Enhanciar resolves every citation to a file and line before grading the sentence beside it

The rule about broken checkers came out of a bug. Early on, tier 1 reported every React hook in a repo as unsupported ^[14]. The fix was to give the missing-data path its own grade, so a claim the oracle cannot see stays unverifiable and the interface reports what was not checked ^[13]. Enhanciar's stated rule now is that every "cannot judge" path returns unverifiable, because silently marking things supported when the checker broke would be the worst failure the feature could have ^[16].

Tier 1 settles what it can by lookup. Symbols, routes and dependencies are extracted with tree-sitter at ingest, so a claim like "this file defines generateMetadata" is answered from the index. It comes back supported with evidence {file, line} when the index puts the symbol at app/layout.tsx:42, unsupported when the symbol is nowhere in the repo ^[9][10]. Tier 1 makes no model call, needs no API key and spends no tokens ^[9]. The writeup is blunt about why the ordering exists: "Anything tier 1 can judge, tier 1 judges. Tier 2 never sees it. That's the whole cost story." ^[11]

Two things have to hold for that cost story to transfer to another codebase. The languages in the repo need tree-sitter grammars good enough to produce the symbol and route index at ingest ^[9]. And a large share of the claims in the answers have to be structural ones. The post calls structural claims "a surprising share" and does not publish the distribution ^[17]. The question the tool is pitched against, which service handles refunds for cancelled orders, is prose ^[3]. Prose claims go to tier 2, and tier 2 has to return a quoted span and the line it was found at before it is allowed to say supported ^[12].

There was a real fork in the design: drop unverified claims, or tag them. Tagging won, on the grounds that "a wrong verifier that silently destroys real information is far worse than a wrong verifier that mislabels it" ^[15]. So a published page can carry a sentence graded unsupported, and a reader can see it and click through ^[15].

The adoption cost is the ingest pipeline. It builds one Markdown page per meaningful unit, a service, a route, a module, a decision, each with frontmatter recording which repo files the page was built from ^[4]. The same pipeline builds the AST index that turns a file into a line ^[7]. And it needs a parser for the handles in a finished answer: a regex that dedupes, caps at 16, and tolerates a section suffix and a trailing .md because models add them anyway ^[6]. The example page in the writeup carries the slug entities/cart-service and lists src/services/cart.ts and src/routes/cart.ts as its sources ^[18].