| name | llm-wiki-quality-gate |
| description | Quality gate runtime skill for LLM Awesome Wiki. Use when an agent needs to run or record compare reports, source/wiki coverage, claim/modality/contradiction review, review queues, validation notes, workspace checks, or round closure decisions. |
LLM Wiki Quality Gate
Role
Use this subskill for compare, review, validation, and closure. It is part of
knowledge-base construction, not downstream skill + tool generation.
This skill owns the normal quality-gate runtime. It does not rewrite wiki pages
or use model self-evaluation as a pass.
Detailed rules are loaded only when needed:
rules/wiki/source-wiki-coverage-protocol.md for source/wiki coverage,
omission, disposition, and anchor tables
rules/claims/claim-modality-contradiction-review-protocol.md for claim
support, generated evidence, modality, unsupported statements, and
contradictions
rules/review/review-queue-workflow.md for review lifecycle and blocking
level
Inputs
- round plan and fixed scope
- source inventory and source packets in scope
- semantic draft or wiki construction analysis when wiki pages changed
- source outline or coverage plan for document/PPT wiki rounds
- changed or inspected wiki pages
wiki/overview.md, wiki/index.md, and wiki/log.md
- claim/evidence records when important claims are in scope
- current review queue
- validation notes and tool output
Missing expected inputs must be recorded. Missing input is not an implicit
pass.
Runtime
- Run available validators for the touched artifact family.
- Write or inspect a compare report. It should answer:
- what was in scope
- whether the round was
full-round, selective, or overview-only
- which source packets and anchors were checked
- whether core source units are covered at the right detail level, weak,
deferred, omitted, or routed to review
- whether the accepted wiki preserves the semantic draft's important
definitions, formulas, derivations, examples, tables, and implications
- whether core formulas and derivations are represented or explicitly block
knowledge coverage
- whether useful candidate-derived content is grounded, narrowed, or routed
to review instead of silently discarded
- which wiki pages changed or were inspected
- whether important claims have source, evidence, accepted review, or review
routing
- which generated, skipped, or uncertain modalities matter
- whether links, index, overview, and log are current
- which omissions, contradictions, unsupported statements, or review items
block advancement
- Use only compare status values:
- Do not treat
compare gate not enabled, model confidence, or "looks
complete" as pass.
- Write or inspect the validation note.
- Decide closure:
close-pass: compare is pass, validation exists, no blocking review
remains, knowledge coverage is pass, generated evidence is marked or
reviewed, semantic draft richness is not materially lost, and
navigation/log state is acceptable
close-with-review: compare is needs-review, artifacts are structurally
reviewable, and carried-forward review has reason, next action, target
round or condition, and blocking level. This can carry nonblocking
modality review or explicit overview/selective deferrals, but not a
full-round core coverage hole or missing core formula/derivation coverage.
do-not-close: compare is fail, required artifacts are missing,
blocking review remains, full-round core knowledge coverage is weak or
missing, accepted wiki is materially poorer than the semantic draft, or
index/overview/log/claim/source coverage is materially stale
- Update workspace logs and commit only accepted
close-pass or explicitly
allowed close-with-review state.
Outputs
- compare report or explicit validation note that compare is not enabled
- source/wiki coverage, claim support, modality, unsupported statement,
contradiction, and review findings at the level needed for the round
- review queue updates for unresolved judgment
- validation note
- closure decision
- next action or next round
Validation
Prefer the narrowest relevant check first, then broader checks:
python -m llm_wiki_tools workspace-check --workspace . --mode reports
python -m llm_wiki_tools workspace-check --workspace . --mode closure
python -m llm_wiki_tools workspace-check --workspace . --mode all
If checks are unavailable, record the missing checker state. Do not convert
missing validation, model confidence, or "looks complete" into pass.
Minimum manual checks:
- report status is
pass, fail, or needs-review
- accepted wiki changes remain discoverable from index/log/overview state
- important claims have source anchors, evidence, accepted review, or review
routing
- accepted wiki pages did not drop important semantic draft content without
grounding, narrowing, deferral, or review routing
- core document/PPT source units have anchor disposition or explicit
overview/selective deferral
- knowledge coverage review is separate from modality review
- formula/equation/derivation gaps that support accepted knowledge affect
knowledge coverage status, not only modality status
- generated evidence is visible and not treated as source-derived truth by
default
- unresolved review is carried forward with reason and next action
close-pass is never used with missing compare, blocking review, or stale
navigation