Verifiable quotes

Type: kb/types/note.md · Status: seedling · Tags: document-system

A verbatim-marked citation asserts that a quoted span is copied exactly from a source. That assertion is currently hand-maintained and unchecked: an author (or agent) writes "quote" plus a citation saying verbatim, and nothing confirms the quote actually appears in the referenced source. That is precisely the state a derived copy of recomputable truth must be checked or absent forbids — a copy of information recomputable from a ground-truth source, trusted by hand instead of machine-checked. This proposal designs the check.

Current state (as of 2026-07-09)

  • The grounding-layer marker convention — every proposition cites a source with a prose locator plus one of verbatim / paraphrase layer / second-hand — is live in the sibling epistack-casebooks repo's kb/lhc/notes/COLLECTION.md, as part of the dialectical/evidential profile's mandatory-evidentiality clause. It is now also recorded as a shipped default in this repo's text-contract profile catalogue (dialectical/evidential entry), so any future collection adopting that profile inherits the same three-part citation shape.
  • The convention is prose-shaped, not machine-structured. Reading the five live casebook notes: a verbatim quote appears either as a blockquote (> "...") or an inline quoted span, followed within a sentence or two by a parenthetical citation carrying a markdown link to the source snapshot and the literal word verbatim (e.g. ([LSAG report](../sources/lsag-2008-review-safety-lhc-collisions.md), Abstract, verbatim.)). No locator syntax or anchor format was ever specified beyond this pattern — the sibling repo's own COLLECTION.md says as much: "A structured source-span locator type remains unadopted; revisit if review tooling ever needs machine-checkable spans (e.g. a staleness gate that re-verifies quotes against snapshots)."
  • Five casebook notes with roughly a dozen verbatim citations already exist as a ready-made prototyping corpus. This proposal did not originate from a felt friction case — no author has yet been burned by an unverified quote drifting from its source — it originates from applying the derived-copy rule to a convention that already exists. The existing corpus means it can still be exercised without waiting for new casework.
  • Factored dependency pairs for review freshness names source-as-gate as a different, not-yet-adopted mechanism: a (note, source) freshness pair for judging whether a note's distillation of a source is still consistent — a Level B, LLM-judged review gate. This proposal is not that. Verbatim-quote matching is fully decidable (does string X occur in file Y), so it needs none of the freshness-pair/review-job machinery — it is a Level A deterministic check, the same tier as link validity or required-heading checks, not Level B judgment.

The design

What gets checked. For every verbatim-flagged citation in a note: extract the quoted span, resolve the linked source file, and test whether the span occurs in that file (after normalization — see below). paraphrase layer and second-hand citations are explicitly out of scope; they never claimed exact text.

Applying the derived-copy rule's four preconditions (from the grounding note) confirms this is enforceable, not just checkable-in-principle:

  1. Derivation rule — a strict substring test, not a judgment call. Met.
  2. Machine-locatability — the quote names its source via the adjacent markdown link; the check follows that link the same way link-health already does. Met, contingent on the extraction heuristic below actually finding the pairing.
  3. Ground truth exists at validation time — the source snapshot is a captured, checked-in file. Met (snapshots are stamped-not-authored and don't change except on an explicit refresh, per kb/sources/COLLECTION.md).
  4. Validator expected to bottom out — the extraction heuristic (below) is itself a hand-authored pattern-match, not a strict grammar. Per precondition 4's own allowance (the tag-README complete mark's scoped rg sweep is the precedent), a centralized, versioned heuristic amortized across every note is still far cheaper than N hand-trusted claims, even though the heuristic itself isn't provably complete.

Extraction heuristic (no new citation syntax required). Rather than inventing a structured citation format, detect the pattern already in use: a quoted span (blockquote or inline "...") paired with the nearest citation in the same or next sentence/paragraph that contains both a markdown link and the token verbatim. This will not resolve every case — nested quoting inside a paraphrase-layer citation, or a verbatim citation separated from its quote by more than a sentence, will misfire. Three outcomes per candidate, not two: match (quote found in source), mismatch (quote not found — the claim is false), unresolved (couldn't confidently pair a quote with a citation — a parser limitation, not necessarily a defect, but worth surfacing so an author can tighten the citation).

Normalization before comparison. Exact byte match is too strict: smart vs. straight quotes, whitespace collapsing (line-wrapped PDF extraction), and minor capture-tool artifacts are expected even for genuinely verbatim text. The check needs a normalization pass (quote-mark folding, whitespace collapse, at minimum) before substring comparison — otherwise it produces false mismatches on correct quotes, which is worse than not checking at all (it trains authors to distrust the tool).

Where it runs. A free choice — see below — but the check's shape (cross-file body-text comparison, not frontmatter/structure) doesn't fit commonplace-validate's existing JSON-schema machinery, which validates one document's frontmatter and structure, not body content against a second file. It is closer in kind to link-health (which already parses markdown links and resolves paths) than to schema validation.

Forces

  • For: this is exactly the rule the KB already committed to. An unchecked "verbatim" claim is the textbook hand-maintained-and-trusted state the derived-copy rule forbids — the false-copy failure mode (silent, unbounded wrongness) is worse than the absent-copy one (a bounded recomputation), and "verbatim" is precisely the kind of claim a reader stops independently verifying once the KB asserts it.
  • For: cheap once built, and the corpus already exists. Level A checks are fast, deterministic, and reusable — no per-note review cost the way a semantic gate would add. Five existing notes mean it can be validated against real data on day one.
  • Against: no felt need yet. No worked case has actually produced a wrong verbatim citation; this is anticipatory, which cuts against the KB's own build-local-first, prove-it-first discipline. Mitigation: the existing corpus substitutes for a live-friction trigger, and the check is cheap enough (Level A, no new type or schema) that building it doesn't accrue the kind of speculative-machinery debt heavier proposals would.
  • Against: the extraction heuristic isn't complete. Some verbatim citations will land in unresolved rather than being checked at all, especially as citation prose style varies across authors and cases. Mitigation: report unresolved as its own category rather than silently passing it, so coverage gaps are visible rather than assumed away.
  • Against: false mismatches from under-normalization would be worse than no check. A checker that cries wolf on correct quotes teaches authors to ignore it, which is the outcome the derived-copy rule exists to prevent one level up. Mitigation: normalization is not optional polish here, it's load-bearing; ship nothing until the five-note corpus passes clean.

Free choices

  • Where it lives. A flag on commonplace-validate, a dedicated command (commonplace-verify-quotes, following the existing per-purpose command pattern), or a scripts/ accumulation-substrate script per ADR 040 that graduates to a package command once it proves durable. Leaning toward starting in scripts/ — this is unproven machinery on an unproven need, exactly what that substrate is for.
  • Severity on mismatch. fail by default per ADR 024's stance (softness is the marked case), or warn initially given the extraction heuristic's imprecision risk, tightening to fail once the false-positive rate is known from real runs.
  • Whether unresolved blocks anything. Treating it as pure signal (never blocks, always reported) versus a soft nudge (warns after some grace period) is undecided; leaning toward pure signal until there's evidence authors respond to it.
  • Scope beyond the dialectical/evidential profile. The grounding-layer marker convention is currently specific to that profile. Nothing about the check is profile-specific in principle — any collection citing sources with a verbatim flag could use it — but there's only one instance to design against today.

Adoption criteria

Prototype against the sibling repo's existing five casebook notes before writing anything into this repo's shipped tooling. Adopt once the prototype (a) achieves a low unresolved rate on real citations without hand-tuning per note, and (b) produces zero false mismatches after normalization — both are cheap to measure against the existing corpus, so this doesn't need to wait for a sixth note or a new case.

Relevant Notes: