Bundle log

A feed reader can follow this log: log.xml holds its newest 20 days.

2026-10-29

Changed: ktl-curator, released to the librarian by the person who wrote its served-member sentences, gives the digest commands as references/domain-bundle.md does, shasum -a 256 on macOS, and says the curator sets sensitivity before confirming. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Domain schemas, released to the librarian by the person who wrote two of its sentences, opens its paragraph on a published vocabulary with the rule that a domain schema imports the copy lokf adapt makes, as ktl-librarian’s references/domain-schema.md says, and its copy sentence reads in one voice. Edited since a person confirmed it; the label says so until the curator looks again.
Curation: human:noelmcloughlin sent 2 back and released the text they wrote in both, the ktl-curator skill and Domain schemas, with the note “rewrite as needed”, so the librarian may rewrite each from its sources.
Curation: human:noelmcloughlin rewrote sentences in 2 concepts, the ktl-curator skill and Domain schemas, confirmed neither, and cleared their “correct facts” notes on both. Both edits were to text the librarian re-derived the same day after a send-back.
Changed: ktl-sidecar splits its paragraph on the preflight’s lines in two, the warnings and the skills installed for one user, with no fact changed.
Changed: ktl-sidecar names the preflight’s new core line, which warns when .lokf/lokf.yaml is from another lokf release than uv.lock pins, as references/prerequisites.md lists it.
Changed: Domain schemas splits its paragraph on a published vocabulary in two, how it is imported and what it lacks, with no fact changed. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator compares a served member’s file with its checksum_value too, as references/domain-bundle.md says; the earlier change of the same day fixed only the release check. Sent back by a person; the note waits for the curator. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Domain schemas says LinkML keeps the last definition, so a published vocabulary’s id, type, name or Person would replace LOKF’s, and that the pinned lokf.yaml is copied again when the floor moves or the lock pins another release, as ktl-librarian’s references/domain-schema.md says. The undeclared-type projection keeps its “from lokf 0.9.0”. Sent back by a person; the note waits for the curator. Edited since a person confirmed it; the label says so until the curator looks again.
Curation: human:noelmcloughlin confirmed none and sent 2 back, the ktl-curator skill and Domain schemas, each with the note “correct facts” on the librarian’s edits of the same day. The same send-back twice in one session points at the librarian’s instructions, not only at these two concepts.
Changed: Domain schemas splits its paragraph on the recipe into three, on the schema’s place, the plugins and the pinned copy, with no fact changed. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator compares a release’s .sha256 with a member’s checksum_value, as references/domain-bundle.md says since the member checksum became LOKF core’s Checksum. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Domain schemas says the pinned lokf.yaml follows the release uv.lock pins, not the floor, as ktl-librarian’s references/domain-schema.md now says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Curating a domain bundle names a member’s checksum as LOKF core’s Checksum, sha256 alone, with a checksum_value, and counts the domain schema’s new ReleaseChecksum among its five classes, as the curator’s references/domain-bundle.md and the sidecar’s templates/domain.yaml now say.
Changed: ktl-sidecar: the release paragraph is split in two, and two long sentences are shortened, with no fact changed.
Changed: ktl-docent widens along fourteen typed relations, isReplacedBy among them, as skills/ktl-docent/SKILL.md lists since 0.59.0. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator counts reliance through fourteen relation fields, and revision is described as released in lokf 0.9.0, as trust-fields.md and review-session.md say. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the three-lines page describes revision as released in lokf 0.9.0, as docs/three-lines.md says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Trust label: revision is described as released in lokf 0.9.0, as trust-fields.md says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Skills repositories: the excerpt field is released in lokf 0.9.0, and docs/skills-repositories.md names --check-excerpts. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: What KTL took from SkillWiki: isReplacedBy and the excerpt check are released, as CHANGELOG.md says for 0.59.0 and 0.60.0. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Domain schemas: the convert and serve recipes pass the schema too, adapt is released in lokf 0.9.0, and a composite of adapt copies is released as one file. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: LOKF toolkit: 0.9.0 is the latest release and the floor, --schema reaches convert, query, serve and export, and the check-refs recipe adds --check-excerpts; the cited cli.py and no-Python schema move to the v0.9.0 commit.
Changed: ktl-librarian: the check-refs recipe adds --check-excerpts, the W012 note names the v0.9.0 commit and wheel, revision is released, and coverage counts exclusions apart. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar: the variables paragraph gives automation.md’s corrected reasons, the conventions count is eighteen, the release merges a composite schema, and the preflight’s floor and schema warnings are described.
Resolved: automation.md now gives the job-if: reason for KNOWLEDGE_LIBRARIAN_ENABLED alone, and one place for every switch for the rest.
Changed: Releasing: since 0.59.0 a schema-only change packs again, and a composite schema is released with its local imports merged in.
Changed: Knowledge outside the repository: the example manifest is .lokf/ingest.yaml.sample, as docs/ingest.md says.
Changed: ktl-librarian: all four schema recipes pass --schema, and lokf adapt is released in lokf 0.9.0, as references/domain-schema.md says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar: the preflight’s schema line follows the row 0.60.0 gives it: a recipe reads [tool.lokf] schema when the justfile’s line is empty, and the release, the website and the librarian read only that line.
Changed: ktl-website: the sync’s lokf export runs under the justfile’s domain schema, as SKILL.md says since 0.60.0.
Changed: Domain bundle: a confirmation admits a member’s explicit ids outside its namespace, as the curator’s domain-bundle.md says since 0.59.0.

2026-10-28

Changed: the threat model concept follows docs/threat-model.md as #292 and #294 left it: through a proxy the proxy resolves every name the federation build and the ingest fetch, and the ingest section gives the page’s four findings, with the address rule for a git entry, the script’s user and why a command is trusted, what the lock never records, what leaves the job, and the force-push. It drops the link-local wording and the 0.56.1 date the page does not give. Edited since a person confirmed it; the label says so until the curator looks again.
Prose: reworded 3 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: the ktl-curator skill concept says the report lists each retired concept whose successor no person has confirmed, and samples concepts checked by automation only when the curation policy sets Automation re-check:, its fourth switch, as SKILL.md and references/review-session.md say since 0.58.0. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-librarian skill says that every concept names a resource or a sources entry, which conventions rule 15 checks since 0.58.0, and that the librarian cites a source by its id in a footnote, which rule 16 checks (SKILL.md, CHANGELOG.md). It also says where a confidential ingest entry’s items are recorded, as references/ingest.md does. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-sidecar skill concept says the federation workflow’s nudge job reads KNOWLEDGE_LIBRARIAN_ENABLED in its own if:, as a person’s note says and templates/github/knowledge-federation.yaml shows. It also carries the seventeen conventions of 0.58.0 from references/gate.md, and the ingest’s git versions and time limit of 0.57.0 from references/prerequisites.md and references/automation.md.
Changed: the ktl-website skill concept says the home page offers a Schema button when the site shows the schema’s pages, as CHANGELOG.md records under 0.57.0 and references/customising.md lists among the default buttons.
Changed: Repository validation names tests/test_docent_holdout.py of 0.58.0, which checks that each concept the holdout expects is still in the bundle and that evals/ asks the same questions. It also carries what 0.57.0 added to the federation, ingest and website tests and to the website smoke test.
Changed: the security policy concept lists docent-eval.yml among this repository’s workflows, with its own Environment and a read-only token, and adds a proxy’s own egress to what the policy does not cover, as SECURITY.md does since 0.58.0. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the threat model concept follows docs/threat-model.md since 0.57.0 and 0.58.0. It carries the ingest’s limits in time and per file, its git version needs, its own home folder for git, the confidential lock and the proxy it hands git, adds the docent eval, and names the variation selector among the characters the hand-off loses. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the versioning policy dates the Conventional Commits preset to 0.57.0, since CHANGELOG.md folded the unreleased 0.56.3 entries into 0.57.0 and no v0.56.3 tag exists. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Releasing dates the Conventional Commits preset to 0.57.0, since CHANGELOG.md folded the unreleased 0.56.3 entries into 0.57.0 and no v0.56.3 tag exists.
Changed: Knowledge outside the repository follows docs/ingest.md and the manifest template since 0.57.0: a git entry fetches only changed files and needs git 2.31, or 2.37 for a host name, file names are spelled for the workflow, the lock gains skipped, a confidential entry’s records stay in the ignored cache unless lock_details is true, a run stops starting work after 25 minutes, and a proxy must refuse private addresses itself.
Changed: the knowledge sources map gains a row for the docent holdout, consciously excluded as concepts so the docent never reads a question beside its answer, and the two website rows account for the site’s build configuration, as docs/repository-layout.md and docs/threat-model.md describe them since 0.58.0. Both rows still exclude the whole of src/, as a person’s note asked. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Skills repositories says the ktl-librarian concept cites eighteen sources and that four of its confirmations were overtaken, as docs/skills-repositories.md now counts them, and cites that page with a footnote. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: The pen for readers aged 13 to 18 now retells four things glossary/the-pen.md says that it left out, as a person’s note asked: the lokf validate run, moving feedback into .lokf/questions.md, the scheduled run’s wrapper and KNOWLEDGE_PEN_DRY_RUN_ONLY, and the handoff lines.

2026-10-27

Curation: human:noelmcloughlin confirmed none and sent 1 back, and cleared 4 of their own notes the librarian had answered, on the Threat model, the Versioning policy, the Security policy and the ktl-sidecar skill. The teen retelling of the pen went back for four things the adult page says and it leaves out, and the ktl-sidecar skill’s question on automation.md’s reason now carries the person’s answer.
Changed: the knowledge sources map’s external-URLs row names the Conventional Commits preset’s parser in place of the Angular preset pages, as policies/versioning.md and playbooks/releasing.md cite it since 0.56.3. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the threat model concept counts seven workflow templates under the two template folders, as docs/threat-model.md does since fc02830. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the versioning policy says a ! after the type gives a major release, since .releaserc.json names the Conventional Commits preset and semantic-release.yml pins it (0.56.3). The preset’s parser replaces the Angular parser among its sources. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the versioning question on ! is settled, since .releaserc.json names the Conventional Commits preset from fc02830 on, the change a person’s answer asked for.
Changed: Releasing says the analyzer reads commits with the Conventional Commits preset, so a feat!: commit gives a major release, as .releaserc.json and semantic-release.yml set it up in 0.56.3. The preset’s parser replaces the Angular parser among its sources.
Changed: the federation and ingest paragraphs follow docs/threat-model.md as the proxy-connect-check fix leaves it. Every address that is not public counts as private, a proxy must refuse a private address itself, and an ingest git entry is held to its checked addresses, follows no redirect, and is refused when git’s own configuration rewrites its address. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-curator skill concept now says the report’s last line either offers Step 2 or states the one stop (never elsewhere in the report), that a harness claiming the person cannot answer is a weaker sign the skill says in that line, and that Step 2 runs one item per turn, as SKILL.md has carried since references/review-session.md 0.56.2 (CHANGELOG.md). Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the knowledge sources map’s two website rows now say the whole of the site’s src/ is consciously excluded as concepts, not only its components, styles and scripts; website/src/ and ktl-website’s template copy both hold pages/, assets/, content.config.ts, routeData.ts, socialCards.ts and trust.ts alongside the components and styles, all site code rather than knowledge. sources now names each src/ tree whole in place of its components and styles paths. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: both website rows now say the whole of src/ is consciously excluded, checked against the working tree of website/src/ and skills/ktl-website/templates/website/src/.
Changed: the threat model concept splits its sync and ingest paragraphs where the topic changes, so each stays under the length the prose rules set. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the threat model concept now carries each limit and check a person’s note named as dropped from docs/threat-model.md: the workflow write to main, the ingest’s limits, the sync’s committer check, the wrapper’s own limit, what build decides, the pen’s caps and what the page does not cover. It no longer names a commit that is not on main. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: the security policy concept follows SECURITY.md again, as a person’s note asked: five surfaces instead of four, the website’s templates and preflight among them, the sidecar’s fourteen scripts and six workflows, and the Environment, federation and sync guards. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-sidecar skill concept says the librarian workflow reads KNOWLEDGE_LIBRARIAN_ENABLED in step conditions of the refresh job, as a person’s note and .github/workflows/knowledge-librarian.yaml say, not in the job’s own condition.
Resolved: the question on the ktl-sidecar skill concept assumed the switch is read in a job’s condition; the workflow reads it only in step conditions, so the question is asked again on that footing.
Changed: the ktl-sidecar skill concept asks whether automation.md’s reason for keeping the librarian’s variables off an Environment holds when the workflow reads its switch in step conditions.
Curation: human:noelmcloughlin confirmed none and sent 3 back, after reading each concept beside its sources, and cleared all 20 of their own notes. The ktl-sidecar skill and the Threat model went back with notes on where they part from their sources, and the Security policy went back with its text released to the librarian for a refresh from SECURITY.md, its four librarian questions cleared. Two librarian questions now carry the person’s answer: the release preset becomes conventionalcommits (Versioning policy), and all of website/src/ is excluded as concepts (Knowledge sources).

2026-10-26

Prose: reworded 15 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: Context layer follows docs/architecture.md’s wording since 62e1ac5: the next person or assistant does that work again.
Changed: Curator spells judgements as docs/architecture.md does since 62e1ac5.
Changed: Desk says the pen refuses a change that would fake or remove a confirmation, as docs/architecture.md puts it since 62e1ac5.
Changed: Domain schema says the curator flags concepts that no longer fit, as docs/for-the-curious.md does since 62e1ac5, and names the linkml-sidecar skills for a schema that outgrows a few classes (skills/ktl-curator/references/domain-schemas.md, 0.55.1).
Changed: Exhibition lists the two plugins’ sites and KTL Federation’s at their w3id addresses, as the README’s Exhibitions section links them since 038c6ae.
Changed: Trust ladder attributes the quoted aim to OKF, as docs/architecture.md does since 62e1ac5, with KTL as the runtime in which those fields are the rungs.
Changed: Domain schemas carries what the two domain-schema pages say since 0.55.1: the three layouts the schema variable allows, lokf adapt for a published vocabulary that cannot be imported as it is, and the linkml-sidecar skills for a schema that outgrows a few classes (skills/ktl-librarian/references/domain-schema.md, skills/ktl-curator/references/domain-schemas.md). Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Hosts and doorways lists the package-mirror, proxy and old-python3 row of skills/ktl-sidecar/references/portability.md, which 0.55.2 extended with the lock and the proxy’s certificate. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Three lines of defence follows docs/three-lines.md since 62e1ac5: the one logged exception is for a repository that cannot sign and has nobody else to approve, no KTL check runs the SHACL shapes yet, and the auditor’s Copilot form is now a link to docs/m365.md, which joins sources. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Why the skills live in their own repository drops the sentence docs/install.md cut in 62e1ac5, since the commands say their scope themselves.
Changed: What KTL took from WikiSkill says why a run in a month’s first seven days goes ahead, and that a person merges whatever the retrieval score, as docs/for-the-curious.md puts both since 62e1ac5.
Changed: The docent in Microsoft 365 Copilot names the two routes where it said either way, as docs/m365.md does since 62e1ac5.
Changed: Knowledge outside the repository says the ingest checks the address each connection opens, against DNS rebinding, as CHANGELOG.md records under 0.56.1.
Changed: Knowledge sources says the threat model’s headings are deep-linked by ai-linkmo as well as the two plugins, since check 9 of scripts/validate-repository.sh recorded two more anchors in 3114239. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator skill says a verdict is committed on its own, signed, never amended into the change it confirms, and how to run rule 14 before the push (0.55.1, SKILL.md, references/review-session.md); it carries a member’s move to another namespace from references/domain-bundle.md, and names the linkml-sidecar skills from references/domain-schemas.md. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-prose skill follows 0.55.1: rule 10 cuts sentences too, the agent reads every file in scope whether the script cleared it or not, and the rules page names the dead sentence, either way with no alternatives, and the two cases only an author can settle (SKILL.md, references/rules.md, CHANGELOG.md).
Changed: ktl-website skill follows 0.56.0 and 0.56.1: the preflight’s mkdocs line, REDIRECTS and DOCS.exclude, Step 5b’s retirement list, a docs page’s Starlight address, the genre on cards and in search, a type worded as the bundle spells it, the lowercased host, the reload after a deploy, and the just the lock pins (SKILL.md, the four references, CHANGELOG.md).
Changed: Releasing says the plan job reads a pull request’s commits from the merge-base, since HEAD is GitHub’s trial merge, as .github/workflows/semantic-release.yml does since 373e27e.
Changed: Repository validation follows validate.yml, tests/ and the smoke test since 0.55.1: the test matrix runs the just the lock pins, pypdf is 6.21.2, actionlint names knowledge-ingest.yaml, the website smoke test covers an mkdocs site’s redirects, and the federation, ingest, sync and website tests cover DNS rebinding, a namespace move, a bracketed route and the redirects.
Changed: Website says the build runs the just the lock pins, that cards and search show a concept’s genre, and that a page kept from before a deploy reloads once, as knowledge-website.yaml and CHANGELOG.md say under 0.56.1. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-librarian skill carries the three schema layouts and lokf adapt from references/domain-schema.md (0.55.1), the curator section of the hand-off with its separate verdict commit from SKILL.md, and the just the lock pins from .github/workflows/knowledge-librarian.yaml (0.56.1). Edited since a person confirmed it; the label says so until the curator looks again.
From reader feedback: ktl-sidecar skill names the command that makes a Claude Code subscription token and the two names its credential takes, from references/automation.md, and reads the variables sentence against the workflow. It also follows 0.55.1 to 0.56.1 and the federation changes on this branch: the preflight’s index, lock and just-pin lines, the rust-just pin, the sync’s bracketed routes, the ingest’s connection check, and the build and members script’s temporary downloads, pruned-release day and namespace move (references/portability.md, references/prerequisites.md, references/automation.md, CHANGELOG.md).
From reader feedback: a reader reports that references/automation.md’s reason for Repository variables is wider than the workflow bears out, since three of the variables are read in a step of the refresh job, inside its Environment. The advice stands; the concept records both readings for the curator.
Changed: Threat model carries what docs/threat-model.md says since 0.56.1: the federation build and the ingest check the address each connection opens, against DNS rebinding, with the proxy and git exceptions; the sync’s hardening bullet reads as the page now puts it; and check 9 records the sync and website headings too. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: LOKF toolkit says the sidecar’s pyproject.toml template pins rust-just beside the toolkit since 0.56.1. PyPI still lists 0.8.0 as the latest lokf, the floor the template names.
Changed: Curating a domain bundle says a release that names another namespace moves the member’s base_iri, as skills/ktl-curator/references/domain-bundle.md says.
Prose: reworded 1 concept body in plain English. No fact, link, number or frontmatter changed.

2026-10-25

Changed: docs/three-lines-critics.md is folded into docs/three-lines.md, so the concepts that cited the critics page now cite the three lines page. explanation/three-lines-of-defence.md was edited since a person confirmed it; the label says so until the curator looks again.
Changed: docs/three-lines-critics.md is folded into docs/three-lines.md, so the concepts that cited the critics page now cite the three lines page. explanation/why-four-roles.md was edited since a person confirmed it; the label says so until the curator looks again.
Changed: docs/three-lines-critics.md is folded into docs/three-lines.md, so the concepts that cited the critics page now cite the three lines page. playbooks/knowledge-sources.md was edited since a person confirmed it; the label says so until the curator looks again.

2026-10-20

Changed: ktl-prose skill: the check reads a comment’s layout, a table of options row by row and no command, and --against compares a script as its comments (references/other-files.md, scripts/prose-check.py).
Prose: reworded 6 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: ktl-prose skill: the check reads a script’s comments and docstrings, reports a paragraph past eight sentences, lists “outright” and “at all”, and runs through uv run; the rules page gains four methods (0.55.0, CHANGELOG.md, scripts/prose-check.py, references/other-files.md).
Changed: ktl-sidecar skill: the preflight runs from the repository’s root and warns about an old python3, the scripts run with .lokf/.venv before uv run behind a package mirror, and the release attaches a schema from anywhere in the repository (0.55.0, CHANGELOG.md, references/portability.md, references/automation.md).
Changed: ktl-librarian skill: the pen validates every concept a patch would write against the schema when .lokf/.venv holds the toolkit (0.55.0, references/patch.md). Edited since a person confirmed it; the label says so until the curator looks again.
Changed: The pen: it validates every concept a patch would write against the schema when .lokf/.venv holds the toolkit (0.55.0, references/patch.md).
Changed: Releasing: the schema := line may name any file in the repository since 0.55.0, and the released schema imports LOKF core as lokf (CHANGELOG.md, knowledge-release.yaml).
Changed: ktl-website skill: the graph page’s full screen and filter controls (0.54.0, CHANGELOG.md).
Changed: Website: the graph explorer’s side panel and filter controls (0.54.0, CHANGELOG.md). Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-docent skill: sources names the skill’s own scripts/ folder, the copies of the report and feedback scripts the concept describes, which the coverage sweep found under no concept. Edited since a person confirmed it; the label says so until the curator looks again.

2026-10-19

Question: policies/security.md asks the curator to bring the person-written policy in line with SECURITY.md as #253 left it, since the pen does not rewrite a person’s text.
Changed: the threat model concept now carries what docs/threat-model.md says since #253: the sync’s --apply and held workflow changes, the federation build’s address, attestation and file-name rules, the website section, and the ingest’s joiner exception. A person asked for a fresh check against the sources, and these were the places where the concept had dropped or blurred what the page says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-sidecar playbook now says where each workflow’s secrets go since 0.53.0, how the sync’s publish job applies a release and holds its workflow changes, and the federation build’s new fetch and attestation rules, as references/automation.md says. The points of the person’s 2026-10-13 note were already carried and still match the sources.
Changed: the ktl-website playbook now says the build installs packages with npm ci --ignore-scripts and Astro’s telemetry off, that deploy also runs harden-runner, and how raw HTML and copied files are held since 0.53.0, as the skill’s deploy.md and inputs.md say.
Changed: the website playbook now gives npm ci --ignore-scripts as website/README.md does, says deploy also runs harden-runner and build runs the site’s own code with telemetry off, and notes the 0.53.0 rules for copied files and raw HTML. A person asked for a refresh against the sources. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the domain-bundle playbook now names the Domain concept’s fourth setting, private_addresses, which templates/domain.yaml added in #253.
Changed: the ingest playbook now says each credential goes in the knowledge-ingest Environment, as docs/ingest.md says since #253.
Changed: the validation playbook now lists what #253 added to tests/test_federation.py, tests/test_sync.py and tests/test_website_template.py.
Changed: the source map’s ktl-website row names scripts/raw-html.mjs, which #253 added under the template’s and the site’s scripts/, folders the map already covers. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-curator concept’s checks before a member is confirmed now carry two steps skills/ktl-curator/references/domain-bundle.md gives, the members script’s own checks and the digest check for a member its site serves. The rest of the concept, the MemberDomain sentence included, matches its sources at HEAD. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the ktl-librarian concept now says the agent’s secrets go in the knowledge-librarian Environment limited to the default branch, as scheduled-task.md and the sidecar’s automation.md say since 0b2911a. Its paragraph on sources outside the repository now carries ingest.md’s qualifiers on a gone source’s question and on --prune. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the Copilot docent concept was re-checked against its four sources and still matches them. A question records that the zip limit is settled in the text and that the toolkit’s zip input still waits for a test.
Changed: the releasing concept now says the sync’s branch carries no workflow a release wrote, and the release’s workflow changes wait for a person, as CHANGELOG.md says under 0.53.0.
Changed: the skills-repositories concept, refreshed against docs/skills-repositories.md and ktl-librarian’s references/agent-skills.md, now says that five of the eight captured docent answers are about the skills, and that a scheduled run leaves an unclear project skill out and names it in the hand-off. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the SkillWiki concept, refreshed against its sources, now says that docs/for-the-curious.md holds what KTL took from the paper since 0.53.0, and that the concept stays the paper’s record. The paper and its code at 050d540 still bear out the rest. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the lokf toolkit reference, refreshed against PyPI (0.8.0 still the latest), the pinned cli.py and its other sources, now says that the federation build registers and exports each member release with the toolkit.
Changed: the context-layer concept follows docs/federation.md since 0b2911a: the build fetches members over http or https from public addresses only, checks each attestation against the member’s own release workflow, refuses a zip with a link, a hidden file or a file a browser runs, and admits a local path only with --allow-local.
Rewrite: the knowledge bundle glossary entry, refreshed against the sidecar README, the architecture page, the Obsidian page and the librarian skill at HEAD, as a person’s note asked. It now names the files that appear in the sidecar as needed, diataxis.md as reserved, the missing-doorway cases and the Obsidian vault settings. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the domain schema entry now lists CHANGELOG.md among its sources, since only the changelog’s 0.44.0 section states the version in which domain.yaml was released with the sidecar.
Changed: the health line entry now lists CHANGELOG.md among its sources, since only the changelog’s 0.45.0 section states the version in which the line began counting reopened concepts apart from drafts.
Changed: the threat model now names ktl-website among the five skills that run the repository’s scripts and the six skills that read content they did not author, counts six pinned workflow templates, carries the KNOWLEDGE_FEDERATE_TOKEN control and the Environment-scoped secrets and federation-build measures, and adds the sync, website and ingest workflows and the mirror’s prompt-injection guard, from docs/threat-model.md at 0b2911a. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the concept now names ktl-website among the five skills, and the six skills that read content they did not author, as docs/threat-model.md does.
Resolved: the concept now says ktl-website’s build also runs the commands site.config.mjs names, or a gen-doc recipe it finds, as repository code.
Resolved: the concept now counts six pinned templates, says the release step moves the skill pin with knowledge-sync.sh and its workflow, and adds the sync, website and ingest workflows and the mirror’s input-path guard.
Resolved: the agent-step bullet now carries the KNOWLEDGE_FEDERATE_TOKEN control a domain host’s members script uses.
Curation: human:noelmcloughlin corrected 1 and sent 1 back. The Security policy’s counts now match SECURITY.md: four groups of surfaces, nine scripts and five workflows under the sidecar’s templates, six skills’ prose, and five skills that run .lokf/scripts/. Three librarian questions on it stay open. The Threat model went back with its text released to the librarian, for a refresh from docs/threat-model.md.
Rewrite: the librarian playbook was re-derived from its sources at f86f5e4. The job token now serves Copilot CLI or a domain host’s members script, as scheduled-task.md step 4 says; skills/ktl-sidecar/references/automation.md joins sources for the file-writing-tool sentence; the .agents/ ignore check (0.53.0) and the ingest’s --prune are added, an unsourced date -u claim and dated history asides are dropped, and the dashes and wrapped lines are gone. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the scheduled agent writes its patch with its own file tool and never through the shell, as skills/ktl-librarian/references/scheduled-task.md says since #251; the ingest prunes only a file an earlier run recorded, as references/automation.md says. The skill install runs on every run, as .github/workflows/knowledge-librarian.yaml says, and the thirteenth convention is named as the conventions script’s.
Rewrite: ktl-curator skill refreshed against its sources at f86f5e4, after a person’s two notes asked for it. The MemberDomain sentence now carries what skills/ktl-curator/references/domain-bundle.md says such a confirmation attests (that the domain belongs under the admission rule, and who answers for it), and the wrong version stamp on it is gone; the release of text a person wrote, the mirror, the gate check’s conditions and the domain host’s checks before confirming are carried from SKILL.md, references/review-session.md and references/domain-bundle.md. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: ktl-docent skill refreshed against its sources at f86f5e4, after a person released the text and asked for it. It now carries the Reopened label, the mirror’s header in the footer and the domain host’s workspace with MemberBundle and MemberDomain members from skills/ktl-docent/references/answering.md, and the Obsidian answer from references/obsidian.md; sources swaps README.md and CHANGELOG.md for docs/architecture.md and references/obsidian.md, and a claim about the preflight that no source makes is gone. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the docent concept’s sources now name docs/architecture.md, whose table of roles holds the Runs column, in place of README.md.
Resolved: once the text was released to the librarian, the rewrite carried the Reopened label from skills/ktl-docent/SKILL.md and the domain host’s workspace from references/answering.md.
Resolved: once the text was released to the librarian, the rewrite carried the mirror’s header in the footer and the MemberDomain members from skills/ktl-docent/references/answering.md.
Changed: Domain schemas quotes docs/architecture.md, which now holds the fixed-taxonomy sentence, in place of the README, and names the domain.yaml the sidecar ships for a bundle that federates other bundles, as both domain-schema references say. The title loses its dash. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the concept now quotes docs/architecture.md and lists it in sources in place of README.md.
Resolved: the concept now names domain.yaml and the curator’s domain-bundle.md, as skills/ktl-curator/references/domain-schemas.md and ktl-librarian’s references/domain-schema.md state.
Changed: Hosts and doorways follows skills/ktl-sidecar/references/portability.md: Codeberg is named beside GitLab, Forgejo and Gitea, the GitHub Actions workflows (ktl-website’s deploy included) are what those forges lose, and Copilot holds the docent and the auditor once it exists. It adds the many-hosts vault docs/obsidian.md describes, and the title loses its dash. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the concept now says those forges lose the GitHub Actions workflows, ktl-website’s deploy included, as the portability matrix states.
Resolved: the concept no longer counts three workflows and names Codeberg beside GitLab, Forgejo and Gitea, as skills/ktl-sidecar/references/portability.md and CHANGELOG.md (0.48.0) do.
Rewrite: Skills repositories follows docs/skills-repositories.md as it stands: six skills, the ktl-librarian concept’s eleven sources and three overtaken confirmations, how often anthropics/skills changes, and what the docent answers from the bundle. The conservator’s description the page no longer gives is gone. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the concept now says six skills, as docs/skills-repositories.md does.
Changed: What KTL took from SkillWiki carries the paper’s own qualifiers (agent execution experience among its inputs, autonomous governance by default with user decisions first), rule 14 as CHANGELOG.md states it, and the person’s new note that reopens a declined proposal, from docs/for-the-curious.md. That page now holds what KTL took from the paper and joins sources. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: The registrar page was refreshed against docs/architecture.md, which now holds the fifth-role section and becomes the resource. KTL Registrar does the registrar’s work and KTL Curator assists the curator, as that page and docs/obsidian.md say, and the plugin links use their w3id addresses. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: The rewrite names docs/architecture.md as the resource and gives each plugin the role that page and docs/obsidian.md give it, which settles the question.
Rewrite: The knowledge-bundle term was refreshed against the sidecar’s README template, docs/architecture.md, docs/obsidian.md and the librarian skill. It now names the reserved index.md header keys as Golden Rule 2 lists them, gives the arrangement that puts exhibits in a vault’s index as docs/obsidian.md describes it, and adds the template README to sources. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: The website playbook was refreshed against website/README.md and knowledge-website.yaml. It now gives the weekly run and the full trigger paths, the Atom feed, the Markdown copies, the release links, the schema and data pages, the w3id address from CHANGELOG.md 0.42.0, and every file website/README.md names under What to edit. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: The rewrite gives the weekly run and the full trigger paths as knowledge-website.yaml lists them.
Resolved: The rewrite says when a promoted release’s changelog page appears, as website/README.md does.
Resolved: The rewrite’s What to edit section names every file website/README.md names, and node scripts/config.mjs.
Resolved: The rewrite gives the w3id address that CHANGELOG.md 0.42.0 records.
Resolved: The rewrite names the Atom feed, the Markdown copies, the release links and the two components that website/README.md names.
Changed: The toolkit reference now gives every lokf convert and lokf tables format that the pinned cli.py accepts, where validate finds its schema, and that --check-refs leaves targets outside the bundle’s namespace alone. PyPI still lists 0.8.0 as the latest release, with [build] and [tables] extras.
Changed: A Disagreement line names the version the answer gave or says it gave both, as skills/ktl-sidecar/templates/m365/ktl-docent-m365.md says. The 50 MB Agent Builder limit is already in the text, as skills/ktl-sidecar/references/m365.md gives it.
Question: docs/threat-model.md now names how a domain host’s members script holds KNOWLEDGE_FEDERATE_TOKEN, which the threat model concept omits.
Curation: human:noelmcloughlin confirmed 3 and sent 14 back. Trust label, Three lines of defence and Why four skill roles are confirmed, each due for review on 2027-10-19, and their notes are cleared. Knowledge bundle, What KTL took from SkillWiki, When the built-in vocabulary stops fitting, Hosts and doorways, Knowledge Trust Ladder in a skills repository, Why a registrar role, the ktl-docent skill and Build, preview and edit the website went back with their text released to the librarian. The ktl-librarian, ktl-curator and ktl-sidecar skills and the LOKF toolkit went back to the librarian. The Security policy and the Threat model went back for a fresh check against their sources, with their text kept. The same send-back came up thirteen times: refresh the concept against its sources, after the heavy development of 0.50.1 to 0.53.0.
Changed: Knowledge sources names the Archify companions in .assets/companions/ and scripts/companion-diagrams.mjs, which draws them, in the .assets row, as docs/repository-layout.md lists them since 2026-10-19. The papers’ row no longer cites docs/wikiskill.md and docs/skillwiki.md, which 0.53.0 removed, and the docs/for-the-curious.md row says that page now holds what KTL took from both papers. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Desk now says, as docs/architecture.md does, that the pen is the only way the librarian writes the bundle and refuses a change that would forge or remove a confirmation.
Changed: The pen now says that in the agent’s step of a scheduled run it only checks the patch, and a run without --dry-run exits 2 and writes nothing, as skills/ktl-librarian/references/patch.md says since 0.53.0.
Changed: Curator now says that a send-back can release text the person wrote to the librarian, through generated.by: process:ktl-curator, as skills/ktl-curator/SKILL.md says.
Changed: Trust ladder now says each label is a signal for the reader to weigh, never a score, as docs/architecture.md says.
Changed: Work list now says, as the new section of docs/for-the-curious.md on how readers and verdicts steer the next run does, that a run starts from the list and the list is ordered by history, never by a clock.
Changed: What KTL took from WikiSkill now says a declined change makes work again when a person leaves a new note, as skills/ktl-sidecar/references/automation.md says. Its sources now name docs/for-the-curious.md, which holds what KTL took from the paper since docs/wikiskill.md was removed in 0.53.0.
Changed: Three lines of defence now carries what docs/three-lines.md says and the concept had left out: the librarian’s process: re-checks and open questions, the curator’s five verdicts, what revision holds for a file and a URL, the rule behind stale_after, and the SPARQL projection. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Knowledge outside the repository says the script refuses a confidential entry when git does not ignore .lokf/.cache/mirror/ or tracks a file there, and what a confidential copy costs, as docs/ingest.md and CHANGELOG.md (0.53.0) say. Its security line follows the page’s shorter wording.
Changed: ktl-librarian skill now says the pen only checks the patch in the agent’s step of a scheduled run, under KNOWLEDGE_PEN_DRY_RUN_ONLY (0.53.0, references/patch.md and CHANGELOG.md). Its paragraph on sources outside the repository now follows references/ingest.md, and it says the source map holds no run account, as SKILL.md does. The prompt-flag, file-writing-tool and pull-request-setting sentences already match knowledge-librarian.sh and scheduled-task.md steps 2 and 5. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator skill now covers the release of text a person wrote on a send-back (0.51.0, references/review-session.md), the Reopened count in the health line, and the domain.yaml schema the sidecar ships (references/domain-schemas.md). That finishes the changes to skills/ktl-curator/ since 0c3ea0a. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar skill now says the wrapper sets KNOWLEDGE_PEN_DRY_RUN_ONLY for the agent (0.53.0), and its domain-host part follows references/automation.md: members a site serves, the sensitivity gate, the gate on links to no concept of a member, what --verify-only checks, and diff. The points of the 2026-10-13 note were already carried.
Changed: ktl-website skill now names the graph page’s workspace view for a domain host and the WORKSPACE option, added in 0.53.0 (CHANGELOG.md, references/customising.md).
Changed: sources now names the file the concept already describes, so knowledge-report.sh coverage counts it as looked at.
Changed: Repository validation: validate.yml now runs the tests as a two-part matrix (the federation tests, and the rest) with pytest -n auto, and tests/requirements.txt pins pytest-xdist==3.8.0. tests/test_federation.py now checks the in-process lokf run that 0.53.0 brought in, and makes the sidecar’s environment once under a lock.
Changed: the auditor’s list now says apart whether a confirmation is tied to the named person and whether a person’s record has been removed, as docs/three-lines.md’s table does. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-curator skill now cites docs/architecture.md for the Runs column, names the Reopened label and how a person confirms a domain bundle’s member, and the mirror sourcing the curator quotes since 0.49.0, following the human release of 2026-10-18 that allows this concept to be rewritten from its sources. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: sources now names docs/architecture.md instead of README.md.
Resolved: the concept now names the Reopened label and how a person confirms a domain bundle’s member.
Resolved: the concept now covers how the curator quotes a .lokf/mirror/ source, now that the human release of 2026-10-18 lets this concept be rewritten.
Changed: the job-token sentence now says the job’s own token is billed to the repository owner’s seat, as scheduled-task.md step 4 does. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: sources now names docs/context-layer.md, where the context layer the toolkit serves is described, beside the technology page already cited.
Changed: the three lines figure is named with its dark counterpart once; .assets/ holds ktl-three-lines.svg and ktl-three-lines-dark.svg. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the Reopened label and the Still-a-draft rule now name themselves, from skills/ktl-curator/references/trust-fields.md, answering the open question a person left on 2026-10-18. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the Reopened label is now named, from skills/ktl-curator/references/trust-fields.md.
Changed: the fifth-role sentence now names docs/architecture.md, the diagram note says dark rather than dimmed, and what remains to do now names fetching a federation member from a private repository, matching docs/three-lines.md since its latest commit. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: the sentence and the sources now name docs/architecture.md.
Resolved: the concept now says dark, matching docs/three-lines.md’s -dark source; CHANGELOG.md confirms the -dimmed set stays unreferenced.
Resolved: docs/three-lines.md no longer lists the MCP server there; it now names fetching a federation member from a private repository, which the concept follows.
Changed: What KTL took from WikiSkill now cites the paper, CHANGELOG.md and the skill files where each change lives, instead of docs/wikiskill.md, so the concept no longer rests on that page.
Rewrite: The curated context layer now summarises the three pages instead of mirroring each section, since 0.53.0 moved the whole federation design to docs/federation.md and kept the technology survey to what KTL needs.
Changed: Context layer says a domain’s owner confirms each bundle the domain admits, as docs/context-layer.md puts it since 0.53.0.
Changed: LOKF toolkit says what convert, export, serve, mcp, tables and registry write or serve, as docs/context-layer-technology.md and docs/federation.md list them, and cites the federation page, where the registry commands moved in 0.53.0.
Changed: Curating a domain bundle follows the curator’s page after 0.53.0: the sensitivity settings, a member a site serves by its concepts.jsonld, and a gate that also refuses a link to no concept of a member. It now says when the build holds a member at its last good state, and how a reader loads the workspace lazily or for some members alone. A member that waits in a member domain joins at the release that domain’s own build took.
Changed: Knowledge sources keeps its rows and standing notes, and drops the account of each past run, which git keeps. The rows follow 0.53.0’s federation page, the two paper concepts and .lokf/.gitignore, the sixteenth template this repository keeps a copy of, and each exclusion now gives its reason in its own row. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Repository validation names the five repositories whose links check 9 records, and says check 9a skips a domain host’s ignored snapshot/ since 0.53.0.
Changed: Contributing no longer names only the two Obsidian plugins among the repositories that link into this one, since CONTRIBUTING.md names none and check 9 now records five.
Curation: human:noelmcloughlin corrected 1: What KTL took from SkillWiki now cites the paper, arXiv:2606.16523, its code at 050d540 and CHANGELOG.md instead of docs/skillwiki.md, and is due for review on 2027-04-19.

2026-10-18

Changed: Curating a domain bundle says the diamond rule’s stop lasts until both owners pin one release, as the curator’s domain-bundle.md now does.
Changed: the source map names the site’s src/components and src/styles folders instead of all of src/, in website/ and in ktl-website’s template, since its rows exclude only the components, styles and scripts. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: Knowledge sources: The website rows exclude the site’s components, styles and scripts. Is the rest of src/, in website/ and in ktl-website’s template, excluded too: its pages, assets/, content.config.ts, routeData.ts, socialCards.ts and trust.ts? If not, playbooks/website.md and playbooks/ktl-website-skill.md should cite them.
Changed: the source map now lists in sources every path its rows consciously exclude, so knowledge-report.sh coverage counts those files as looked at. Each row still gives the reason, and the site’s components, styles and scripts are taken as its src/ and scripts/ folders. Edited since a person confirmed it; the label says so until the curator looks again.
Curation: human:noelmcloughlin released 3 concepts they wrote, so the librarian may rewrite them from their sources: Trust label, ktl-curator skill and Three lines of defence, each sent back earlier today. Each keeps its confirmation and note, and stays a draft until a person confirms the new text.
Curation: human:noelmcloughlin corrected 1 now and sent 4 back. Knowledge bundle now lists docs/architecture.md as a source and says the bundle opens as a vault of its own, beside a person’s vault and never inside it. The ktl-librarian skill, Trust label, ktl-curator skill and Three lines of defence went back for not following their sources’ recent changes. The same send-back came up three times: Trust label, ktl-curator skill and Three lines of defence carry text a person wrote, which the pen does not let the librarian update when its sources change.
Changed: The curated context layer: the workspace build takes each member and the domain’s settings as a person last confirmed them, as docs/federation.md says since 0.50.1.
Changed: ktl-sidecar skill: the federation build takes each member as a person last confirmed it and leaves retired concepts out, as references/automation.md says since 0.50.1.
Changed: Repository validation: the federation and sync tests cover what 0.50.1 fixed.
Changed: Knowledge sources: the thirty-fourth pass, after 0.50.1. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the domain bundle playbook says which release a member joins at while a move waits, what a send-back and a removed confirmation do, and that the Domain concept’s settings count as a person confirmed them, as domain-bundle.md now says.
Changed: Curating a domain bundle: the diamond rule gives its reason apart from what the two owners then do, as docs/federation.md puts it, and a member leaves when the curator retires its concept, as the curator’s domain-bundle.md says.
Prose: reworded 4 concept bodies in plain English. No fact, link, number or frontmatter changed.
Added: Knowledge outside the repository, from docs/ingest.md and the librarian’s references/ingest.md, which 0.49.0 and 0.50.0 brought; the coverage sweep found the guide under no concept.
Rewrite: The curated context layer: since 0.48.0 docs/federation.md holds the tiers of domains, the workspace, what exists and the decisions, which docs/context-layer.md no longer carries. The concept now follows both pages, and lists one named graph per bundle, the lokf-registry MCP server and the SPARQL endpoint as existing, with lazy loading still a design.
Resolved: The curated context layer: the table in docs/federation.md lists one named graph per bundle under knowledge-federate.py build as existing, and lazy loading as a design; the MCP server now exists as knowledge-federate.py mcp.
Open question: Hosts and doorways: skills/ktl-sidecar/SKILL.md now counts six workflow files and CHANGELOG.md adds Codeberg to the forges, while the concept, which a person wrote, still says three workflows and three forges.
Open question: Three lines of defence: docs/three-lines.md now lists lazy loading, a lokf-registry MCP server and remote harvesting as what remains of the context layer, where the concept says the domain tiers and the workspace; docs/federation.md lists the MCP server as built.
Changed: Context layer: docs/context-layer.md now says why KTL federates bundles through domains, and docs/federation.md holds how a federation runs and what remains a design. The concept no longer cites a release date or a table that the context-layer page has dropped.
Changed: Health line: the definition said seven trust labels, while knowledge-report.sh health prints eight, as the description and body already say.
Changed: LOKF: skills/ktl-librarian/SKILL.md no longer contrasts standard tooling with scripts that only work in the repository that grew them, and no other source says so, so the concept drops that contrast.
Changed: Registrar: docs/for-the-curious.md now says that on another forge, or with none, knowledge-provenance.sh checks each confirmation’s signature against the keys under .lokf/curators/, and the concept says so too.
Changed: Work list: skills/ktl-librarian/SKILL.md now says only that history orders a source’s last commit and the stamp, and no longer adds that no clock does.
Open question: ktl-curator skill: since 0.49.0 references/review-session.md says how the curator quotes a source mirrored under .lokf/mirror/, and the concept, whose text a person wrote, does not say it.
Open question: ktl-docent skill: since 0.49.0 references/answering.md says how the docent cites a source mirrored under .lokf/mirror/, and the concept, whose text a person wrote, does not say it.
Changed: ktl-librarian skill: the concept now carries the mirror of sources outside the repository (references/ingest.md, new in 0.49.0), the orphan sweep with knowledge-report.sh coverage (SKILL.md), one log line for a whole patch (references/patch.md), and the members script’s use of the job token (references/scheduled-task.md). Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar skill: skills/ktl-sidecar/SKILL.md and references/automation.md now install the sync templates (0.47.0) and the ingest templates (0.49.0), the report script counts tracked files no concept names (0.48.0), and a domain follows member domains tier by tier (0.48.0). The concept now says so, and the points the note of 2026-10-13 names are carried as SKILL.md states them.
Changed: ktl-website skill: CHANGELOG.md says since 0.48.0 the site’s figures stay in the text column and open at full size on a click, which a host gets with its next website repair.
Changed: Contributing: CONTRIBUTING.md now says each sibling’s knowledge-sync.yaml opens the sync pull request within a week, and scripts/sync-sidecar.sh --pr and --finish open and finish it by hand.
Changed: Releasing: knowledge-release.yaml attaches the domain schema the justfile names since 0.46.0, publish.yml asks the hosts in KNOWLEDGE_SYNC_HOSTS to sync, and docs/releasing.md and scripts/sync-sidecar.sh say each sibling now syncs itself through knowledge-sync.yaml, with --pr and --finish for the day of a release.
Changed: Repository validation: scripts/validate-repository.sh check 9 now reads the site’s w3id addresses, check 11 reads knowledge-sync.manifest first, and check 15 names the sidecar’s knowledge-sync.sh; validate.yml lints knowledge-sync.yaml, and tests/test_ingest.py, tests/test_sync.py and the pypdf pin are new.
Changed: Knowledge sources: a row each for the sync (0.47.0) and the ingest (0.49.0, 0.50.0), the templates row names the federation, sync and ingest templates, the README row KTL Federation’s site, and the context-layer row docs/federation.md. The map now lists under sources the files its rows account for, so the coverage sweep counts them. Edited since a person confirmed it; the label says so until the curator looks again.
Resolved: Knowledge sources: the templates row now names domain.yaml, knowledge-federate.py and knowledge-federation.yaml from skills/ktl-sidecar/templates/, the README row names KTL Federation’s site, and the context-layer row names docs/federation.md.
Open question: Security policy: SECURITY.md now counts nine scripts and five workflows in the sidecar’s templates and covers the federation build and the sync, which the concept, written by a person, does not carry.
Open question: Threat model: docs/threat-model.md now covers the sync workflow, the federation build, the ingest workflow and the mirror, which the concept, written by a person, does not carry.
Added: Curating a domain bundle, from skills/ktl-curator/references/domain-bundle.md as docs/domain-name leaves it, so a domain host can link the curator’s page by its w3id IRI.
Changed: the context layer explanation and the exhibition term no longer name a domain by its position. As docs/context-layer.md and README.md now say, the domain a build starts from bounds the workspace, and KTL Federation is the domain that federates member bundles.

2026-10-16

Prose: reworded 8 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: The curated context layer follows docs/context-layer.md, which says the domain tiers and the workspace were released on 2026-10-16 as the sidecar’s federation templates, with ktl-federation as the first host. The concept no longer says the diagrams draw them dashed, and now carries the page’s table of what exists and what remains a design.
Open question: The curated context layer: one bundle, and a federation of bundles: docs/context-layer.md’s table under What exists today lists one named graph per bundle as design, beside lazy loading and a lokf-registry MCP server. Since 0.44.0 knowledge-federate.py build loads each member’s graph.nt into a named graph named by its base_iri and writes them all to graph.nq, as CHANGELOG.md and the description in .assets/ktl-context-workspace.svg say. Is one named graph per bundle now built, with lazy loading and the MCP server still a design?
Changed: Context layer: docs/context-layer.md says the federation’s domain tiers and workspace were released on 2026-10-16 as the sidecar’s templates, so the term no longer calls them a design drawn dashed.
Changed: Health line: since 0.45.0 the line counts eight labels, with Reopened for a confirmed concept a question put back to draft, as knowledge-report.sh and the curator’s trust-fields.md now say.
Changed: Trust ladder: docs/architecture.md and its ladder figure now say that a question after a confirmation reopens a concept without dropping its rung, and only an edit drops it.
Changed: Domain schema: the curator’s domain-schemas.md says since 0.44.0 that the sidecar ships one domain schema, domain.yaml, for a bundle that federates other bundles.
Changed: Exhibition: the README’s Exhibitions section lists KTL Federation since 0.44.0, so the term names five.
Changed: Why the skills live in their own repository: docs/install.md now installs all six with one gh skill install --all --agent line, says npx skills can fail behind a proxy, and no longer records the old repository name, so the concept cites CHANGELOG.md for the rename.
Changed: gh skill: the manual’s --agent, --scope and --dir options and their defaults, which docs/install.md now relies on.
Changed: ktl-sidecar skill: the three files a domain host gets since 0.44.0, domain.yaml, knowledge-federate.py and knowledge-federation.yaml, and what each does, as SKILL.md and references/automation.md now say.
Changed: ktl-website skill: since 0.45.0 a concept’s card shows its open questions beside the label, which reads reopened for a confirmed concept a question put back to draft.
Changed: Releasing: since 0.45.0 sync-sidecar.sh keeps a sibling’s own schema := line, reports a failure of any check it runs, and writes the sibling’s changelog entry in full.
Changed: Repository validation: check 9a skips a sibling’s website build output, check 10a counts no table pipes, check 11 covers a domain host’s three files, check 17a keeps every tracked SKILL.md under skills/ktl-*/, actionlint reads five templates, and test_federation.py joins the tests, as 0.44.0 and 0.45.0 brought them.
Open question: When the built-in vocabulary stops fitting - domain schemas: Since 0.44.0 skills/ktl-curator/references/domain-schemas.md says one domain schema ships with the sidecar, domain.yaml, for a bundle that federates other bundles, and ktl-librarian’s references/domain-schema.md points a domain bundle at that template and at the curator’s domain-bundle.md. This concept does not say so.
Open question: Trust label: Since 0.45.0 skills/ktl-curator/references/trust-fields.md has a label Reopened: status: draft on a concept with a human: verified event, naming the earliest question asked on or after the confirmation’s day, its asker and how many wait. A reopened concept still counts as confirmed, and Still a draft now needs no human: event. This concept does not name the label.
Open question: ktl-curator skill: Since 0.45.0 the curator’s label table reads Still a draft / Reopened, and the health line counts Reopened, a confirmed concept a question put back to draft, apart from Drafts (SKILL.md, references/trust-fields.md). Since 0.44.0 references/domain-bundle.md says how a person confirms a member of a domain bundle. This concept carries neither.
Open question: ktl-docent skill: Since 0.45.0 the docent’s label table reads Still a draft / Reopened: for a confirmed concept a question put back to draft, the docent says the confirmation first, then when it was reopened, by whom and how many questions wait (SKILL.md, references/answering.md). Since 0.44.0 references/answering.md also says how to answer on a domain host from the workspace under snapshot/, labelling each concept with its bundle. This concept carries neither.
Open question: Knowledge sources: Since 0.44.0 ktl-sidecar’s templates also hold a domain host’s domain.yaml, scripts/knowledge-federate.py and github/knowledge-federation.yaml, which sync-sidecar.sh carries and check 11 holds to their templates where a copy exists. The templates row counts nine scripts and three workflows and names none of them. The README’s Exhibitions list KTL Federation since 2026-10-16, and docs/context-layer.md says the federation design was released that day. The README row and the context-layer row say neither.
Prose: reworded 4 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: Exhibition names the four sites the README’s Exhibitions section lists since 2026-10-16, the two Obsidian plugins’ sites among them.
Changed: Knowledge sources says the README links the site’s pages since 0.43.1 and lists the two Obsidian plugins’ exhibitions since 2026-10-16, and records the thirty-second pass. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Releasing says sync-sidecar.sh carries .lokf/justfile since 0.43.1 and keeps the sibling’s first line, which names the host, as docs/releasing.md now says.
Changed: Repository validation says check 11 holds .lokf/justfile to its template below the first line since 0.43.1.
Open question: Three lines of defence - where each role sits, and what an auditor can check: Since 0.43.0 docs/three-lines.md shows .assets/ktl-three-lines-dark.svg in GitHub’s dark theme, and CHANGELOG.md says the -dimmed set stays in .assets/ unreferenced. This concept says the diagram has a dimmed counterpart for dark mode.

2026-10-15

Changed: Why the skills live in their own repository: docs/install.md no longer says one install serves every repository or that a repository needs no copy; it says a user install keeps the agent’s instructions under the user’s control.
Changed: Why four skill roles: ktl-website now builds the schema’s pages and the data too, as docs/architecture.md says since 0.41.0, and the concept no longer says the roles lose nothing without it, which the page says only of ktl-prose.
Changed: ktl-website skill: what 0.43.0 added to the site (the full-text and Markdown files, the log feed, concept history and licence, hover previews, release links and reading levels), and the site’s address; the architecture page does not say the roles lose nothing without it.
Changed: ktl-sidecar skill: the scheduled librarian’s AGENT_PROMPT_FLAG, the five conditions for any other agent, the patch file written with the agent’s own file tool, the pull-request setting the publish job needs, and rule 13’s base_iri exception, as references/automation.md and references/gate.md now say.
Changed: ktl-librarian skill: the scheduled run’s prompt flag, the rule that the agent writes its patch with its own file tool, and the pull-request setting, as knowledge-librarian.sh and references/scheduled-task.md now say. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Releasing: sync-sidecar.sh works on a sync-<tag> branch in each sibling and sets its next steps apart, and the changelog fold lists each entry once, as 0.40.0 and 0.42.0 brought them.
Changed: Repository validation: check 11w also holds the website template’s logos equal to .assets/, check 14 proves the fold lists each entry once, and the website smoke test builds the schema’s pages and the data, as validate-repository.sh and smoke-test-website.sh now do.
Changed: The docent in Microsoft 365 Copilot: docs/m365.md now says an auditor would be ktl-auditor-m365.md, not that it is planned, and no longer says the identity record is held in reserve.
Changed: Accession: the quotation follows docs/architecture.md, which now says the registrar “keeps the records in order”.
Changed: Registrar: the quotation follows docs/architecture.md, which now says the registrar “keeps the records in order”.
Changed: Domain schema: docs/architecture.md and docs/for-the-curious.md no longer say “not a looser bundle”, so the term no longer does.
Changed: Doorway: docs/obsidian.md no longer says to open the link itself and never the host root, so the term cites ktl-docent’s references/obsidian.md, which still does.
Changed: Workshop: docs/obsidian.md no longer says what keeps the workshop clean, so the term cites ktl-docent’s references/obsidian.md, which still does.
Changed: Exhibition: the README’s Exhibitions section, added in 0.42.0, lists the sites that show a repository’s collection.
Changed: The source map: the README’s row names the Exhibitions and the site’s address, and the thirty-first pass is recorded. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: Security policy: Since 0.41.0 the ktl-website row of SECURITY.md says the build job also runs just when the site builds a folder through a recipe, and that the sync runs the commands site.config.mjs names to build the schema’s pages and the data. Since 2026-10-14 SECURITY.md also counts five surfaces, matching its table. This concept carries neither.
Open question: Threat model: Since 0.41.0 docs/threat-model.md says that ktl-website’s build also runs the commands its site.config.mjs names, or a gen-doc recipe it finds, to build the schema’s pages and the data, and that they are the repository’s code too. This concept does not say so.
Open question: Knowledge Trust Ladder in a skills repository: Since 2026-10-14 docs/skills-repositories.md says each of this repository’s six skills has a Playbook concept. This concept still says five.
Open question: Knowledge bundle: Since 2026-10-14 docs/obsidian.md sends readers to each plugin’s usage page, not its README, and says the plugins’ architecture pages name the cost of putting exhibits in a vault’s index. This concept says the plugin READMEs name the cost of a folder inside a vault. Which page names that cost now?
Open question: Build, preview and edit the website: Since 0.43.0 website/README.md says the site also serves the bundle log as an Atom feed, knowledge/log.xml, and the home, docs and extra pages as Markdown beside them, and that each release page links the repository, the docs and the bundle at the release’s tag. It names src/components/ReadingLevels.astro and src/components/ConceptPreviews.astro for the reading levels and the hover preview. This concept names none of them.

2026-10-14

Changed: ktl-website skill: the Schema topic, the data pages in Datasette Lite, the preflight’s recipe and just lines, the weekly run and its new triggers, and the smoke test’s schema and data, as 0.41.0 and 0.42.0 brought them.
Open question: Build, preview and edit the website: Since 0.41.0 (2026-10-14) knowledge-website.yaml also runs every Monday at 05:23 UTC from the default branch, and on a change to .linkml/, a justfile, a pyproject.toml or a uv.lock; the scheduled run deploys too. This concept lists the push and pull-request paths up to the workflow itself, and no weekly run.
Open question: Build, preview and edit the website: website/README.md now says a promoted release’s changelog page appears with the next build, the workflow’s weekly run at the latest, or a run by hand. This concept says the next build, or a run by hand.
Open question: Build, preview and edit the website: website/README.md now names site.config.mjs for the site’s values and options, src/styles/custom.css, this repository’s own, for its look, and src/components/Mermaid.astro for how a diagram is drawn, and says node scripts/config.mjs prints what the sync reads. This concept’s What to edit section names none of them.
Open question: Build, preview and edit the website: Since 0.42.0 (CHANGELOG.md) the README and the architecture page link the site at its persistent address, https://w3id.org/ktl/, which redirects to the GitHub Pages address this concept names.
Changed: base_iri moved from the knowledge-trust-ladder.example placeholder to https://w3id.org/ktl/knowledge/, now that the w3id rule resolves it. Every concept id, the publisher id and every typed-relation target moved with it; no verified event changed.
Changed: Knowledge sources: base_iri is now https://w3id.org/ktl/knowledge/, no longer the knowledge-trust-ladder.example placeholder. Edited since a person confirmed it; the label says so until the curator looks again.
Added: ktl-website skill, from skills/ktl-website/SKILL.md and its references, the sixth skill that 0.39.0 added.
Changed: Versioning policy: six skills under one tag, as docs/architecture.md and CONTRIBUTING.md now count since ktl-website. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Why the skills live in their own repository: six skills, as docs/install.md, plugin.json and marketplace.json now count since ktl-website.
Changed: Contributing: only website/ builds, two helpers (ktl-prose and ktl-website) and six skills under one tag, check 11w among the template pairs, and the website’s role boundary, as CONTRIBUTING.md now says.
Changed: Repository validation: six skill directories, ktl-website’s word budget, checks 11w and 12w, check 17 over every skill’s templates, the two website test modules, four workflow templates for actionlint and the smoke-test-website job, as validate-repository.sh and validate.yml now say.
Changed: Releasing: six skills under one tag since 0.39.0 (CHANGELOG.md), and sync-sidecar.sh also carries ktl-website’s templates to a sibling that has a site.
Changed: ktl-sidecar skill: Step 6 names ktl-website as the skill that can publish the bundle, and the scheduled librarian’s and the release’s variables go under Repository variables, never on an environment, as SKILL.md and references/automation.md now say.
Changed: Why four skill roles: the architecture page now names a sixth skill, ktl-website, which publishes the bundle and is optional like ktl-prose.
Changed: The source map: six skill playbooks, a row for ktl-website’s templates and preflight, the website workflow’s new name, the website smoke test, and the thirtieth pass’s notes. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: Security policy: Since 2026-10-14 SECURITY.md lists ktl-website’s templates and its preflight as a surface of their own. It names knowledge-website.yaml among this repository’s copies of the templates, counts the six skills’ prose, and names five skills that run .lokf/scripts/, ktl-website among them. This concept carries the earlier table and four skills. SECURITY.md itself still says four surfaces above a table of five rows.
Open question: Threat model: Since 2026-10-14 docs/threat-model.md names ktl-website among the skills that run the repository’s scripts from .lokf/scripts/, counts four workflow templates under the two skills’ templates/github/ as pinned, and says six skills read content they did not author. This concept still names four skills that run those scripts.
Open question: Hosts and doorways - where the bundle’s real folder lives: Since 2026-10-14 the portability matrix in skills/ktl-sidecar/references/portability.md says that on GitLab, Forgejo and Gitea the GitHub Actions workflows are lost, ktl-website’s deploy included. This concept says the three GitHub Actions workflows are the only loss there.

2026-10-13

Curation: human:noelmcloughlin corrected 1: Build, preview and edit the website, now naming knowledge-website.yaml, the default branch, the variable gate, docs/architecture.en-simple.md and DOCS.groups.
Curation: human:noelmcloughlin sent 1 back: Build, preview and edit the website, for the website.yml changes the ktl-website skill brings.
Changed: The context layer now says the architecture page uses the term and links the page, since “Why libraries have catalogues” and “Prose, Structure, Meaning, Context” moved there from the README on 2026-10-13. It is about the glossary term of the same name.
Changed: Intended uses takes the context layer and the question of who is responsible from docs/architecture.md, where the README’s sections moved on 2026-10-13, and keeps the README for what the project is.
Changed: Why a distribution repository now cites docs/install.md, which holds the quick start, the plugin commands and the docent-only route the README carried until 2026-10-13, and docs/architecture.md for the one-tag rule.
Rewrite: Why four roles follows docs/architecture.md, where the README’s roles moved on 2026-10-13, and answers the note sent back that day. Only KTL Registrar does the registrar’s work in Obsidian, and the concept now carries the Runs column, the curator always being a person, what ktl-prose never changes and the three lines of defence.
Changed: Why four roles cites docs/architecture.md and docs/obsidian.md, and relates to the three lines of defence its body links.
Changed: WikiSkill cites docs/architecture.md for the paper’s credit, which moved there from the README on 2026-10-13.
Changed: Accession cites docs/architecture.md, where the README’s fifth-role section moved on 2026-10-13, and relates to the collection an accession joins.
Changed: Desk cites docs/architecture.md, where the README’s fifth-role section moved on 2026-10-13, and gives the registrar’s work in Obsidian to KTL Registrar alone, as that page says.
Changed: Docent cites docs/architecture.md, where the README’s table of roles moved on 2026-10-13, and the retelling beside it.
Changed: LOKF cites docs/architecture.md, where the README’s format section and credits moved on 2026-10-13, and relates to LinkML, the language its schema is written in.
Changed: OKF cites docs/architecture.md, whose credits name Google Cloud as the specification’s creator since they moved from the README on 2026-10-13.
Changed: Registrar cites docs/architecture.md, where the README’s fifth-role section moved on 2026-10-13.
Changed: Sidecar cites docs/architecture.md, where the README’s table of roles moved on 2026-10-13.
Changed: The pen cites docs/architecture.md, where the README’s fifth-role section moved on 2026-10-13; the README no longer names the pen.
Changed: The pen for younger readers follows the pen, which now names the architecture page instead of the README.
Changed: ktl-prose skill cites docs/architecture.md for the fifth skill, which moved there from the README on 2026-10-13.
Rewrite: Versioning answers the note of 2026-10-13: it says what .releaserc.json does with ! beside what docs/releasing.md says, from commit-analyzer’s default preset and rules at the version the workflow pins. The one-tag rule cites docs/architecture.md, where the README’s reading list moved. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Versioning cites .releaserc.json, the release workflow and commit-analyzer’s README, default rules and Angular parser. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: Versioning policy: docs/releasing.md and the plan job in .github/workflows/semantic-release.yml treat ! after the type as breaking. .releaserc.json names no preset, and the default Angular preset of commit-analyzer, which the workflow pins at 13.0.1, has no place for ! in its header pattern, so a feat!: commit releases nothing. Which should change: the preset, or the page and the plan job?
Changed: Releasing now says which rules are .releaserc.json’s and which are commit-analyzer’s defaults, and that the Angular preset it reads commits with has no place for !.
Changed: The lokf toolkit answers the note of 2026-10-13. It no longer says the SHACL shapes catch violations, which CHANGELOG.md withdrew, calls the template’s lokf[build]>=0.8.0 a floor, and carries the librarian’s version check and the byte-identical schema. Its commands come from the toolkit’s cli.py at v0.8.0, and its extras and their uses from the pages that name them.
Changed: ktl-sidecar skill answers the note of 2026-10-13. The wrapper looks under .claude/skills/, .github/skills/ and .agents/skills/ before bare skills/, references/m365.md dates from 2026-09-24, and the concept now carries the repair rule, the base IRI rule, the four things gitignoring changes and why placeholders are substituted with Python.
Changed: Repository validation says check 9 now also checks each heading a sibling repository links to, as scripts/validate-repository.sh does since 2026-10-13.
Open question: When the built-in vocabulary stops fitting - domain schemas: Since 2026-10-13 the section “Prose, Structure, Meaning, Context” is on docs/architecture.md, not in README.md. This concept says “The README says” for its quotation and lists README.md in sources; both should name docs/architecture.md.
Open question: Why a registrar role, and why it is not a skill: Since 2026-10-13 the fifth-role section is on docs/architecture.md, not in README.md, so resource, sources and “The README calls it the fifth role” should name that page. That page and docs/obsidian.md now say KTL Registrar does the registrar’s work and KTL Curator assists the curator, while this concept and its description say both plugins do the registrar’s work.
Open question: Three lines of defence - where each role sits, and what an auditor can check: Since 2026-10-13 the fifth-role section that points at docs/three-lines.md is on docs/architecture.md, not in README.md. The sentence “The README’s section on the fifth role” and the README.md source should name docs/architecture.md.
Open question: Knowledge bundle: Since 2026-10-13 README.md keeps the poem, but “Why libraries have catalogues”, which says a bundle is the catalogue, is on docs/architecture.md. sources should list docs/architecture.md beside README.md.
Open question: ktl-curator skill: Since 2026-10-13 the Runs column (“a little, regularly”) is in the table of roles on docs/architecture.md, not in README.md. sources should list docs/architecture.md instead of README.md.
Open question: ktl-docent skill: Since 2026-10-13 the Runs column (“whenever anyone asks”) is in the table of roles on docs/architecture.md, not in README.md. sources should list docs/architecture.md instead of README.md.
Open question: Build, preview and edit the website: Since 2026-10-13 website/README.md says docs/architecture.en-simple.md becomes the page for readers aged 13 to 18, and that the home page’s tagline is the first line of docs/architecture.md that is bold throughout; .github/workflows/website.yml no longer lists README.en-simple.md. This concept still names README.en-simple.md. CHANGELOG.md changed too, only in its unreleased notes, and still records the site under the 0.38.1 release.
Changed: The LinkML reference is titled LinkML (linkml.io), as its siblings name their sites, so the index tells it from the glossary term.
Added: Librarian, a role the glossary lacked, from the table of roles on docs/architecture.md and ktl-librarian’s SKILL.md.
Added: Curator, a role the glossary lacked, from the table of roles on docs/architecture.md, the verdicts in docs/for-the-curious.md and the plugin table in docs/obsidian.md.
Added: Collection, the word the README’s opening paragraph sets in bold, from the README and docs/architecture.md.
Added: Context layer, the term docs/architecture.md sets in bold, beside the explanation of the same name, with docs/context-layer.md’s definition.
Added: Trust ladder, the project’s own name for how far each claim has been checked, from docs/architecture.md.
Added: Workshop, one of the two vaults docs/obsidian.md sets in bold.
Added: Exhibition, one of the two vaults docs/obsidian.md sets in bold.
Added: Doorway, the name docs/obsidian.md gives the knowledge_bundle link, with ktl-sidecar’s rule for a gitignored .lokf/.
Added: Health line, from the curator’s trust-fields.md and SKILL.md, the report script and the architecture page.
Added: Work list, from ktl-librarian’s SKILL.md and docs/for-the-curious.md.
Added: LinkML, the schema language LOKF is written in, from docs/architecture.md, docs/for-the-curious.md, ktl-librarian’s SKILL.md and the curator’s domain-schemas.md. The LinkML reference defines it.
Added: Domain schema, the term docs/for-the-curious.md defines, with what the architecture page says it gives a domain.
Added: Three lines of defence, the model docs/architecture.md sets in bold, from docs/three-lines.md.
Changed: The source map follows the README’s move: docs/architecture.md joins the README’s row, the retelling has its new path, the install count, the card and the release rows follow their pages, and the twenty-ninth pass is recorded. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Docent cites the retelling for readers aged 13 to 18 at its new path, docs/architecture.en-simple.md.
Curation: human:noelmcloughlin confirmed 2 concepts, corrected 16, sent 5 back and marked 1 for later. Two corrections added files the concept drew on to sources, two moved a resource to the page that holds the facts, ten restored facts a summary had dropped from its source, and two send-backs name claims from files the concept does not list, so the librarian may be citing too few sources and shortening too far.

2026-10-12

Prose: reworded 2 concept bodies in plain English: the context layer, whose pin-or-float detail now follows the list of decisions, and the website playbook. No fact, link, number or frontmatter changed.
Changed: The context layer follows docs/context-layer.md again: its new title and federation heading, and the pin-or-float decision, which the page now settles by pinning and which absorbs the staleness bullet. Trust at read time moved from the decisions into the workspace section, as on the page.
Changed: Domain schemas quotes the README as it now reads, which says a domain “extends the schema” without naming LinkML.
Added: Build, preview and edit the website, from website/README.md and .github/workflows/website.yml, which the 0.38.1 release added.
Changed: The source map gains a row for website/ and website.yml, which feed the new website playbook. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: docs/context-layer.md and docs/context-layer-technology.md now write catalogue, as the README and the rest of the bundle do. The concept and its description follow them; product names such as Unity Catalog keep their own spelling.
Changed: Knowledge sources records the twenty-seventh pass. Edited since a person confirmed it; the label says so until the curator looks again.
Prose: reworded 2 concept bodies in plain English: WikiSkill and SkillWiki, after the librarian rewrote both. No fact, link, number or frontmatter changed.
Changed: WikiSkill takes the page’s new title, and its description names what the page now holds.
Changed: SkillWiki takes the page’s new title, and its description names what the page now holds.
Rewrite: docs/wikiskill.md was trimmed to what KTL took from the paper and what the paper leaves open, without the part-by-part table or the programs, models and people table. The concept follows it, under the page’s new title.
Rewrite: docs/skillwiki.md was trimmed to the four changes, the excerpt check and where the designs part, without the side-by-side table, the programs, models and people table or the paper’s open points. The concept follows it, under the page’s new title.
Changed: docs/skills-repositories.md no longer reads the two papers against KTL’s design, and now gives the handling before the two runs. The overview says so.
Changed: docs/context-layer.md no longer quotes Tellius, Andreessen Horowitz, Forrester, Gartner and the other vendors, nor the analogies in its workspace steps. The concept drops the two quotations it carried.
Changed: docs/three-lines-critics.md is no longer marked as preliminary research, and the concept no longer says it is.
Changed: Knowledge sources: the docs/wikiskill.md, docs/skillwiki.md and docs/skills-repositories.md rows follow the trimmed pages, and the twenty-sixth pass is recorded. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Domain schemas now relates to the curator skill instead of naming it as its subject.
Changed: Three lines of defence now relates to the two role explanations instead of naming them as its subject.
Changed: Versioning now relates to the releasing playbook instead of naming it as its subject. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar skill now cites commit 61b9355, which split gate.md out of automation.md on 2026-09-24.
Changed: The source map’s templates row now names every file the sidecar writes, and re-checks the justfile, which carries the domain-schema variable since 0.38.0. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Trust label keeps definedBy as an empty list, as every relation field is a list.
Changed: Context layer now names all three gaps docs/context-layer.md says OKF leaves open, and the page’s recommendation on cross-bundle links.
Changed: Hosts and doorways now gives both halves of what Obsidian’s Excluded files does, as skills/ktl-sidecar/references/portability.md states them.
Changed: Intended uses now says, as the README does, that the context layer is what is missing and the bundle is its catalogue.
Changed: Three lines of defence now matches docs/three-lines-critics.md on how its sources are cited and on who produces evidence that oversight works, and relates to the context layer it links.
Changed: Why a distribution repository’s description now names the Microsoft 365 route and the rename its body covers, and its sources list .claude-plugin/plugin.json.
Changed: Why four roles no longer calls every handoff a frontmatter fact, since one is a feedback entry, and says the pen marks new concepts as drafts.
Changed: WikiSkill now quotes docs/wikiskill.md on the concepts readers keep arriving at, and relates to SkillWiki, which it sets itself apart from.
Changed: OKF now lists README.md, the source that names Google Cloud as the specification’s creator.
Changed: Trust label now references the OKF specification rather than naming it as the definition, since the labels are defined in trust-fields.md and the specification defines only the fields.
Changed: The pen for younger readers now relates to the registrar its body links.
Changed: ktl-librarian skill now covers its plain-English rules and the offer of a prose pass before hand-off, and carries typed relations to the security policy and ktl-prose that its body names. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-prose skill now numbers its steps 0 to 5 as SKILL.md does, says step 0 runs the conventions script too, and drops an override claim no source makes.
Changed: ktl-sidecar skill now says the validate job passes the merge-base, as knowledge-registrar.yaml has done since 0.36.1.
Changed: Open the bundle in Obsidian now lists Obsidian’s symlinks help page, the source of the sentence it quotes.
Changed: Releasing now points at the paragraph that describes check; the page has no next section.
Changed: Threat model now lists knowledge-conventions.sh, the source of its claim about conventions rule 10.
Changed: Code of conduct’s description no longer promises the behavioural standards, which the body does not list.
Changed: Open Skills CLI’s description now says what the body covers: the installer and its local-path use as a pre-publication check.
Changed: The source map no longer says all three release files sit behind the release Environment; only the release job does. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: Knowledge bundle: the body’s library and museum terms, the workshop and the log and index details come from README.md, docs/obsidian.md and skills/ktl-librarian/SKILL.md, but sources names only skills/ktl-sidecar/templates/README.md. Should those three be added as sources?
Open question: Why a registrar role, and why it is not a skill: the body’s account of the pen (stamps generated, keeps index bullets equal, files the log line, refuses some writes) and of the wrapper comes from skills/ktl-librarian/SKILL.md and skills/ktl-sidecar/SKILL.md, which are not in sources. The pen also refuses deleting a concept a person left a note on, not only one a person confirmed. And about names the knowledge bundle while glossary/registrar describes this role. Should these change?
Open question: Agent Skills specification: the body says the description is kept near ~100 tokens, but the specification’s ~100 tokens is a startup budget for name and description together, and this repository enforces only the 1024-character limit (check 3c); its descriptions run to about 150 to 230 tokens. The resource is agentskills.io/home while the facts come from agentskills.io/specification. Should both change?
Prose: reworded 22 concept bodies in plain English. No fact, link, number or frontmatter changed.
Changed: the README’s context layer now links docs/context-layer.md, which the concept now cites. That page, like the README, names no intended use or non-goal, so the open question stays.
Changed: docs/three-lines.md now lists the context layer’s domain tiers and workspace among what remains for this project, and the concept follows it.
Changed: README.md no longer calls a docent a museum’s name for a guide. README.en-simple.md does, and the concept now cites it.
Changed: README.md no longer says each skill stands alone, and no other source does, so the concept no longer says it.
Changed: scripts/sync-sidecar.sh no longer counts ai-linkmo among the siblings it brings up to a release, and the concept no longer does either.
Changed: the source map names docs/context-layer.md and docs/context-layer-technology.md, which the new context-layer concept follows. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: OKF specification (v0.2): the resource is the copy under knowledge-catalog/okf/, whose folder’s README now says OKF lives in its own repository and asks readers to stop using that copy. The README, the sidecar’s README template and CHANGELOG.md now link https://github.com/GoogleCloudPlatform/open-knowledge-format/blob/main/SPEC.md. A person wrote this concept, so moving its resource is the curator’s to record.
Added: The context layer, from docs/context-layer.md and docs/context-layer-technology.md, which the README now links.
Changed: check 16 is described without the repository’s old name, which the check itself refuses outside the files that record history.
Changed: ktl-librarian now names the domain schema by one schema line in the justfile, has pages for a host of Agent Skills and for retellings, and the scheduled run validates through just lokf-check-refs. references/domain-schema.md, references/agent-skills.md, references/retellings.md and knowledge-librarian.yaml say so. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: ktl-sidecar’s page now names the fourteenth convention, the curator’s queue in the report script, and what is tested before a release. references/gate.md, references/portability.md and the CHANGELOG entries for 0.36.0 and 0.37.1 say so.
Changed: ktl-curator’s page now says the script ranks the queue and counts reliance through thirteen relation fields, reads the justfile’s schema line, offers a Younger readers: switch, and asks what replaced a retired concept. SKILL.md, references/trust-fields.md and references/review-session.md say so.
Changed: ktl-docent’s page now follows all thirteen typed relations, names a retired concept’s successor and the label for a reader’s dispute, passes --concept with a Disagreement, and answers a younger reader. SKILL.md, references/answering.md and references/feedback.md say so.
Changed: ktl-prose’s page now covers retellings for younger readers and the pass for dyslexic readers, with prose-check.py --audience, --derived-from and the VOA word book. references/young-readers.md, prose-check.py and the 0.38.0 CHANGELOG entry say so.
Changed: changelog-release.mjs check now passes an empty [Unreleased] for a version already promoted, as .releaserc.json and the script say since #114, and publish.yml runs the pytest suite under tests/ before it publishes.
Rewrite: the behavioural checks moved out of scripts/validate-repository.sh into pytest modules under tests/, which validate.yml runs in a job of its own (0.37.1). The page now lists the contract’s checks as the script numbers them, says which module carries each moved check, and adds checks 7c, 7d, 10a and 22.
Changed: CONTRIBUTING.md now asks for pytest beside the contract script before a pull request, lists tests/ in its layout, and says the skills are written partly in Python, not in one Python script.
Changed: the limits list now names skills/ktl-sidecar/references/m365.md as the source of Agent Builder’s 50 MB zip limit, beside the toolkit’s 10 MB app package the builder enforces.
Changed: A domain schema is now one schema variable in the sidecar’s justfile, which the validate and check-refs recipes and the three workflows use, as skills/ktl-librarian/references/domain-schema.md and the justfile template say since #116.
Changed: docs/skills-repositories.md now says what KTL does and leaves alone in a repository of Agent Skills; no source yet states intended uses or non-goals in general, so the question stays open.
Changed: The install commands now install for the user (--global, --scope user), and docs/install.md gained a section on when a team installs into the repository instead (#115).
Changed: docs/wikiskill.md now names WikiSkill apart from SkillWiki, the 2026 paper docs/skillwiki.md reads, and the concept says so.
Changed: the trust labels now include a reader disputed this, read from .lokf/feedback.md, and derived from a concept edited since its confirmation, and a retired concept names its successor from log.md; skills/ktl-curator/references/trust-fields.md says so.
Changed: SECURITY.md now counts four skills that run the repository’s .lokf/scripts/, ktl-prose among them, and says prose-check.py also reads the skill’s own word list. Every other claim was checked against SECURITY.md and scripts/validate-repository.sh check 10 and still holds. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: docs/threat-model.md now names ktl-prose among the skills that run .lokf/scripts/, adds conventions rule 14, says what a Disagreement’s concept path is used for, and covers a skills repository. The concept follows all four.
Changed: the source map names the files tracked since the 2026-10-06 refresh (tests/, pytest.ini, README.en-simple.md, docs/skillwiki.md, docs/skills-repositories.md, docs/examples/prose.md, ktl-prose’s word list and three new references) and follows the README’s renamed agents callout. Edited since a person confirmed it; the label says so until the curator looks again.
Added: SkillWiki’s skill governance and KTL, from docs/skillwiki.md, which no concept covered.
Added: KTL in a skills repository, from docs/skills-repositories.md, which no concept covered.
Prose: after the librarian’s refresh, reworded 3 concept bodies in plain English: repository validation now says which test module tests each moved check, and two long sentences in the SkillWiki and skills-repository concepts became lists. No fact, link, number or frontmatter changed.

2026-10-11

Added: The pen, retold for readers aged 13 to 18 from glossary/the-pen.md, the first retelling for younger readers in this bundle.

2026-10-06

Prose: after the librarian’s refresh, reworded 2 more concept bodies: a long sentence in the ktl-prose skill split in two, and the lokf toolkit’s last paragraph split where it turns to the no-Python fallback. No fact, link, number or frontmatter changed.
Changed: ktl-prose now lays out what it rewords the way the file’s project does, one line for each paragraph and list item in the bundle, as skills/ktl-prose/SKILL.md and references/rules.md now say. The concept says so in its rules, its steps and its note on files outside the bundle.
Changed: since 2026-10-05 the no-Python fallback reads LOKF’s schema from the commit tagged v0.8.0, never from the tag, as skills/ktl-librarian/SKILL.md and CHANGELOG.md say; both are now among the concept’s sources. The description now has no dash.
Changed: check 11’s synthetic bundle also breaks rule 12 with an index bullet left behind by an edited description, as scripts/validate-repository.sh shows, and the list of breaks now includes it. The description now has no dash.
Changed: the body cites check 9 of scripts/validate-repository.sh, which lists docs/threat-model.md among the paths a sibling depends on, so that script is now among the concept’s sources. The description now has no dash.
Changed: the install manual does not say the command group may change without notice; the v2.90.0 release notes do, and the concept now credits them. The publish manual and the two workflows that call gh skill publish are now among its sources.
Changed: the contributing playbook’s description, which both index bullets copy, now reads in plain English, with no dash.
Changed: the ktl-curator skill’s description, which both index bullets copy, now reads in plain English, with no dash.
Changed: the ktl-docent skill’s description, which both index bullets copy, now reads in plain English, with no dash.
Changed: the code of conduct’s description, which both index bullets copy, now reads in plain English, with no dash.
Changed: The docent records a miss through its own copy of knowledge-feedback.sh, as skills/ktl-docent/SKILL.md now says. The concept’s quotation that a docent explains the exhibition without moving anything on the shelves now comes from the README’s table of roles.
Changed: docs/install.md no longer counts four skills or calls ktl-prose a helper, and points to the README’s install section instead. The concept now says what README.md says: each skill stands alone, and ktl-prose is optional and installs the same way.
Changed: README.md no longer says that a person confirms drafts a few at a time, so the run order drops the phrase. Its docent row now says that, like a real docent, the docent explains the exhibition without moving anything on the shelves, and the docent bullet says so too. The sidecar bullet says the skill installs .lokf/, the README’s own verb.
Changed: skills/ktl-curator/references/trust-fields.md now reads a verified key with no events as nobody has checked this yet, as knowledge-report.sh does, and the concept says so.
Changed: README.md now makes the taxonomy point in one sentence: LOKF supplies the fixed taxonomy that OKF leaves to the producer, and a domain that outgrows it extends the schema. The concept follows it, and no longer says that any domain can bring its own, which no source now states.
Changed: The concept now lists docs/three-lines-critics.md, which it summarises, and CHANGELOG.md, which dates the critics page’s arrangement, as its sources. It drops a date that no source gives, and says holds, installs and released where the pages do.
Changed: docs/wikiskill.md now leaves the pen’s refusals, conventions rule 13 and revision to the threat model and the three-lines page, and says once what it said twice. The concept follows it, takes the page’s labels for two findings, and says merges where it said lands.
Changed: skills/ktl-sidecar/SKILL.md says that every skill addresses .lokf/knowledge, where the table counted four skills, and that Step 2 adds the doorway.
Changed: CONTRIBUTING.md now has a template change copied over each of this repository’s copies, ktl-docent’s included, under checks 11 and 11a, and names the two plugin repositories in the sync command.
Changed: SECURITY.md now counts ktl-docent’s two scripts among the scripts a skill runs in place, and says which skills run the repository’s .lokf/scripts/ and that ktl-docent runs only its own. Checked against SECURITY.md again, as the note of 2026-10-03 asks. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: scripts/validate-repository.sh adds check 11a, which holds ktl-docent’s own copies of two sidecar scripts to their templates and runs them as an installer leaves them, and check 11b, which holds the no-Python schema URL to one commit. The smoke test now checks that the docent’s scripts come with the install and work in a reader’s repository.
Changed: docs/threat-model.md now says which skills run the repository’s .lokf/scripts/, that ktl-docent runs only its own copies of two of them, and that the no-Python schema is read by commit. The concept follows it, and says that the release Environment, not the gate, exists in name only without reviewers.
Changed: skills/ktl-curator/SKILL.md now keeps only the username from glab api user, because the whole profile includes the person’s email, and the concept says so.
Changed: skills/ktl-librarian/SKILL.md now reads the no-Python schema from a commit, not a tag, and answers Snyk’s W012 finding, so the concept adds that note. Its W011 paragraph now says, as the skill does, that an entry can neither add a confirmation nor remove one. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: skills/ktl-sidecar/SKILL.md now says that ktl-docent runs its own copies of the feedback recorder and the report script, so the copies Step 5 installs serve a person, an earlier ktl-docent, ktl-curator and ktl-librarian.
Changed: Since 0.34.1 the Copilot docent’s gap line starts a Miss with the reader’s question and says which of two cases it is, and a Disagreement says which version the answer gave, as ktl-docent-m365.md states. Where a release has no zip, the concept now says the repository’s sidecar has no .lokf/m365/, as docs/m365.md does.
Changed: The source map now names skills/ktl-docent/scripts/, the docent’s own copies of two sidecar scripts added in 0.34.2, which no row accounted for. Edited since a person confirmed it; the label says so until the curator looks again.
Rewrite: ktl-docent now runs its own copies of knowledge-report.sh and knowledge-feedback.sh, never the repository’s, names the asker in an entry only when the asker agrees, and labels a retired or edited concept as the report script does. The concept is rewritten from skills/ktl-docent/SKILL.md and its references as they now stand, with the skill’s new scanner note.
Prose: reworded 4 concept bodies in plain English: the ktl-docent skill, three lines of defence, WikiSkill’s loop and the threat model. No fact, link, number or frontmatter changed.
Prose: reworded 21 more concept bodies in plain English, among them repository validation, releasing and the three skill playbooks, with long runs of checks, steps and conditions turned into lists. No fact, link, number or frontmatter changed.
Prose: a full pass over the 25 concept bodies reworded today puts each paragraph and list item on one line, as ktl-librarian writes them, and makes 30 sentences plainer. No fact, link, number or frontmatter changed.

2026-10-05

Changed: the skills pin is two values now, TRUST_LADDER_SKILLS_REF and TRUST_LADDER_SKILLS_SHA, and the release step in semantic-release.yml and scripts/sync-sidecar.sh each move both, as docs/releasing.md says. The page no longer says that check 11 ignores one value.
Changed: scripts/validate-repository.sh has twenty-one numbered checks, and checks 11, 12, 15, 18, 19 and 20 now test more, such as the commit pinned beside the skills tag and the .lokf/uv.lock the gate installs from. The page now also describes check 16b, and check 12 counts five other scripts under sh.
Changed: the row for this repository’s own sidecar copies names both values of the skills pin, since check 11 and the preflight now ignore TRUST_LADDER_SKILLS_SHA too. Three rows follow README.md: docs/for-the-curious.md also says who writes the bundle, the README no longer shows the lifecycle diagram, and its pointer for agents is in the Prefer to ask? callout. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: skills/ktl-curator/references/portability.md says the key that signs is often a subkey, and names subkeys only in GitHub’s key listing. The concept said usually, and gave the field to every forge.
Changed: skills/ktl-docent/SKILL.md says the scheduled workflow commits feedback.md often into a public pull request. The concept said only that the pull request can be public.
Changed: the hand-off now gives the words before and after from the OK line of prose-check.py, as step 5 of skills/ktl-prose/SKILL.md asks.
Changed: the page now names two more refusals of the pen, the commit the skill pin checks, and the conventions run in refresh, as skills/ktl-librarian/references/patch.md and .github/workflows/knowledge-librarian.yaml state. It also says that a scheduled run stops while a pull request the workflow opened is still open, and what the quiet check no longer counts, as .lokf/scripts/knowledge-report.sh states. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the page now says that a scheduled run stops while a pull request the workflow opened is still open, and proposes nothing again that a person declined, as skills/ktl-sidecar/references/automation.md states. It also says that the skill pin names a commit beside its tag, which the install step checks and the preflight’s copies line leaves out of its comparison, as .github/workflows/knowledge-librarian.yaml and .lokf/scripts/knowledge-preflight.sh state.
Changed: the pen now refuses a patch that would change a person’s record in any concept it writes, and a patch that handles more than ten feedback entries. skills/ktl-librarian/references/patch.md and the header of knowledge-apply.py say so.
Changed: the concept now names the inputs of the label a source moved since the confirmation, which skills/ktl-curator/references/trust-fields.md lists as resource, sources[].resource and git history.
Changed: the surface table now says separate jobs, not two jobs, mean the agent never meets a write token, as SECURITY.md says. Every other claim still matches SECURITY.md, check 10 of scripts/validate-repository.sh and CHANGELOG.md. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the concept now names four measures docs/threat-model.md added: the skills pin named by commit, the toolkit and parser installed as reviewed, the librarian’s three jobs, and the guards on the earlier job’s reading of pull requests. It also names the pen’s check of a person’s record and the hand-off written again after the retrieval call. The pen’s date now follows CHANGELOG.md 0.29.0, and the pen is named as the librarian’s only way to write the bundle rather than the bundle’s only writer.
Changed: scripts/sync-sidecar.sh now also moves TRUST_LADDER_SKILLS_SHA to the commit the tag names, as CONTRIBUTING.md says.
Changed: three phrases now match skills/ktl-curator/references/domain-schemas.md. That page says LOKF has a small vocabulary and never says it is small on purpose, and its signs are many producer-defined keys and a wrong or ambiguous field.
Changed: three claims that no source states are removed: that most readers of a code repository never open Obsidian, that the vault lacks the bundle on mobile, and that the problem showed in the maintainer’s own vault. The notes-vault bullet now ends with the reason docs/obsidian.md gives, that this keeps the workshop clean.
Changed: the context layer is defined as README.md defines it, a governed layer that persists from one task to the next. The open question stays, because neither README.md nor any page under docs/ states which uses KTL is meant for and which it is not.
Changed: the loop now ends as README.md says, with the curator reviewing what is new or changed. The README no longer shows .assets/ktl-lifecycle-loop.svg, so the concept no longer points at it.
Changed: docs/wikiskill.md no longer says the paper’s wiki keeps every trace: the paper keeps them, and its wiki’s log names the errors that came back. The concept now says the same, and adds what the page now says about the pen’s check of a person’s record, declined pull requests, the pen’s limit of ten feedback entries and the questions the retrieval score leaves out.
Changed: step 2 of the two-vault setup now says what docs/obsidian.md says: there is nothing to configure, because the root index.md carries the bundle’s header. No repository source holds the phrase the concept quoted as the plugins’ default.

2026-10-04

Added: Accession, a museum’s word for an item added to its collection, from README.md, which now says so where it describes the registrar.
Changed: Registrar: its quotation of the README carries the README’s new gloss on accession.
Added: Registrar, the person in a museum who keeps the collection’s records, from README.md, which now says so where it first names the role.
Added: Sidecar, the .lokf/ folder beside the code and the role that installs it, from README.md, which now says so where it first names it.
Added: Docent, a museum’s name for a guide, from README.md, which now says so where it first names the docent.
Added: Desk, the README’s word for the place where a role does its work, from its section on the registrar.
Changed: ktl-prose skill: rule 3 covers a paragraph past 150 words, rule 9 asks for the literal verb, and the check reports both and a character no reader sees.
Changed: repository validation: check 18 exercises the paragraph, figure-of-speech and unseen-character findings, and check 21 searches every tracked file for a character a reader cannot see.
Changed: threat model: this repository’s contract fails on a character a reader cannot see, and ktl-prose’s check refuses a rewording that adds one.
Changed: security policy: SECURITY.md now says what guards each surface, and that a security fix is published as a patch release. Checked against it again, as the note of 2026-10-03 asks. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the pen: the README now names it as the registrar for the librarian.
Changed: contributing: CONTRIBUTING.md now says to merge this side first, and that a curator’s key is added first in a pull request of its own.
Changed: the skill page gives the Golden Rules in brief and keeps their tables, the first run’s sweep and the audit’s detail in three new references; a run’s hand-off reaches its pull request, a quiet week runs no agent, and a note the run reads leaves its stamp. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: the pen also takes the librarian’s hand-off, which it cleans and writes nowhere in the bundle, as references/patch.md now says.
Changed: check 11 now proves conventions rule 13, check 19 the pen’s hand-off and check 20 the quiet check, and the layout tests the wrapper’s quiet skip and hand-off; the playbook counts thirteen conventions, eight of them in the Python half.
Changed: the sidecar’s references now describe the quiet week, the hand-off’s path to the pull request, and the gate’s thirteenth convention.
Changed: the threat model adds the librarian’s hand-off as a guarded path to the pull request, and conventions rule 13 under human attribution.
Changed: docs/wikiskill.md now names conventions rule 13, the hand-off that carries a repeated send-back to the pull request, and the quiet week that bounds a run further.
Changed: an auditor’s questions now include whether a confirmed text is the one its person saw, which conventions rule 13 answers.
Prose: Releasing; the sentence on the sync script is split into six; no fact changed.
Changed: WikiSkill’s loop and Knowledge Trust Ladder; docs/wikiskill.md now names the design KTL and keeps the bundle for its corpus and the wiki for the paper’s, and the concept follows it.
Changed: Three lines of defence; the roles are KTL’s and the pages now say so, where they named them after LOKF, the format.
Changed: Threat model; the three repositories that share it are KTL’s, not LOKF’s.
Changed: Contributing; the sibling repositories are KTL’s, as CONTRIBUTING.md now says.
Changed: Releasing; the three repositories are KTL’s, and the sync script’s every-host templates now include one from ktl-sidecar’s Step 1.
Added: The pen, the name the README and docs/for-the-curious.md give knowledge-apply.sh, the librarian’s only way to write the bundle.
Changed: Releasing; the sync script now lays down a template a sibling lacks when Step 5 puts it on every host or a copy there needs it, as scripts/sync-sidecar.sh and docs/releasing.md say.
Changed: Repository validation; check 19 now also holds the pen to a root index shaped by hand, which keeps its ## sections and its lines that list several concepts.

2026-10-03

Prose: Threat model: one sentence split; no fact changed.
Prose: ktl-sidecar skill: one sentence split; no fact changed.
Prose: ktl-docent skill: one sentence split; no fact changed.
Changed: ktl-librarian skill: a refresh starts from the work list, a handled entry moves to the ledger with its question, the pen has eight operations and prints its own format, and publish refuses a change to a person’s record, as skills/ktl-librarian/SKILL.md, references/patch.md and references/scheduled-task.md now say. Edited since a person confirmed it; the label says so until the curator looks again.
From reader feedback: ktl-librarian skill: step 6 of skills/ktl-librarian/SKILL.md names .lokf/patch.yaml as the one file the scheduled agent writes, so the skill page no longer lags the wrapper and the question is withdrawn.
Changed: ktl-sidecar skill: Step 5 lays down nine scripts and fourteen files, knowledge-report.sh among them and among the three that land regardless; the gate reference covers a removed confirmation and the unattended check, and the automation reference the retrieval score.
Changed: ktl-curator skill: Step 1 quotes knowledge-report.sh, lists apart the questions a later confirmation answered and names the confirmed concepts whose source moved, as skills/ktl-curator/SKILL.md and references/trust-fields.md now say.
Changed: ktl-docent skill: a label comes from knowledge-report.sh labels, and a handled entry moves to the ledger, as skills/ktl-docent/SKILL.md and references/feedback.md now say.
Changed: Threat model: a removed confirmation needs its person, publish reads a person’s record off the patched tree, the pen refuses to delete a concept a person left a note on, and a handled entry moves to a ledger only programs read, as docs/threat-model.md now says.
Changed: Repository validation: twenty checks and nine sidecar scripts, as scripts/validate-repository.sh now counts; checks 13, 13a, 19 and 20, the twelfth conventions rule and the layout tests’ new cases are described.
Changed: Knowledge sources: the sidecar lays down nine scripts, knowledge-report.sh among the fifteen dogfooded copies, and the docs/wikiskill.md row follows the page’s present sections. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Three lines of defence: the second line names knowledge-report.sh, the gate asks the person behind a confirmation that is removed, and the report lists a confirmed concept whose source moved, as docs/three-lines.md now says.
Changed: WikiSkill’s loop and the bundle’s ladder: the ablation ran on four benchmarks, and the rows and findings name the ledger, the work list, the report script, the removal rule and the retrieval score, as docs/wikiskill.md now does.
Changed: Why four skill roles: the registrar’s parts include knowledge-report.sh at the reader’s desk and the curator’s, as README.md now says.
Changed: Trust label: knowledge-report.sh computes the labels, and a concept edited since its confirmation is counted once, as skills/ktl-curator/references/trust-fields.md now says.
Open question: Why a registrar role, and why it is not a skill: README.md now says the gate ties each confirmation that is added or removed to its person, and names knowledge-report.sh as the registrar at the reader’s desk and the curator’s; this text, which a person wrote, says each new confirmation and names neither.
Changed: Security policy, re-checked against SECURITY.md as the curator’s send-back of 2026-10-03 asked: the body follows the file section by section and carries qualifiers it had dropped (a first reply, harden-runner in audit mode, actions pinned to commit SHAs, the release Environment, with scope and reproduction, the gate proving who vouched, not what they read); the word counts are the ones check 10’s comment gives. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Knowledge sources: the issue-template row says both templates name five skills, as the curator’s send-back of 2026-10-03 asked, and the plugin row counts five skills, as README.md does; the other “four skills” in the bundle follow sources that still say four, with ktl-prose as a helper. Edited since a person confirmed it; the label says so until the curator looks again.
Curation: human:noelmcloughlin confirmed 2 concepts (ktl-librarian skill, versioning policy), corrected 1 (Why a registrar role: the title now ends “why it is not a skill”, since the README’s fifth skill is ktl-prose) and sent 2 back (Knowledge sources: the issue-template row still says “all four skills”; Security policy: check it against SECURITY.md again). Two of the three verdicts were the same kind, a skill count the move to five skills left behind, which is worth a line in ktl-librarian’s instructions.
Changed: The docent in Microsoft 365 Copilot: the release asset’s <repository> may be the name KNOWLEDGE_RELEASE_NAME sets, as docs/m365.md says since #86, and the atk validate line quotes the error the page quotes.
Open question: The docent in Microsoft 365 Copilot: Microsoft’s page on custom skills (updated 2026-09-04) says the Agents Toolkit takes a skill directory and that .zip isn’t supported. skills/ktl-sidecar/references/m365.md says atk add skill --from takes a zip from any path, as checked against the toolkit’s source, and this concept follows it. Which holds for the toolkit a reader installs?
Changed: Why the skills live in their own repository: the old repository name is left to docs/install.md, which records it; the contract keeps it out of every other file.
Changed: ktl-librarian skill: seven pen operations, with reindex, as references/patch.md now gives. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: When the built-in vocabulary stops fitting: the description is the catalogue entry, cut to one sentence.
Changed: Hosts and doorways: the description is the catalogue entry, cut to one sentence.
Changed: What Knowledge Trust Ladder is for: the description is the catalogue entry, cut to two sentences.
Changed: Three lines of defence: the description is the catalogue entry, cut to one sentence.
Changed: The docent in Microsoft 365 Copilot: the description is the catalogue entry, cut to one sentence.
Changed: Open the knowledge bundle in Obsidian: the description is the catalogue entry, cut to one sentence.
Changed: Releasing: the description is the catalogue entry, cut to one sentence.
Changed: Why a registrar role: the description no longer says “rather than by an agent”, which README.md does not say (the curator’s send-back of 2026-09-26); the body follows the README’s registrar section as of f1f0d51, which dropped the two plugins’ “levels of checking” and folded the three-lines paragraph under it, and the undated pen is no longer dated. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: The docent in Microsoft 365 Copilot: the limits sentence now gives both sizes the sidecar’s references/m365.md states, the toolkit’s 10 MB app package and Agent Builder’s 50 MB zip (the curator’s send-back of 2026-09-26); four qualifiers docs/m365.md does not state are dropped (“fifth skill”, “at the bundle’s commit time”, “the toolkit’s preview”, “check 17 word for word”).
Changed: Knowledge sources: the rows name bug_report.md, feature_request.md, docent.md and curator.md instead of their globs, and .gitignore joins the LICENSE row (the curator’s send-back of 2026-09-26); the skills row counts five playbooks and the references row admits prose-check.py; a row for .markdownlint-cli2.jsonc and lychee.toml, which had none; a note for this pass. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: WikiSkill’s loop and the bundle’s ladder (was “What WikiSkill says about a knowledge bundle”): docs/wikiskill.md was restructured in f1f0d51 around a part-by-part table, eight findings, and where each design puts programs, models and people; the determinism section and the four gaps it named are gone from the page, so they are gone here.
Changed: Three lines of defence: the second line names knowledge-apply.sh as the librarian’s only pen and its refusal of a human: actor, as docs/three-lines.md does since f1f0d51, and the README pointer names the section that now holds it.
Changed: gh skill (GitHub CLI): the manual’s --from-local and its preview notice are carried, the v2.90.0 claim now cites the release notes that state it, and a dependency claim no page supports is dropped.
Changed: LinkML: the class count is dropped; the specification site lists fourteen types and the 0.8.0 schema’s Concept descendants plus Person and Organization make fifteen, so the page states neither.
Changed: LOKF toolkit: the 0.8.0 command line’s other subcommands and the [tables] extra are named, and an unverified claim about upstream main is dropped.
Changed: Why four skill roles: the README now names ktl-prose as the fifth skill that is not a role and the registrar as toolkit, gate, pen and plugins, and its “in short” summary line is gone since f1f0d51.
Changed: Why the skills live in their own repository: five skills, as README.md, docs/install.md and plugin.json now count; two dates and two check numbers no named source states are dropped.
Changed: When the built-in vocabulary stops fitting: the OKF sentence follows the README’s wording since f1f0d51, and the graph note carries how 0.8.0 and 0.9.0 project an undeclared type, as the librarian’s domain-schema.md states.
Changed: ktl-prose skill: “the roles lose nothing without it”, as the README says, not “the four roles”.
Changed: Contributing: one Python script and five skills, as CONTRIBUTING.md says; the pull request’s why, the ktl-prose boundary, the sync script’s own checks and the “land this side first” condition are carried; two dates and a word count no named source states are dropped.
Changed: Releasing: five skills, as docs/releasing.md counts; the title check is its own workflow step; the release asset’s name follows KNOWLEDGE_RELEASE_NAME; the sync script reports a template a sibling never laid down; the pin is never bumped by hand, the sibling pull request waits for the tag, and the Environment and branch settings are stated, all from docs/releasing.md; the Copilot skill is no longer “not yet released”, and two dates no source carries are dropped.
Changed: Repository validation: nineteen checks, five skill directories, eight sidecar scripts and four sibling scripts, as scripts/validate-repository.sh now counts; checks 18 (prose-check.py) and 19 (knowledge-apply.sh) are described; the files exempt from checks 16 and 16a are all named; the job’s step order follows validate.yml.
Changed: Versioning policy: five skills under one tag, as README.md, CONTRIBUTING.md and docs/releasing.md now say, and publish.yml is listed as the source for the v refusal it carries. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Security policy: the surface table’s fourth row, prose-check.py, and “five skills’ prose” as SECURITY.md now reads, and a report says why a file is exploitable. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Threat model: docs/threat-model.md‘s statements on how the agent is chosen and armed, Dependabot and the templates’ pins, the Environment’s reviewers and --ignore-scripts, the curator’s refused verbs, the attestation job, a gitignored bundle, and the ktl-prose guard are carried; the gate description names the one job that reads events whole; a claim about stable section headings and a word count no source states are dropped.
Changed: Code of conduct: “officially represents”, as CODE_OF_CONDUCT.md says.
Changed: ktl-curator skill: two claims about the KTL Curator plugin that no source carries are replaced by what trust-fields.md says; revision is left out with no version control too, and records the source’s state, as review-session.md says; README.md and CHANGELOG.md are listed for the cadence and the dates.
Changed: ktl-docent skill: evidence-first also covers a concept edited since its confirmation, as SKILL.md says; README.md and CHANGELOG.md are listed for the cadence and the dates.
Changed: ktl-sidecar skill: “missing or broken”, as Step 0 says; the sources now list the librarian workflow, scheduled-task.md and the preflight, which carry the install step and the copies exception, and CHANGELOG.md for the dates; the pen’s undated arrival is no longer dated.
Changed: ktl-librarian skill: the refresh job’s --check-refs step continues on error and reports in the pull request body, as knowledge-librarian.yaml has it; revision is left out with no version control too; three undated changes are no longer dated, and two qualifiers no source states are dropped. Edited since a person confirmed it; the label says so until the curator looks again.
Open question: ktl-librarian skill: the scheduled wrapper knowledge-librarian.sh lets the agent write only .lokf/patch.yaml, and this concept says so, but skills/ktl-librarian/SKILL.md step 6 still says the wrapper permits edits under .lokf/knowledge/, knowledge_bundle/ and .lokf/feedback.md; the skill page lags the script and is outside the bundle.
Changed: Open the knowledge bundle in Obsidian: the sidecar’s portability.md, which carries the reconciler, the Windows junction detail and the version, and CHANGELOG.md, which dates the retired layout, are listed as sources; a reading date no source records is dropped.
Changed: Hosts and doorways: three workflows and three calls to rewrite on another forge, as the sidecar’s portability.md counts; a paragraph on an alternative no source records, a “never inside the host’s build” claim and a markers detail are dropped; portability.md and CHANGELOG.md, which carry the matrix and the retired layout’s dates, are listed as sources.
Changed: Trust label: the section reference and the reason nothing is scored follow trust-fields.md’s wording.
Changed: LOKF: the tools sentence follows the README’s wording since f1f0d51.

2026-10-02

Prose: What WikiSkill says about a knowledge bundle: two sentences split; no fact changed.
Added: ktl-prose skill, the fifth skill, from skills/ktl-prose/SKILL.md; re-created after the bundle was restored from main.
Added: What WikiSkill says about a knowledge bundle, from docs/wikiskill.md, the new page that reads the paper against this design.
Changed: ktl-librarian skill: the run now reads the curator’s verdicts first, caps feedback at ten entries, treats a miss on a covered question as a description defect, and writes the bundle only through knowledge-apply.sh; skills/ktl-librarian/SKILL.md and references/patch.md say so. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Knowledge sources: the sidecar lays down eight scripts, the apply script and its Python half among the fourteen dogfooded copies, and docs/wikiskill.md feeds a new Explanation concept. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Why a registrar role: knowledge-apply.sh is the registrar at the librarian’s desk, as the README’s fifth-role section now says. Edited since a person confirmed it; the label says so until the curator looks again.
Changed: Threat model: the hardening list gains the librarian’s pen, knowledge-apply.sh, as docs/threat-model.md does.
Changed: Three lines of defence: the independent checker is the generated JSON Schema; the SHACL shapes exist upstream and nothing here runs them, as docs/three-lines.md now says.
Changed: ktl-sidecar skill: Step 5 lays down eight scripts and thirteen files, the apply script and its Python half among them.

2026-09-26

Curation: human:noelmcloughlin confirmed 5 concepts (ktl-librarian skill, LOKF specification, versioning policy, security policy, knowledge bundle), corrected 3 (Agent Skills specification, OKF specification, AI covenant) and sent 3 back (Why a registrar role: its description says “rather than by an agent”, which the README does not; Knowledge sources: name the files covered only by glob and add .gitignore; The docent in Microsoft 365 Copilot: Agent Builder’s 50 MB zip limit). All three corrections were the same kind: a paraphrase that drifted from its source’s wording, adding a qualifier the source lacks (“case-exact”, “100 words”, generated as provenance) or dropping one it has (“in that session”, “batched”), which is worth a line in ktl-librarian’s instructions.
The skill playbooks carry the Snyk W011 acknowledgements. #83 added a note to the librarian’s and docent’s SKILL.md saying what contains each finding: the scheduled workflow’s publish job for the librarian reading feedback, and the read-only bundle and scripted feedback entry for the docent’s repository fallback. ktl-librarian-skill and ktl-docent-skill now say so, and the docent playbook gains a typed reference to the threat model it cites.

2026-09-25

The Copilot playbook says what a tenant without custom skills sees. docs/m365.md and the sidecar’s references/m365.md record a test in such a tenant: atk provision refused agent_skills at Microsoft’s service, and Agent Builder sent the skill to Copilot Studio, which needs its own licence. docent-in-m365-copilot now checks for the preview before the routes, adds the Copilot Studio step, and has two troubleshooting lines for those errors in place of the fallback to Agent Builder.

2026-09-24

The Copilot playbook names both toolkit flags. docs/m365.md and the sidecar’s references/m365.md now say atk add skill is hidden behind TEAMSFX_AGENT_SKILLS in the 1.1.17 release and ATK_FRONTIER from the 1.1.18 betas on, and set both. docent-in-m365-copilot follows, in its toolkit route and its troubleshooting list. The toolkit route also changes into the project folder, passes --env dev, and skips atk validate, which rejects agent_skills against the v1.8 schema; two troubleshooting lines cover InvalidProjectError and that rejection. Both variables are exported for the session, since provision packs the skill’s files only while one is set.
The Copilot playbook follows the reworked docs/m365.md. docent-in-m365-copilot now gives the three steps in order, says what a release with no docent zip means, adds the first-answer check and where its gap line goes, and carries the page’s troubleshooting list. The build’s need for bash and zip is new, taken from the builder.
Three open questions closed. docs/install.md now says the skills took the ktl- prefix in 0.22.0, so the distribution concept states it again. references/scheduled-task.md lists the refresh job’s install, Node and write-scope steps, closing the librarian playbook’s question. Obsidian’s settings help confirms that Excluded files hides a folder from search, graph view and unlinked mentions, so the Obsidian playbook cites it, and the sidecar’s portability.md and the docent’s obsidian.md no longer understate it.
Threat model follows its source. docs/threat-model.md and the concept now say the release workflow’s attach job checks each zip, not one.
The bundle follows the Microsoft 365 Copilot docent. A new playbook, docent-in-m365-copilot, derives from docs/m365.md and the sidecar’s references/m365.md: how to get or build the docent’s Copilot skill, add it to a declarative agent and keep it current, and why only read-only roles go there. The sidecar, releasing and repository-validation playbooks now describe the .lokf/m365/ folder, the release workflow’s second zip and check 17. The three-lines, domain-schemas, hosts and distribution explanations carry the auditor’s planned Copilot form, the reserved Microsoft 365 identity shape, the new portability row and the fourth install route.
Sibling sync described. scripts/sync-sidecar.sh brings a sibling repository’s template copies and skills pin up to one release and leaves the diff for review; the releasing and contributing playbooks and the source map now say so. CONTRIBUTING.md no longer asks for a hand bump of the template’s pin or calls the librarian workflow an exception to check 11, so the contributing playbook’s open question closes and the concept follows its source again.
Curation: at human:noelmcloughlin’s instruction, his four confirmations of 2026-09-09 (knowledge bundle, Agent Skills, OKF and LOKF specifications) are re-dated from 18:30-18:38Z to 16:35:55Z, the commit that recorded them, so they satisfy conventions rule 11. What was confirmed is unchanged.
Glossary terms described. The four glossary terms gain a description and their index bullets a summary. The knowledge bundle’s is the first sentence of its human-authored definition, word for word, so its generated record stays true.
Conventions rule 11 reflected. The repository-validation playbook counts eleven rules and the new time check among check 12’s cases, and the librarian playbook says every at now comes from date -u.
Audit: impossible timestamps corrected. Sixteen generated.at values and eight process:ktl-librarian verification times were later than the commit that recorded them (local time written as UTC on 2026-09-24, round placeholder times on 2026-09-09 and 2026-09-14). Each now carries that commit’s time, the latest moment git attests. No human: event was touched.
Audit: three body links typed. lokf-specification now references the toolkit, the librarian playbook references the curator playbook, and the sidecar playbook hasPart the Obsidian playbook, the inverse of that playbook’s existing isPartOf.
The bundle follows 0.24.0 and 0.25.0. The sidecar, librarian, releasing and repository-validation playbooks and the threat model now describe the third sidecar workflow, knowledge-release.yaml, which publish.yml dispatches to attach the bundle zip, the scheduled librarian’s credential slot and --check-refs validation, and the gate detail’s move into the sidecar’s new gate.md. The source map lists the release workflow among the dogfooded copies; the rest of the diff was rewording that changed no concept.
Two concepts cut back to their sources. why-a-registrar-role drops the four-skill contrast and “or agent”, neither of which the README says, closing the curator’s open question. policies/versioning.md drops an uncited history of the old rule and the reasoning behind one shared tag, keeping the README’s conclusion, and dates semantic release to 0.15.0 (2026-09-12) from the changelog instead of 2026-09-14.
Reader feedback consumed. Two docent misses from 2026-09-24. The governance question becomes explanation/domain-schemas.md, derived from the curator’s and librarian’s domain-schema references, including the AI Risk Ontology pointer no concept carried. The DevSecOps question becomes explanation/intended-uses.md, a draft placeholder, because no source names KTL’s intended uses or non-goals.
Curation: human:noelmcloughlin confirmed 2 concepts (LOKF specification, AI covenant), corrected 1 (Knowledge bundle’s definition), and sent 2 back (ktl-librarian skill, Why a registrar role).
Audit: playbooks/index.md had no bullet for the Obsidian playbook, and the sidecar playbook linked hosts-and-doorways in its body with no typed relation; both fixed. Every other check passed - schema, relation targets, conventions, the projected graph, and the manual lint pass.
README-sourced explanations re-derived. why-four-roles, why-a-registrar-role, why-a-distribution-repository and glossary/lokf had kept prose no current source carries (an origin story, an adversarial review analogy, a single-plugin framing), and missed the run order, the plugins’ levels of checking and the Claude Code plugin route. Each now cites every file it rests on; the registrar concept’s open question closes by dropping the uncited rename detail.
The contract and release playbooks follow #69 and #70. Check 11 now covers the librarian workflow apart from its skills pin, which the release commit moves in the template only; checks 3c, 16 and 16a were missing, and the count is sixteen, not fifteen.
Skill and policy playbooks follow their sources. The sidecar has seven steps and two scripts that land regardless; the curator’s report counts unsigned confirmations from git history; the librarian consumes feedback first; the threat model gains the registrar’s pull-request-metadata guard and the publish job’s exact pathspecs; contributing, security and the covenant gain the #60 rules. Two curator open questions close by citing the files the claims came from.
Source map: rows for .claude-plugin/ and .lokf/curators/, and the dogfooded-workflows row no longer says the librarian workflow omits its install step.

2026-09-23

The docent records a reader’s gap without reading the file it goes in. .lokf/feedback.md is kept newest first, so adding an entry meant reading it, editing it and writing it back, which put other readers’ free text into the docent’s session, held off by prose alone (Snyk W011). knowledge-feedback.sh, a new sidecar script, does the insertion instead: ktl-docent passes its own entry as an argument and never opens the file. The guard is the call rather than a rule an agent has to keep, the same reason the conventions the librarian kept breaking became a script. The preflight reports a sidecar that predates it; check 12a exercises it. Nothing changed for ktl-librarian, which still reads them.
A scaffolded host installed a librarian that was not in the tag. The librarian template’s TRUST_LADDER_SKILLS_REF read v0.21.0, where the skills are named lokf-*; its install step clones that tag and copies skills/ktl-librarian, a path that arrived with the rename in v0.22.0. Every scheduled run failed there, armed or not. The pin moves to v0.22.0, and check 15 now reads the pinned tag for the path the step copies - a version check alone reported both halves as current.

2026-09-19

The relation audit moved onto the toolkit’s own flag. just lokf-check-refs ran a hand-written SPARQL query naming ten predicates, which had drifted against the schema and reported an external source: or definedBy: URL - correct usage for both slots - as a dangling target. Both justfiles and both registrar workflows now call lokf validate --check-refs, so the slot list comes from the schema and the gate validates once instead of twice. The two plugin repositories had already made this change in their own copies; the template was the last holdout, which is what their preflight copies line was reporting. references/lokf-toolkit.md no longer says the recipe runs a SPARQL query.
Repository renamed to knowledge-trust-ladder, from lokf-agent-skills, because upstream LOKF now ships its own bundled skills and a third-party repository named after the format read as their official home. The bundle’s base_iri moves with it, re-iding all 30 concepts and every typed-relation target in one mechanical pass; no verified event changed, so nothing a person confirmed was touched. Concepts naming the repository in prose, its GitHub URL, or the card asset were rewritten by the same pass. The four skill names stay as they are: they are installed paths in every host. Entries below this day keep the name the project had when they were written, as CHANGELOG.md does: a log that renames its own past is no longer a record of what happened.
The README states the architectural claim, and a logo joins the assets. A short section near the top quotes OKF’s fourth goal - fields that make a corpus trustable “without prescribing any runtime” - and says this is such a runtime, then says the domain schemas OKF puts out of scope are in scope here by construction, pointing at the curator’s domain-schemas page for the flag and the recipe. New .assets/knowledge-trust-ladder-logo.svg, three rungs in the trust ladder’s own grey, amber and blue, shared byte-for-byte with both plugin repositories; lokf-trust-ladder.svg renamed trust-ladder.svg, since the old name now reads as the project’s rather than the diagram’s. The source-map row for .assets/*.svg covers both.
The README became a front door. Much the same length - 1,792 words to 1,864 - and a different shape: the opening sentence carried four link-bearing clauses before saying what the project does, and the install commands - two code blocks, a prerequisites paragraph and the pinning rule - sat in the middle of the narrative. Commands moved to a new docs/install.md, leaving one npx skills add line; the closing sections collapsed into one Read on table. What the space bought: the runtime claim, the context layer the bundle is, and the schema-first order of work. The source-map row now names the new page.
The logo carries knowledge and trust, not just a ladder. The first mark was three rungs alone: it drew ladder and left the other two words to the filename. The ladder now rises out of an open book (knowledge, in the agent-green of the diagrams’ key) and its top rung is the check mark in the curator’s blue - so the mark reads draft, checked, vouched-for from bottom to top. Rails lightened to #A8A69D so the grey draft rung is not read as structure. Every diagram now carries the mark and the wordmark in its top-left corner, from one shared definition.
The three-lines diagram was keyed against its own legend. Its bands coloured the first line green and the second blue, while the legend the other diagrams share reads green for agent work, amber for deterministic checks, blue for a named person - so the registrar, a program, sat in the person’s colour and the curator, a person, sat in the agent’s. The bands are now neutral panels and each card takes the colour of who acts. It also gained the title and legend the rest of the family has.
Text that collided with its own box. A headless-Chrome pass measured every <text> against the smallest rect containing it: six labels across four diagrams overflowed or came within 2px, including one in the siblings’ two-vaults copy that was hidden behind a card it was painted before. Fixed by shortening the labels, widening one card, and correcting the paint order.
Assets restacked. Every diagram’s canvas was pure white while its cards were tinted, so the largest area was the brightest and the cards read as sunken rather than raised. The canvas is now #E4E1D7 across all seven assets, and the faint ink #888780 is now #6E6D66, which also lifts the 10px grey labels past 4.5:1 - a contrast failure that predated the change. The same restack was applied to both plugin repositories’ assets.

2026-09-18

Curation: unauthenticated session (no gh login, so only send-back/retire/later were available - no human: events written) sent back 5 concepts flagged as edited since a prior human confirmation: Knowledge bundle, ktl-librarian skill, Why a registrar role, Security policy, Versioning policy. Four of the five shared one root cause - the concept cited a single resource but the disputed detail actually came from a different file the librarian had read (a workflow wrapper, scripts/validate-repository.sh, or a sibling doc) - worth ktl-librarian tightening resource/sources citation discipline rather than re-deriving each concept the same way again.
Pin check given the publish window: check 15 required the pin to equal the newest released heading, which no publish can satisfy - the heading is promoted on merge and the tag created later, so the pin can only name the release below it. It now accepts either of the two newest, failing the 0.19.3 publish is what showed this.
Retitle now re-runs the check: the squash-title check asks the author to retitle, but semantic-release.yml listened only for the default pull-request types, so a title edit fired no run and the check could not be cleared. The trigger names edited. playbooks/releasing.md follows.
Tracked resources only: Rule 5’s field notes now say a local resource must be a path git tracks. A gitignored install path under .agents/ resolves only on the machine holding it, and failed the conventions gate in a clean CI checkout.
Registrar template least-privilege: the workflow had no top-level permissions: {} and left the checkout credential on disk in both read-only jobs, while knowledge-librarian.yaml and two host copies already set the first. Both added to the template and this repository’s copy, so the gate’s jobs opt into scope rather than inherit it.
Skills pin held to the release: the librarian template’s LOKF_SKILLS_REF sat at v0.9.0 while this repository released v0.19.2, so every host scaffolded from it installed a librarian ten minor versions old. Bumped, and check 15 now holds the pin to CHANGELOG.md’s top released heading so it cannot fall behind unnoticed. playbooks/repository-validation.md follows.

2026-09-17

Wrapper restore made unconditional: knowledge-librarian.sh restored .git/config and .git/hooks/ only after a clean agent return, so a failing agent or a cancelled job left a poisoned config in the checkout. An EXIT trap now restores on every way out; the layout tests prove the failing and cancelled cases. playbooks/ktl-librarian-skill.md follows.
Schema fallback pinned: the no-Python fallback URL in both skills’ Sources notes, the sidecar README and references/lokf-toolkit.md now points at the v0.8.0 tag matching the lokf floor, not main, which has already diverged from it. The librarian’s tooling-version step moves the pin with the floor.
Duplicate-release fix (seventeenth pass): two merges to main between publishes computed the same next version twice, promoting ## [Unreleased] onto a second ## [0.19.0] heading above the first and orphaning the second push’s entries. changelog-release.mjs promote now folds into the top released section, by subsection, when its version carries no tag yet; the plan job also runs changelog-release.mjs check directly, since semantic-release skips its own plugins’ hooks on a pull request and the contract’s claim that an empty [Unreleased] fails a pull request was not true. Check 14 proves both, and repairs the changelog this bug had already written twice. playbooks/releasing.md follows.
Third audit (sixteenth pass): the sidecar and contract playbooks and the threat model follow the audit’s fixes - the sidecar lays down the conventions script’s Python half, both gates read merges against every parent, human generated records and only the frontmatter, and rule 10 keeps the fields the gates read to plain spellings.
Conventions script split (fifteenth pass): knowledge-conventions.py, run through uv run, takes over the five frontmatter-shape rules (quoted at, verified shape, open questions, duplicate id, closed frontmatter) with a real YAML parse; the contract and sources playbooks follow.
Second audit (fourteenth pass): the three lines explanation follows its pruned page and the shortened critics page; the trust label and the curator and librarian playbooks now say revision is proposed for lokf 0.9.0 and unreleased; the contract playbook adds the id the gate used to skip, and the gate’s reading of a confirmation whole - a re-dated event or a flow-style one counts, a renamed concept does not.
Audit follow-up (thirteenth pass): the sidecar, curator and docent playbooks, the contract and the threat model follow the forge-free gate’s SSH and subkey support, the linked-bundle and commit.gpgsign fixes, and the new prerequisites.md that puts each preflight line in plain words for a person who cannot act on it.
Portability work (twelfth pass): the contract, the four skill playbooks, the threat model, the three lines page and the hosts explanation follow their sources - a preflight every skill runs first, a forge-free provenance gate, .lokf/.gitattributes, three new conventions (one file per id, lowercase paths, a readable frontmatter block), each skill’s compatibility field, and three identity routes for the curator.
Steady-state refresh (eleventh pass): twenty concepts re-checked against their sources. The librarian, curator and docent skills and the trust label now describe revision, and the librarian skill no longer counts its classes.
Releasing corrected: semantic-release runs with --dry-run and @semantic-release/exec calls the changelog script’s check and notes; publish.yml compares against the top released heading, not Unreleased.
Provenance fixed: repository-validation and three-lines-of-defence carried a generated.at later than the change that wrote them, and now carry the time that change merged.
Repository validation follows check 11: the conventions script’s sixth rule (a commit-shaped revision must name a commit holding the resource) is proved in a throwaway repository, failing and passing.
Three lines of defence follows the page: the state a confirmation was checked against is recorded in revision (lokf 0.9.0+), the gate resolves a commit hash against the tree, and OKF’s adoption of the field (knowledge-catalog#437) is the remaining upstream item.

2026-09-16

Steady-state refresh (tenth pass), against the uncommitted working tree: fifteen concepts re-checked against their sources and found to match (verified refreshed), the four re-generated today carry today’s generated.at, and the source map’s excluded-files and images rows follow docs/examples/ and the new diagram. explanation/three-lines-of-defence.md follows the page’s renamed closing heading, “What remains to do, and who does it”. explanation/index.md’s hosts-and-doorways bullet no longer calls the layout an open question. .lokf/feedback.md has no entries; lokf on PyPI is still 0.7.0.
explanation/three-lines-of-defence.md refreshed against a rewritten docs/three-lines.md and its new docs/three-lines-critics.md: the page names The Institute of Internal Auditors in full, cites the 2026 Statement of Position, places each role in a diagram and ends with what remains open and whose it is; the critics page quotes the model’s critics and the research behind two partial critiques, marked preliminary, with a bundle’s answer to each and every remaining gap labelled by kind. index.md, playbooks/knowledge-sources.md and docs/repository-layout.md follow, and the source map’s excluded-files row follows EXAMPLES.md to docs/examples/docent.md.
playbooks/ktl-curator-skill.md, playbooks/ktl-docent-skill.md and playbooks/repository-validation.md follow their skills: the conventions script’s fifth rule (a local resource must exist), the curation policy’s Independent re-check: and Evidence first: switches, and that the docent reads the second from the policy alone.
playbooks/ktl-sidecar-skill.md no longer says “the solo maintainer’s signing setup”, matching references/automation.md and the registrar workflow’s comments, which describe the case - the confirming person opens the pull request - rather than the person.

2026-09-14

Maintainer polish after the ninth pass: the README’s tree moved to docs/repository-layout.md and says three repositories, the count the ninth pass flagged, and index.md, policies/index.md, policies/threat-model.md, playbooks/releasing.md and playbooks/knowledge-sources.md now count the same three. explanation/three-lines-of-defence.md names the README’s renamed section; prose counts of LOKF’s classes and relations left the README, docs/ and the skills’ references, with Rule 3 and trust-fields.md still enumerating them under check 7.
Steady-state refresh (ninth pass), against commit b94a299. policies/security.md rewritten to match the slimmed SECURITY.md (reporting, supported versions, a surface table), and a new concept, policies/threat-model.md, added for the shared design page docs/threat-model.md moved it to - hardening, the human: attribution gate, prompt-injection guards - keeping that page’s section headings stable since the other three repositories deep-link them. playbooks/knowledge-sources.md and playbooks/repository-validation.md updated to match; index.md and policies/index.md gained the new concept’s bullet. Flagged, not fixed (outside this skill’s .lokf/-only scope): README.md’s docs-layout listing still says “four repositories release” where the same commit corrected docs/releasing.md itself to “three”. See playbooks/knowledge-sources.md’s note for detail.
Steady-state refresh (eighth pass), against the two commits landed since the seventh pass. Twelve concepts re-checked against their current sources and found unchanged (content matches; verified refreshed only): explanation/why-four-roles.md, why-a-registrar-role.md, why-a-distribution-repository.md, glossary/lokf.md, glossary/okf.md, glossary/trust-label.md, glossary/knowledge-bundle.md, explanation/hosts-and-doorways.md, playbooks/open-bundle-in-obsidian.md, playbooks/ktl-sidecar-skill.md, playbooks/ktl-docent-skill.md, playbooks/ktl-librarian-skill.md. Orphan fixed: playbooks/knowledge-sources.md’s dogfooded-workflows row had no entry for .lokf/scripts/knowledge-conventions.sh (added earlier today), now listed alongside knowledge-librarian.sh as byte-identical to its template. .lokf/feedback.md had no entries.
playbooks/repository-validation.md: check 11 added (template copies identical, knowledge-conventions.sh exercised), the sibling count corrected to the two plugin repositories, and actionlint’s explicit paths noted.
Steady-state refresh against this session’s uncommitted repository changes: CONTRIBUTING.md was trimmed to a checklist, its release process and commit-signing walkthrough moved to two new pages, docs/releasing.md and docs/signing-commits.md; a third new page, docs/three-lines.md, maps the four-skill-plus-registrar cast onto the Three Lines Model; validate.yml gained an action-pinning check; scripts/validate-repository.sh gained checks 9/9a (the paths sibling repositories deep-link into, and completeness when they’re cloned alongside) and 10 (a 1000-word budget on CONTRIBUTING.md); README.md gained a section pointing at docs/three-lines.md. playbooks/contributing.md rewritten to match the trimmed file, its Code-of-conduct/AI/Releasing sections, and the new docs/ links. policies/versioning.md re-sourced from CONTRIBUTING.md to docs/releasing.md and rewritten: its qualitative patch/minor/major definition (“does not materially change expected behaviour”) no longer exists anywhere in the repository, superseded by the Conventional-Commits table docs/releasing.md/playbooks/releasing.md already state - the 2026-09-10 human confirmation left standing, per rule, since the concept was not human-authored. playbooks/releasing.md’s stale CONTRIBUTING.md citation repointed to docs/releasing.md. playbooks/repository-validation.md gained the new action-pinning check and, closing a pre-existing gap this pass found (checks 6-8 were never described either), the full ten-check enumeration. Added explanation/three-lines-of-defence.md for the new page, linked from why-four-roles.md and why-a-registrar-role.md via about. Orphan sweep, also fixing two pre-existing gaps unrelated to this diff: explanation/index.md was missing a bullet for why-a-registrar-role.md (present in the root index.md and in log.md since 2026-09-10, never added here), and its first bullet’s title had drifted from the concept’s own (“Why four roles” vs “Why four skill roles”); both fixed. index.md gained the new explanation’s bullet. playbooks/knowledge-sources.md: the CONTRIBUTING.md row updated, two new rows added for the docs/ pages, and a run note added. lokf on PyPI is still 0.7.0 (checked via PyPI’s JSON API; uv pip index is still not a subcommand here), matching the sidecar’s floor - no bump. .lokf/feedback.md has no entries. Not re-checked this pass: concepts whose resource is untouched by this diff.
Two reference concepts corrected on a librarian pass over the whole repository. references/linkml.md still said the vocabulary it describes has fourteen classes - the same stale count fixed in Rule 3 earlier today - and named no way to validate an extension; it now says fifteen and points at lokf validate --schema and the recipe. references/lokf-toolkit.md described lokf validate without the --schema flag the recipe rests on. lokf on PyPI is 0.7.0, matching the floor; .lokf/feedback.md is empty.
Steady-state refresh, against branch docs/domain-schema-extension (uncommitted): reframed the domain-schema-extension material added this session - skills/ktl-librarian/references/domain-schema.md now states the recipe (pin the core schema, write a domain schema, name the class exactly, wire --schema in) without the validator-internals digression or the two citations of a named consumer project the first draft carried. playbooks/ktl-librarian-skill.md gained a paragraph on it; re-verified, no change: playbooks/ktl-curator-skill.md. Correctness bug fixed: Rule 3 in skills/ktl-librarian/SKILL.md listed fourteen classes, omitting Role, against README.md/docs/for-the-curious.md’s fifteen - the open question this concept carried since 2026-09-13. Confirmed against the raw schema (fifteen concept classes, Role among them), fixed Rule 3, closed the question. playbooks/knowledge-sources.md: run note added.
Curation-policy table: how a row binds to a class is now stated. ktl-curator/references/review-session.md defined the template but not the match. The KTL Curator plugin, which implements this document, read it literally, so the template’s own prose rows (“Glossary terms”, “people”) bound nothing. The skill now states the tolerant match - spaces and plural form ignored - and playbooks/ktl-curator-skill.md records it.
The curator skill now reads a host’s domain schema too, closing the asymmetry the ktl-curator plugin’s bundle raised as an open question: the librarian’s Rule 3 widened its class list from .lokf/justfile’s --schema, while the curator still counted vocabulary fit against the fifteen built-ins, so a host with a domain schema saw every domain class reported as a misfit for good. references/trust-fields.md states the widening (read the file; no toolkit, as everywhere else in Step 1), the label becomes “doesn’t fit the known vocabulary” as the plugin already words it, and SKILL.md, references/domain-schemas.md and the librarian’s references/domain-schema.md (step 5: the skills need no telling, the plugins do) follow. playbooks/ktl-curator-skill.md records it.

2026-09-13

Steady-state refresh (librarian pass, no feedback pending), after a ktl-sidecar health check that found nothing to repair here. Withdrawn: the third pass’s additions that described commit 312190e, namely the sidecar’s refresh mode and automation.md’s GPG-or-SSH and required-checks-off text (playbooks/ktl-sidecar-skill.md, and the index.md bullet), the Drift audit heading and Role in Rule 3 (playbooks/ktl-librarian-skill.md), and SECURITY.md’s “documented once, here” paragraph (policies/security.md, which now cites the librarian-PR wrinkle to automation.md, where this tree states it). That commit is on the unmerged branch update, not in the checked-out skills/. playbooks/ktl-librarian-skill.md is status: draft again with an open question: README.md counts fifteen LOKF classes, Rule 3 lists fourteen. Re-verified, no change: explanation/why-four-roles.md, explanation/why-a-distribution-repository.md, glossary/lokf.md. playbooks/knowledge-sources.md: run note added.
playbooks/knowledge-sources.md: the docs/ row now covers the new docs/obsidian.md - the README’s Obsidian section, moved out as a page of its own with the two-vaults picture - and names the two files it must agree with. explanation/why-a-registrar-role.md: the “extensions of the skills’ output” sentence is now cited to docs/obsidian.md, where it lives. No other concept changed: the README polish moved emphasis and order, not facts.
Steady-state refresh (librarian pass, no feedback pending), against the commits since 3399203 and this session’s uncommitted README.md (three more .assets/ images, one sentence on the plugin READMEs). Content corrected or extended: policies/security.md (main is protected but not by a merge gate - required status checks deliberately off; the template’s validate/provenance jobs never run on the librarian’s own PR; a consumer’s SECURITY.md links here rather than restating; the closing “branch protection” claim replaced by review-before-merge, as SECURITY.md now says), policies/versioning.md (tags carry the v, changelog headings never do), playbooks/contributing.md (required checks off; sign your commits), playbooks/releasing.md (the v the cross-check strips), playbooks/ktl-librarian-skill.md (the new Drift audit heading, and Role restoring the fifteenth class), playbooks/ktl-sidecar-skill.md (the new refresh mode; automation.md‘s GPG-or-SSH signing, its advice against required status checks, and its “Keeping a host’s copies in step” section; the Step 6 handoff line on a host’s SECURITY.md) and explanation/why-a-registrar-role.md (the plugins are extensions of the skills’ output, not plugins for a vault in general). Re-verified, no change: explanation/why-four-roles.md, explanation/why-a-distribution-repository.md, glossary/lokf.md, playbooks/open-bundle-in-obsidian.md. playbooks/knowledge-sources.md: the dogfooded-workflows row no longer claims byte-identity - this repository’s knowledge-librarian.yaml deliberately omits the template’s “Install the pinned ktl-librarian skill” step, since it publishes the skills it uses - and a run note added. Drift audit: that omission plus a two-line comment is the only difference; knowledge-registrar.yaml, the wrapper and the justfile match their templates. lokf on PyPI is still 0.7.0, matching the floor.
Steady-state refresh (librarian pass, no feedback pending): re-verified playbooks/ktl-sidecar-skill.md, explanation/hosts-and-doorways.md, playbooks/open-bundle-in-obsidian.md, explanation/why-four-roles.md and explanation/why-a-registrar-role.md against skills/ktl-sidecar/SKILL.md and the compacted README.md from this session’s “For the curious” move and prose-compaction pass. All five already matched - no content changed, only each concept’s own process:ktl-librarian verified event, the two hosts-and-doorways/open-bundle-in-obsidian concepts gaining their first one. playbooks/knowledge-sources.md gained its own docs/for-the-curious.md row (added this session) and its generated/verified refreshed to match. glossary/knowledge-bundle.md was not re-checked - templates/README.md’s wording tweak this session didn’t touch what the term means.
The visible layout is retired; one layout, the doorway by default. Hand-edited in session alongside the ktl-sidecar change, not a librarian run - the librarian’s next pass re-verifies. playbooks/open-bundle-in-obsidian.md and explanation/hosts-and-doorways.md rewritten (generated refreshed, their process verified events dropped, since the text they checked is gone): the bundle is .lokf/knowledge on every host, knowledge_bundle a link beside it; an Obsidian user keeps two vaults - the workshop they have, and the bundle opened as its own vault through the link - and the bundle is never laid down as a real folder inside a vault, because a day of use in the maintainer’s own vault showed the exhibition leaking into the workshop’s link suggestions, quick switcher, graph and search (Obsidian’s Excluded files only makes that less noticeable). The playbook keeps the 1.13.7 reconciler findings and now names the cost of linking a repository’s bundle into a vault; the explanation records the 2026-09-12 attempt, its cost, and what the reversal keeps (dual pathspecs, plugin auto-detection, junction guidance). glossary/knowledge-bundle.md (generated refreshed; its human confirmation of 2026-09-09 stands, so the curator will show it as edited since) no longer claims “three pictures”: a bundle is a folder of notes opened as its own vault, or, at a cost, a folder inside a vault or a whole vault. playbooks/ktl-sidecar-skill.md (generated refreshed) drops the Step 0 layout decision; playbooks/ktl-librarian-skill.md (generated refreshed) restates the second-name rule without the visible layout. index.md’s two TOC lines follow.
Later the same day. The names of the maintainer’s private projects were removed from this bundle’s notes and from this log: the family this bundle describes is the four skills and the two Obsidian plugins, and no host outside it is a reference. README.md’s Obsidian section was cut back to the value of the pairing - Obsidian optional, the exhibition as the workshop’s output - and the mechanics it carried (what the file reconciler does with a link, the second vault, the linked-in arrangement and its cost) moved to ktl-sidecar/references/portability.md, so that playbooks/open-bundle-in-obsidian.md derives from a source that states them. Every passing mention of Obsidian in the skills and templates now says it is optional. README.md’s “For the curious” moved to docs/for-the-curious.md, mirroring the plugin READMEs (playbooks/knowledge-sources.md lists it), and duplicated prose was compacted in the README, the sidecar skill’s Steps 2 and 6, portability.md and this bundle’s open-bundle-in-obsidian.md, whose history section now defers to explanation/hosts-and-doorways.md.

2026-09-12

Steady-state refresh against the security-hardening pass (uncommitted at the time of this run: 19 files touched, none of them .lokf/). policies/security.md rewritten from the current SECURITY.md (generated/verified refreshed): harden-runner now runs on every workflow that installs packages or runs third-party/agent code, not only validate.yml/publish.yml; a new ## human: attribution is a claim, not a credential section records knowledge-registrar.yaml’s provenance gate, ktl-curator’s gh api user-only identity rule, its refusal to run unattended, and the optional attestation job; the blast-radius paragraph now correctly attributes boundary enforcement to the privileged publish job (re-deriving the allow-list from the patch itself on a clean checkout) rather than the wrapper script, and records the wrapper’s new .git/config/.git/hooks/ snapshot-and-restore and its checks moving into a main() called last. playbooks/ktl-docent-skill.md and playbooks/ktl-librarian-skill.md (generated/verified refreshed on both) gained the docent’s new never-repeat-a-secret guard and the librarian’s own scheduled-run-only caveat on the tooling-version check, respectively. playbooks/releasing.md (generated/verified refreshed) gained the env:-var injection-hardening note on publish.yml’s two version checks, and a fact from a previously unlisted source, .releaserc.json: the release-rule mapping is the Angular preset’s defaults plus one addition, security: -> patch. knowledge-sources.md’s source map gained a row for semantic-release.yml / changelog-release.mjs / .releaserc.json, none of which had one. Re-verified without body changes (verified refreshed only, confirmed still matching their current, unchanged-by-this-pass resources): playbooks/ktl-sidecar-skill.md, playbooks/open-bundle-in-obsidian.md, playbooks/contributing.md, policies/versioning.md, explanation/hosts-and-doorways.md (its first verified event), explanation/why-a-distribution-repository.md, explanation/why-a-registrar-role.md, explanation/why-four-roles.md, glossary/lokf.md, glossary/okf.md, glossary/knowledge-bundle.md. .lokf/feedback.md held no real entries (its unfilled template only), so nothing was consumed. lokf on PyPI is still 0.7.0, matching this sidecar’s floor - no bump needed. Not re-checked this run: the policy concepts whose resources this pass never touched (policies/ai-covenant.md, policies/code-of-conduct.md), the two concepts sourced from skills/ktl-curator/ (playbooks/ktl-curator-skill.md, glossary/trust-label.md), playbooks/repository-validation.md, and the seven external Reference concepts - none of their resources appear among this pass’s changed files, so they were left as last verified.
Semantic-release, hardened - version and changelog only (maintainer decision): playbooks/releasing.md rewritten (generated/verified refreshed). semantic-release.yml’s release job, behind the release GitHub Environment, computes the next version from Conventional Commits on every push to main, but only ever runs semantic-release --dry-run - gh skill publish stays this repository’s one tag creator, so nothing here writes, commits, tags, or publishes on the tool’s own initiative. A plain shell step reads the dry run’s computed version, promotes CHANGELOG.md’s ## [Unreleased] section itself, and commits directly. publish.yml gains a cross-check: the maintainer’s typed version must match what got promoted, or the run fails before touching the registry.
KTL Registrar named (maintainer decision): explanation/why-a-registrar-role.md (generated refreshed), playbooks/ktl-librarian-skill.md, playbooks/open-bundle-in-obsidian.md, and playbooks/knowledge-sources.md name the plugin and its repository obsidian-ktl-registrar. The diataxis.md actor the librarian leaves alone reads ktl-registrar/<version>. glossary/knowledge-bundle.md (generated refreshed) now records the README’s three pictures for the bundle - the librarian’s catalogue, the museum’s exhibition as the hall visitors are shown into, and the workshop of sources both stand against - after the README’s opening section was made to say that its catalogue and its exhibition are the same folder.
Corrected playbooks/open-bundle-in-obsidian.md (rewritten, generated refreshed): it now states the one supported route - open knowledge_bundle itself as a vault - and records why the repository-root route cannot work, citing Obsidian’s help on symbolic links (“ignores a symlink … from one folder in the vault to another folder in the same vault”) and its dot-folder rule. The two companion plugin READMEs had claimed otherwise; both were corrected the same day. Added the supported reverse direction (linking a repository’s bundle into a personal vault). Source: skills/ktl-sidecar/SKILL.md Step 2, whose symlink paragraph was rewritten to match.
Re-verified playbooks/ktl-sidecar-skill.md against the updated router (Step 5 heading and Step 6 wording changed, step list unchanged) - verified timestamp refreshed only. Not a full steady-state sweep.
Second pass, same day: playbooks/open-bundle-in-obsidian.md rewritten again after the maintainer asked for the architecture, not the Obsidian community’s habits, to lead. The symlink rule is now stated from Obsidian 1.13.7’s own file reconciler (reconcileSymbolicLinkCreation, read from the installed application bundle): a link is skipped when its resolved path equals, contains, or lies inside a folder already being watched, the vault root included - so the doorway opened as a vault works exactly as Step 2 intends (the maintainer’s own vault does this daily), and a host-root vault merely does not list it, which for a notes-vault host is the property that keeps the sidecar safe inside the vault. Added the Windows junction (mklink /J, no elevated rights) and the OneDrive note (syncs neither symlinks nor junctions; no rule against dot-folders) to the playbook, skills/ktl-sidecar/SKILL.md Step 2, and references/portability.md. Added explanation/hosts-and-doorways.md (status: draft, with open questions for the maintainer): the two-name rule (.lokf/knowledge for tools, knowledge_bundle for people, one of them a link) and the proposal to let a vault or shared-drive host make the visible name the real folder.
Third pass, same day - the proposal adopted: the maintainer asked for the recommended fixes to be implemented. skills/ktl-sidecar/SKILL.md Step 0 now decides the layout by host (code repository: hidden real folder, visible doorway link; notes vault or shared folder: visible real knowledge_bundle/, .lokf/knowledge as the tools’ link), with Steps 1, 2, 3, 5 and 6 marked where the visible layout differs; the wrapper script and both workflow templates name the bundle under both names (a git pathspec never traverses a symlink; git add refuses an empty pathspec, hence a guard), the registrar triggers on knowledge_bundle/**, and the justfile gained just lokf-link. skills/ktl-librarian/SKILL.md gained the layout note and a rule to leave KTL Registrar’s Obsidian affordances ( blocks, diataxis.md) alone. Updated explanation/hosts-and-doorways.md (proposal → adopted, open questions resolved), playbooks/open-bundle-in-obsidian.md (visible-layout section), playbooks/ktl-sidecar-skill.md (Step 0 decision) and playbooks/ktl-librarian-skill.md (also corrected “14-class” to the schema’s 15, matching README.md’s earlier fix). templates/README.md and this sidecar’s README.md explain the layout in one sentence.
Fourth pass - the vault-in-a-subfolder host, and layout tests. ktl-sidecar Step 0 gained the case of a vault one level below the repository root (the maintainer’s own notes vault, vault/ here): the real folder goes inside the vault, vault/knowledge_bundle/, the link is .lokf/knowledge -> ../vault/knowledge_bundle, the justfile’s new visible variable names that path for just lokf-link, and Step 5’s three files name it in place of knowledge_bundle. explanation/hosts-and-doorways.md gained the table row and the note that scripts/test-sidecar-layouts.sh now pins both layouts - the wrapper’s boundary check, both workflows’ pathspecs, and the recipe - from the repository-contract check; playbooks/open-bundle-in-obsidian.md and playbooks/ktl-sidecar-skill.md say the same in a clause. The README’s Obsidian section now states the two-name rule outright and names the vault the workshop and the bundle the exhibition. explanation/why-a-registrar-role.md no longer files KTL Curator under the registrar or calls the plugin a curator: KTL Registrar is the registrar in the editor, KTL Curator is the person’s assistant, and the curator is always a person.
Vocabulary: glossary/knowledge-bundle.md now records the museum synonyms the READMEs use - the bundle is the exhibition, a concept an exhibit, the sources or notes the workshop - so the three words are defined once, at the term of record.

2026-09-11

knowledge_bundle symlink added to ktl-sidecar Step 2: a third root-level pointer, alongside llms.txt and the README aside - ln -s .lokf/knowledge knowledge_bundle, a visible entry point for humans and their tools, chiefly Obsidian’s “Open folder as vault,” which like most OS folder pickers hides dot-directories by default. templates/gitignore now excludes .obsidian/, which Obsidian writes through the link into the real .lokf/knowledge/.obsidian/ when used as a vault. Added playbooks/open-bundle-in-obsidian.md and refreshed playbooks/ktl-sidecar-skill.md’s Overview (Step 2 now three additions, not two). Re-verified glossary/knowledge-bundle.md and playbooks/repository-validation.md against their now-touched resources - no body drift, verified timestamps only. This repository’s own bundle gained the symlink too, plus matching excludes in .markdownlint-cli2.jsonc, lychee.toml, and the codespell step so the aliased files aren’t linted/checked twice - those two config files stay outside the bundle, per the existing convention noted in playbooks/knowledge-sources.md. Targeted pass, not a full steady-state sweep - concepts untouched by this change were not re-checked this run.

2026-09-10

knowledge-validate.yaml renamed to knowledge-registrar.yaml: at a maintainer’s request, so the CI gate that keeps bundle records well-formed (never judging their truth) is named for the registrar role it actually performs, matching explanation/why-a-registrar-role.md. Renamed the dogfooded workflow and its byte-identical template counterpart under skills/ktl-sidecar/templates/github/, and updated every cross-reference across the four skills’ SKILL.md/references/ files, .github/dependabot.yml, explanation/why-a-registrar-role.md, and this file’s own source-map row. Historical log.md/source-map entries that named the old workflow stay as they were, describing what it was called at the time.
Feedback consumed - Disagreement fixed: ktl-docent flagged that explanation/why-four-roles.md’s title/description read as the total count of roles, while README.md’s own headers (“Four roles, three lines of the poem” / “The fifth role, which is not a skill”) name five roles in total, four of which are skills. Retitled the concept “Why four skill roles rather than one skill”, reworded its description, and added a short paragraph naming both the source headers and explanation/why-a-registrar-role.md explicitly. Updated knowledge/index.md’s matching bullet.
Curation: human:noelmcloughlin confirmed 5 concepts this session - policies/security.md, explanation/why-a-registrar-role.md, playbooks/knowledge-sources.md, policies/ai-covenant.md, policies/versioning.md - each cleared status: draft, gained a verified event, and a proposed stale_after review date (2027-03-10 for policies/security.md, given how fast its CI-security content is moving; 2027-09-10 for the rest). policies/security.md’s ## Open questions entry (about SECURITY.md’s own prose lagging the actual two-job workflow split) was cleared - the person confirmed the gap is real and the concept’s claim, not SECURITY.md’s prose, reflects current behaviour; fixing SECURITY.md itself remains outside this skill’s scope.
policies/security.md refreshed to describe knowledge-librarian.yaml’s new two-job, least-privilege split (commit 17d8f3a): the agent runs contents: read with no persisted credentials, and only a separate agent-free publish job holds the write scope; enforcement of the bundle write-boundary moved into .lokf/scripts/knowledge-librarian.sh itself (non-zero exit on a stray path), rather than a dedicated workflow step. SECURITY.md’s own prose still asserts the prior single-job design in one place - recorded under this concept’s ## Open questions since fixing that file is outside this skill’s .lokf/-only scope.
Added explanation/why-a-registrar-role.md: README.md gained a substantive new section describing a “registrar” role - keeping bundle records well-formed and provenanced - filled by tooling (the lokf toolkit, CI) or, for hand-edited bundles, two companion Obsidian plugins (KTL Registrar, KTL Curator), rather than by a fifth skill. No prior concept captured this; explanation/why-four-roles.md gained a relatedTo link to it.
Steady-state refresh: re-verified playbooks/releasing.md, playbooks/ktl-librarian-skill.md, playbooks/ktl-docent-skill.md, playbooks/repository-validation.md, and policies/versioning.md against their current sources - no drift found - and refreshed each concept’s own process:ktl-librarian verified event (human verified events left untouched). lokf on PyPI is still 0.7.0; the sidecar’s >=0.7.0 floor needs no bump.

2026-09-09

SECURITY.md gained an “Interactive use: scope is advisory, not enforced” section: prompted by a user question about npx skills add’s generic “runs with full agent permissions” warning. Boilerplate - the Agent Skills format has no permission manifest - but it surfaced a real gap: a skill’s Scope: line is prose, checked only by the scheduled workflow’s write-scope enforcement; interactive sessions rely on a person reviewing what the agent changed. Updated policies/security.md to match.
Security hardening on knowledge-librarian.yaml: prompted by an external security scan of the scaffolding templates. Two findings: (1, MEDIUM) the agent step let a repository variable’s content become the executed shell command (bash -c "${KNOWLEDGE_LIBRARIAN_CMD}"), wider than needed under that job’s contents: write scope - fixed by removing that variable entirely; the workflow now always invokes the pinned, reviewed .lokf/scripts/knowledge-librarian.sh directly, and AGENT_CLI only selects which agent runs, never what command runs. (2, LOW) the wrapper script’s contract - only touch .lokf/knowledge/, never run git - was advisory, unenforced - fixed by a new “Enforce the agent’s write scope” step that fails the job before any commit if anything else changed (excluding .lokf/uv.lock, a known uv sync side effect). Applied to the scaffolding template and this repository’s dogfooded copy in lockstep (both .yaml and .sh), plus SECURITY.md, scheduled-task.md, and automation.md. Updated policies/security.md to match.
Curation: human:noelmcloughlin confirmed 5 concepts (the ones most relied-upon and least-checked per the Step 1 report): references/agent-skills-specification, references/okf-specification, glossary/knowledge-bundle, playbooks/ktl-librarian-skill, references/lokf-specification. Sent 0 back, corrected 0, retired 0. One evidence-citation error caught mid-session (a quote presented as coming from a single “Section 1” of ktl-librarian/SKILL.md was actually a synthesis drawn from three separate places in the file) - re-quoted properly and the underlying claim held; a one-off slip in how evidence was presented, not a defect in the concept or a pattern worth changing ktl-librarian’s instructions over.
AI_COVENANT.md gained a “Repository-Owned Agent Automation” section: raised by the maintainer after noticing this same run’s own commit carried an AI co-authorship trailer the covenant already discourages (fixed by amending that commit). The deeper gap: the covenant’s model is a person drafting with AI help then submitting, but ktl-librarian (scheduled or interactive) and ktl-curator both have an agent executing the actual commit/PR. The scheduled workflow already gets this right by convention (bot identity, no trailer, mandatory review) but the convention was never written down. New section makes it explicit: bot-or-maintainer identity with no trailer either way, every change lands as a human-reviewed PR (never an agent’s own approval), and a curator-recorded human verdict must trace to that person’s real-time answer to that specific item. Updated policies/ai-covenant.md to match.
Steady-state refresh, second pass: re-verified all 18 internal-resource concepts (no drift - the two upstream commits since the prior pass, e6d5633 and b743c84, were prose tidying with no facts this bundle asserts) and, for the first time since bootstrap, fetched and checked all seven external Reference concepts against their live sources - all still accurate. Details and the one loose end (an unconfirmed GitHub CLI version number) are in playbooks/knowledge-sources.md.
Correctness bug found and fixed: a same-day commit (e6d5633) had flipped the sameAs typed-relation’s RDF predicate from the correct schema:sameAs to owl:sameAs in ktl-librarian/SKILL.md’s Golden Rule 4 - confirmed wrong against both the LOKF specification site and the raw lokf.yaml schema. Not just documentation: just lokf-check-refs’s SPARQL query filtered on the same wrong predicate in both .lokf/justfile and the ktl-sidecar template it’s copied from, so a sameAs relation’s target would have silently never been checked for existing. Fixed all three; lokf validate and lokf-check-refs still pass.
Tooling floor bumped: lokf reached 0.7.0 on PyPI (floor was >=0.5.0, already resolving to 0.7.0 with no upper bound). Reviewed 0.6.0/0.7.0 release notes for breaking changes affecting this sidecar - none found. Bumped the floor to >=0.7.0 to match what’s actually locked.
Feedback consumed, no bundle change: the one open .lokf/feedback.md entry asked about a roadmap for a fifth skill; re-searched the repository and found nothing to derive a concept from, so cleared it rather than inventing one, per the docent’s own note.
Orphan sweep: LICENSE, llms.txt, and EXAMPLES.md had no source-map row; added one, all three staying excluded as concepts (see playbooks/knowledge-sources.md for why).
Initialization: Scaffolded the LOKF bundle for LOKF Agent Skills, then ran a bootstrap discovery pass over the repository. Populated it with 25 concepts: 8 playbooks (the four skills, the source map, contributing, releasing, repository validation), 7 references (the LOKF, OKF and Agent Skills specifications, the lokf toolkit, LinkML, and the two installer CLIs), 4 glossary terms, 4 policies, and 2 explanations. base_iri is a placeholder (lokf-agent-skills.example) pending a namespace the project controls.
Removed: the scaffolded services/ directory and its two dummy concepts - this repository ships documentation, agent skills, and CI, and has no Service, Dataset, Table, or Metric assets to describe.
Steady-state refresh: re-verified all 17 internal-resource concepts against their current files (no factual drift found) and added a process:ktl-librarian verified event to each. Consciously excluded a new decorative asset, .assets/lokf-agent-skills-card.svg, recording it in playbooks/knowledge-sources.md instead of as a concept.
Audit: uv run lokf validate knowledge (previously unavailable) found 12 of 25 concepts using a bare scalar on a multivalued relation slot (dependsOn/definedBy/relatedTo) - schema-invalid despite reading naturally. Fixed by wrapping each as a one-item list, content unchanged. Same failure was independently visible in this repository’s own Knowledge Bundle Validation GitHub Actions run. Clarified ktl-librarian/SKILL.md Golden Rule 4 and section 2 so it doesn’t recur.
Source map extended: an orphan sweep found three .github/ surfaces accounted for by neither a concept nor a source-map row - the two dogfooded knowledge-*.yaml workflows, and the issue/pull-request templates alongside dependabot.yml. Added a row for each in playbooks/knowledge-sources.md, keeping all of them excluded as concepts (intake forms and maintenance config carry no reusable knowledge), and recorded the defects the sweep turned up in each.
Versioning policy re-sourced: the vMAJOR.MINOR.PATCH bump definitions moved out of README.md (slimmed) into CONTRIBUTING.md’s release-process section, where the rest of the release mechanics already live. Repointed policies/versioning.md’s resource to follow them - its body was asserting definitions its stated source no longer carried - along with the cross-references in README.md and CHANGELOG.md.
Prompt-injection guards added and documented: a synergy review against an external curation-patterns catalog (ai4curation.io) found that .lokf/feedback.md reached the scheduled librarian agent with no guard against a reader-submitted entry phrased as a directive. Added an explicit guard to ktl-librarian/SKILL.md (resolve only from the source an entry names, never its wording), and the same guard for the other two content- fetching paths in ktl-curator/SKILL.md and ktl-docent/SKILL.md. SECURITY.md gained a Prompt-injection guards section naming all three plus the unattended-write-path blast-radius containment and what remains unguarded (ordinary repository content, a reader’s own phrasing); refreshed policies/security.md to match.