046-Verbatim quotes are validated against their cited source

Type: ../types/adr.md · Status: accepted

Status: accepted
Date: 2026-07-12

Context

A verbatim citation asserts that a quoted span is copied exactly from a source retained in the KB. The dialectical/evidential profile has mandated the three-part convention — file link, prose locator, and a grounding-layer marker of verbatim / paraphrase layer / second-hand — since ADR 042. Until now nothing checked it: an author wrote a quote, wrote verbatim, and the claim was hand-maintained and trusted.

That is exactly the state a derived copy of recomputable truth must be checked or absent forbids. All four of its preconditions hold here: the derivation rule is a strict substring test, the quote names its source via the adjacent markdown link, the source snapshot is a checked-in file present at validation time, and the extraction heuristic is a single centralized artifact amortized across every citation it protects. The check is therefore Level A — deterministic, near-zero cost — not a judgment-based review gate.

The prototype supplied the missing evidence the proposal initially lacked — it did not originate from a felt friction case, and said so in writing. Run over the sibling epistack-casebooks corpus it found 63 match, 18 mismatch, 6 unresolved across 87 candidates, with no false mismatch on manual audit. The mismatches are real failures of the strict assertion, dominated by editorial omission (...), bracketed substitution, case changes, and punctuation moved inside the quotation boundary.

Decision

commonplace-validate resolves every verbatim-marked quotation against the markdown source it links, as a generic check alongside link health — not a type rule, because the trigger is the citation, not the note's type. A note that asserts a verbatim quote gets checked whatever it is.

Three outcomes, with deliberately asymmetric severity:

• mismatch → fail. The note asserts a falsehood. Per the derived-copy rule the false-copy failure mode is silent and unbounded, so it cannot be a warning.
• unresolved → warn, but only in notes carrying at least one resolvable verbatim quote. An unpaired verbatim citation is a parser-coverage gap worth surfacing where the convention is in use. Warning on it unconditionally would fire in every KB that merely writes about the convention — this repo has ten such candidates and zero real ones — and a check that cries wolf teaches authors to ignore it, which is the failure the check exists to prevent.
• match → pass, reported as a count.

commonplace-verify-quotes gives the corpus-wide view (match / mismatch / unresolved counts over many files) that an audit or a before/after sweep needs; the validator gives the per-note enforcement.

Both referential checks share one notion of code. This is the second check that dereferences — that resolves an element of a note against a second artifact — link health being the first, and the two must not disagree about what counts as content. A fenced code block demonstrating the citation convention is showing it, not asserting it; scanning one reports a false mismatch against whatever source the example links, which the first cut of this check did. Code fences are therefore neutralized in note_parser by a single primitive, blank_fenced_code_blocks, which both checks consume. It blanks rather than deletes: a check that only needs the set of matches (link health) is indifferent, while a check that pairs elements by proximity and reports line numbers (this one) needs the offsets to survive.

Consequences

The verbatim marker becomes load-bearing rather than decorative: it is now the cheapest of the three grounding-layer markers to assert and the only one that can be mechanically falsified. paraphrase layer and second-hand remain unchecked by construction — they never claimed exact text, and no derivation rule exists for them, which is precisely why a citation cannot assert more fidelity than its capture preserved.

Downstream KBs adopting the dialectical/evidential profile inherit the check with no configuration. The 18 live mismatches in the sibling casebooks are not fixed here: those notes are slated for a from-scratch rebuild, and the point of shipping the checker first is that the rebuild's citations land machine-checked rather than hand-trusted.

The extraction heuristic is not complete, and unresolved is how it says so rather than passing silently. Normalization (quote-mark folding, whitespace collapse, emphasis stripping, NFKC) is load-bearing, not polish: an under-normalized checker would produce false mismatches on correct quotes, which is worse than no check.

An open architectural question this decision surfaces rather than settles. The dividing line in validation is not frontmatter versus body — the schema already validates the body. ParsedDocument.to_validation_object() hands the schema body, headings, links, and body_dates, and several type schemas assert required headings. Anything about a note's own text that the parse model exposes is expressible declaratively, and is type-owned.

What the schema cannot express is dereferencing: it has no way to say follow this path and look inside the artifact it names. So the real class is referential checks — does this link's target exist (link health), does this quoted span occur in the source it cites (this ADR). Their ground truth lives outside the document, which is precondition 3 of the derived-copy rule, and it is why they are hand-written imperative passes rather than schema constraints.

That class has no design. It has no shared model of a positioned element, no shared severity policy, and no owner. Sharing the code-fence primitive removes the immediate divergence without supplying one, and the shared parse remains lossy for it: ParsedDocument.links is a tuple of URLs with no positions, which suffices for link health (it needs only the set) but not for proximity pairing, so the quote checker still carries a private link regex. The gap belongs to the kb-graph-loader workshop — and it is the same gap, since a referential check is a graph edge being resolved.

Supersedes and retires the verifiable-quotes proposal, whose content has shipped.