Context
curate_wiki already produces authoring jobs from clustered memories, and the file-doc skeletons already expose curation gaps rather than hiding them. What is missing is the link: a deterministic structural skeleton arriving as the input to that machinery, instead of the curator inferring page decomposition from memory clusters alone.
Provenance is the differentiator worth building for, but the claim has to be stated precisely, because tools in this category already enforce citation discipline. At least one competing generator labels every node of every diagram it emits with file:line, requires two or more file:line citations behind every duplication claim, and asks its own sub-agents for a confidence note plus a list of known gaps. "We cite our sources" is not a differentiator.
The difference is who vouches for the citation:
- theirs — a model writes the citation and another model rejects reports that lack one. Self-policed. A plausible-looking hallucinated
file:line passes.
- ours — the tag is derived.
extracted comes from the AST, inferred comes from the resolver, and the gaps come from a measured missed graph rather than from a model's self-assessment.
Hence three tags rather than a confidence adjective:
extracted — read from the ASTinferred — produced by the resolverauthored — written by a model
The first two arrive in the skeleton (cdeust/ai-architect-mcp-codebase#196) and are machine-checkable. This issue adds the third and, more importantly, guarantees the first two survive curation rather than being paraphrased into unsourced prose. A curated page that quietly downgrades a derived claim to an authored one has destroyed the only property that separates this output from a well-prompted one-shot generator.
Scope
Consume the skeleton through the existing curate_wiki queue and emit curated pages.
Acceptance criteria
- Skeleton in, curated pages out, through the existing queue. No parallel authoring path is introduced.
- Every paragraph making a structural claim carries a provenance tag.
authored is permitted only where no skeleton claim backs the sentence; a test asserts a curated page cannot tag a skeleton-backed claim as authored.
file:line citations from the skeleton are preserved verbatim in the rendered page. A test reconciles a curated page's citations against the skeleton it came from, catching paraphrase drift. This reconciliation is the machine-checkable guarantee the differentiator rests on; if it is only spot-checked, the differentiator is rhetoric.- A citation that does not resolve to a symbol at that location in the indexed tree fails the build rather than shipping. Asserted with a deliberately corrupted skeleton.
- Coverage gaps carried by the skeleton appear on the curated page. A test asserts that a page generated from a skeleton with a non-empty coverage block renders that block, and that an empty block renders as an explicit "no known gaps" rather than vanishing.
wiki_verify passes on the generated set.- Generated pages are scored against the 42 canonical scopes in
mcp_server/core/wiki_coverage.py; uncovered scopes are reported as gaps in the output, not omitted.
- Scoped mutation run on the changed files: zero surviving non-equivalent mutants, or each survivor documented as equivalent.
- Completion Ledger in the PR.
Depends on
Out of scope
Context
curate_wikialready produces authoring jobs from clustered memories, and the file-doc skeletons already expose curation gaps rather than hiding them. What is missing is the link: a deterministic structural skeleton arriving as the input to that machinery, instead of the curator inferring page decomposition from memory clusters alone.Provenance is the differentiator worth building for, but the claim has to be stated precisely, because tools in this category already enforce citation discipline. At least one competing generator labels every node of every diagram it emits with
file:line, requires two or morefile:linecitations behind every duplication claim, and asks its own sub-agents for a confidence note plus a list of known gaps. "We cite our sources" is not a differentiator.The difference is who vouches for the citation:
file:linepasses.extractedcomes from the AST,inferredcomes from the resolver, and the gaps come from a measured missed graph rather than from a model's self-assessment.Hence three tags rather than a confidence adjective:
extracted— read from the ASTinferred— produced by the resolverauthored— written by a modelThe first two arrive in the skeleton (cdeust/ai-architect-mcp-codebase#196) and are machine-checkable. This issue adds the third and, more importantly, guarantees the first two survive curation rather than being paraphrased into unsourced prose. A curated page that quietly downgrades a derived claim to an authored one has destroyed the only property that separates this output from a well-prompted one-shot generator.
Scope
Consume the skeleton through the existing
curate_wikiqueue and emit curated pages.Acceptance criteria
authoredis permitted only where no skeleton claim backs the sentence; a test asserts a curated page cannot tag a skeleton-backed claim asauthored.file:linecitations from the skeleton are preserved verbatim in the rendered page. A test reconciles a curated page's citations against the skeleton it came from, catching paraphrase drift. This reconciliation is the machine-checkable guarantee the differentiator rests on; if it is only spot-checked, the differentiator is rhetoric.wiki_verifypasses on the generated set.mcp_server/core/wiki_coverage.py; uncovered scopes are reported as gaps in the output, not omitted.Depends on
Out of scope