identifiable artifact with a known-wrong sentence is worth more than a third tag cut at the same name in one day.

What 5.0.0 changes

  • Native inventory sharding seam. Per-cell transport artifact naming for every kind, with cell=None keeping the previous names byte-identically, and a union gate that counts cells rather than shards.
  • Sidecar link-sync consolidation. Four linking modules unified.
  • Publication gate selection policy. --changed narrowing is admitted only for evidence that cannot reach another project's publication output; an observation failure raises rather than guessing a clean change set.
  • Diagnosability. A native refusal now names why it refused instead of discarding the reason.
  • Platform 5 admission. The release guards admit platform 5, and every canonical exemplar's publication binding moves to 5.0.0 with its tracked copy and manifest.
  • Native repairs, residue fixes and the gates around them. The whole-PDF control ceiling, the deadline and verdict guards, the retained preflight refusal reason, the installer console-script shim exemption, the declared-capacity reclaim, and the checks that now refuse a stale version on CITATION.cff, a release note's first heading, the release index, the root prose surfaces, an exemplar's platform binding, its lock rows and its artifact manifest.

Evidence, read at this exact commit

| Check | Result | | --- | --- | | publication-audit --all-public --strict --rendered | pass — 25 projects, 0 blocking, 0 review-required | | check_publication_outputs.py --all-public --strict --jobs 4 | pass — 25 projects, 0 blocking, 0 review-required | | check_release_versions.py --tag v5.0.0 | pass — root 5.0.0 + 25 exemplars | | Rendered-provenance strict validation | 25 of 25 PASS on the clean committed tree | | CI Gate on the merge commit | success, 86 checks, zero non-success |

What this release does not claim

  • The hosted required-acceptance lane has not passed at this commit. That lane has never succeeded in this release's history: 39 runs, 0 successes. The merge was admitted on CI Gate, the root release contract, and the rendered-provenance evidence above — not on that lane. Nothing here should be read as satisfying it.
  • native Source reference lacks strict published rendered evidence was never a content blocker. The predicate was correct; the receipts were behind it, because the stage fingerprint hashes every tracked file under infrastructure/ and scripts/, so two implementation files invalidated all 25 receipts at once. Fifteen exemplars whose publication copies moved to platform 5.0.0 were re-rendered and re-receipted; the remaining ten were rebound to unchanged output bytes under the current stage closure. All twenty-five pass the gate unchanged in force. The value of this release is that the instruments tell drift from missing content rather than reporting silence.

Known issue: the rebind treadmill

Relocks and tracked-file edits invalidate rendered receipts that no rendered byte changed. Measured four times in one evening during this release, each requiring a separate rebind pull request, each with output: 0 moved:

| change | fingerprint moved | rendered bytes | | --- | --- | --- | | dependency relock of all 25 exemplars | config on 25/25 | 0 | | one tracked rendering file | stage on 25/25 | 0 | | one exemplar's certifi bump | config on 1/25 | 0 | | four more relocks | config on 25/25 | 0 |

The gate's message — "validation report was not generated for the exact current rendered snapshot" — is false in every one of these cases. The snapshot is current; only a fingerprint input moved. The repair is in review.

🤖 Generated with Factory Droid

Notes

If you use this template in your research, please cite it as below.

Files

docxology/template-v5.0.0.zip

Files (188.6 MB)

Name Size
md5:3fa491722f420a2fcfab533f6478568e
188.6 MB Preview Download

Additional details

Related works