KB Review

Run the canonical governance lane between draft knowledge and the approved live brain. kb-review owns two internal modes:

• gate mode for pending draft decisions, promotion/rejection, and briefing rebuilds tied to the same review pass
• maintenance mode for approved-layer drift, stale briefings, provenance and alias refresh, creator consistency checks, governance-index rebuilds, deterministic navigation-index rebuilds, graph gaps, backlog pressure, and safe mechanical fixes

In Karpathy's LLM Wiki, this maps to two operations: the implicit quality judgment during ingest (our gate mode), and the explicit "Lint" pass he describes — "look for contradictions between pages, stale claims that newer sources have superseded, orphan pages with no inbound links, important concepts mentioned but lacking their own page, missing cross-references" (our maintenance mode). This contract makes both operations explicit and separable.

When a user says output/reports, decide whether they mean a governance report or a deterministic rendered report. Lint, contradiction, orphan-page, and stale-claim work stays here in maintenance mode.

Minimal loop

1. inspect the draft package, cited captures, and overlapping live pages
2. decide approve, reject, or escalate for human judgment
3. promote only durable, well-supported knowledge into live
4. rebuild briefings when the approved change should alter downstream context

When this compounds the wiki

kb-review is where candidate structure becomes durable truth. Good review keeps the brain dense and reusable without letting noisy, redundant, or weakly-linked material accumulate in wiki/live/.

When not to promote

A draft can be technically correct and still fail the promotion bar if it is redundant, weakly provenanced, poorly linked, or better expressed as a relationship or hub update instead of a new live page.

Read before reviewing

Read these files first:

• local AGENTS.md
• local CLAUDE.md if present
• ../obsidian-notes-karpathy/scripts/skill-contract-registry.json
• ../obsidian-notes-karpathy/references/file-model.md
• ../obsidian-notes-karpathy/references/schema-template.md
• ../obsidian-notes-karpathy/references/review-template.md
• ../obsidian-notes-karpathy/references/briefing-template.md
• ../obsidian-notes-karpathy/references/health-rubric.md
• ../obsidian-notes-karpathy/references/provenance-and-alias-policy.md

Treat skill-contract-registry.json as the canonical source for required references, baseline command, and expected write surfaces.

If onkb is available, run:

• onkb --json review queue <vault-root>
• onkb --json query scope <vault-root> when validating downstream boundaries
• onkb --json review lint <vault-root> for maintenance-mode diagnostics
• onkb --json review governance <vault-root> when maintenance mode should refresh QUESTIONS.md, GAPS.md, ALIASES.md, ENTITIES.md, or RELATIONSHIPS.md
• onkb --json review indices <vault-root> when maintenance mode should check the retrieval navigation indices (INDEX.md, CONCEPTS.md, SOURCES.md, TOPICS.md, RECENT.md, plus the managed onkb:indices block in wiki/index.md) for drift; add --write to rebuild them deterministically from the approved live layer. EDITORIAL-PRIORITIES.md stays editorially owned and is never rewritten by this command.
• onkb --json review graph <vault-root> when the user wants machine-readable graph export during maintenance

If onkb is missing, follow the install fallback in ../obsidian-notes-karpathy/references/lifecycle-matrix.md, then rerun the same command.

Load on demand

Load these only when the trigger applies:

• ../obsidian-notes-karpathy/references/chinese-llm-wiki-compat.md — when the request leans on Chinese-LLM-Wiki vocabulary beyond the mappings inlined in this SKILL
• ../obsidian-notes-karpathy/references/archive-model.md — when archive reuse, truth-status, or output-surface posture questions arise
• ../obsidian-notes-karpathy/references/lifecycle-matrix.md — when onkb is missing (install fallback) or a stage handoff is unclear
• ../obsidian-notes-karpathy/references/activity-log-template.md — when appending an entry to wiki/log.md
• ../obsidian-notes-karpathy/references/search-upgrades.md — when retrieval quality tuning or scaling beyond the default ranking is on the table
• ../obsidian-notes-karpathy/references/questions-and-reflection-policy.md — when standing questions, reflection notes, or QUESTIONS.md governance is involved
• ../obsidian-notes-karpathy/references/query-writeback-lifecycle.md — when writeback candidates, writeback_status, or archive backlog work is involved
• ../obsidian-notes-karpathy/references/memory-lifecycle.md — when MEMORY.md or episodic memory boundaries are involved
• ../obsidian-notes-karpathy/references/graph-contract.md — when graph exports, related edges, or relationship candidates are involved
• ../obsidian-notes-karpathy/references/source-manifest-contract.md — when raw/_manifest.yaml fields or manifest anomalies are involved
• ../obsidian-notes-karpathy/references/topic-template.md — when shaping or promoting browse-layer topic pages
• ../obsidian-notes-karpathy/references/profile-contract.md — when account profiles or source_profile metadata is involved
• ../obsidian-notes-karpathy/references/automation-hooks.md — when wiring or invoking onkb review automation modes
• ../obsidian-notes-karpathy/references/episode-template.md — when outputs/episodes/** work is involved
• ../obsidian-notes-karpathy/references/procedure-template.md — when the durable delta is procedural (workflow or playbook)

Independence rule

Review the draft package, the directly cited raw captures, and overlapping live pages only.

Do not preserve a draft merely because the producing agent seems competent or because the generation path looks sophisticated.

Mixed gate policy

• auto-approve when the draft has no blocking flags, no unresolved live conflict, no unresolved alias/duplicate concern, and the review score clears the threshold
• auto-reject when evidence is weak, blocking flags remain, or the draft is clearly unsafe to compound
• send to human review when the draft conflicts with live knowledge, source integrity is suspect, or the score band is ambiguous

When the draft is borderline, weigh both accuracy and whether the page deserves promotion into the long-term brain. A technically correct draft can still stay out of wiki/live/ if it is noisy, redundant, weakly provenanced, or not durable enough to merit reuse.

What reviewers must judge explicitly

• are source claims and compiler inferences cleanly separated?
• is source integrity still credible?
• are aliases aligned with the existing approved vocabulary?
• does this draft duplicate an existing approved page?
• are contradictions recorded rather than silently overwritten?
• does the page deserve durable retention in the long-term brain?
• should the durable delta land in wiki/live/procedures/ instead of another semantic page?
• is the confidence metadata ready to be promoted, reduced, or left unset?
• should the durable improvement be a promoted page, a stronger relationship edge, or inclusion in a curated hub?
• does this draft improve the browseable graph of the wiki, or does it just add more prose?

Procedural promotion

When a draft carries promotion_target: procedural in its frontmatter, promote the durable delta into wiki/live/procedures/{slug}.md instead of wiki/live/concepts/. Use the procedure frontmatter from ../obsidian-notes-karpathy/references/procedure-template.md (procedure_id, confidence_band, decay_class, next_review_due_at, etc.).

Promote procedurally when the durable delta answers "how to do X" rather than "what X is" — a workflow, playbook, or repeated decision pattern. Typical signals: the draft reads as a sequence of steps, the same pattern applies across future situations, repeated question clusters resolve into the same sequence of actions.

Decision rules:

• if promotion_target: procedural is set and the signals hold, promote into wiki/live/procedures/
• if the flag is set but the content is actually semantic, flip to promotion_target: semantic in the review record and promote into wiki/live/concepts/
• if both targets plausibly fit, prefer procedural when the reuse value is in the sequence of actions; prefer semantic when it is in a definition
• if neither target fits cleanly, record the ambiguity in the review record and send to human review

Outputs

kb-review may write:

• read and resolve pending work from wiki/drafts/**
• outputs/reviews/*.md
• promoted pages under wiki/live/**
• promoted browse-layer topic pages under wiki/live/topics/**
• refreshed governance indices under wiki/live/indices/**
• rebuilt navigation indices (INDEX.md, CONCEPTS.md, SOURCES.md, TOPICS.md, RECENT.md) and the managed onkb:indices block in wiki/index.md via onkb review indices --write
• regenerated wiki/briefings/{role}.md
• maintenance reports under outputs/health/**
• graph snapshot exports under outputs/health/graph-snapshot.json
• review and brief entries in wiki/log.md

It should not mutate raw captures.

outputs/qa/** and outputs/content/** are owned by kb-query. kb-review touches them only as writes_mechanical_fix_only targets declared in scripts/skill-contract-registry.json — deterministic, unambiguous mechanical fixes during maintenance mode (broken links, stale frontmatter, scope metadata drift, obvious typos). Any substantive new content or re-synthesis belongs back in kb-query, not here.

Mode selection

Use gate mode when:

• drafts are pending
• a human approval boundary is the next safe step
• briefings should be rebuilt because the current review pass changed approved truth

Use maintenance mode when:

• the approved layer is drifting, contradictory, stale, or weakly linked
• archived answers or content have writeback backlog
• archived outputs have reuse drift, stale claims, or private/shared scope leaks
• confidence metadata, supersession bookkeeping, or audit trails have decayed
• governance indices, retrieval navigation indices, or graph exports need deterministic refresh
• safe mechanical fixes in approved or archived surfaces are clearer than creating new prose
• creator-facing guidance surfaces such as CLAUDE.md, MEMORY.md, account _style-guide.md, or account briefings may be drifting apart
• archived publish outputs are not reusing prior approved coverage cleanly
• important approved sources are staying underused in downstream Q&A or publish artifacts

kb-review is now the only public governance surface. If the user literally says kb-health, treat that as kb-review maintenance mode.

Review records should capture whether fact and inference are separated cleanly enough for safe reuse, plus a short promotion reason explaining why the page should or should not persist.

Archive hygiene in maintenance mode should stay explicit:

• stale archived Q&A relative to newer live pages
• archived publish artifacts that no longer reuse the best approved coverage
• archive backlog with writeback_status: pending
• scope leaks or sensitivity metadata drift in archived outputs

For backlog stuck at pending with followup_route: draft, the standard disposal action is the deterministic writeback lane: run onkb --json compile writeback <vault-root> as a dry-run report, then --write to scaffold drafts and advance those artifacts to drafted. The scaffolds re-enter this skill through the normal gate. Backlog with followup_route: review stays here as maintenance judgment work.

This is maintenance of the artifact archive, not promotion of archive into truth.