Verified Citations

Verified citations

A citation is output you produced, never a reference you recalled. This skill exists because recall is confident and wrong in a narrow, repeatable way, and because the model committing the error reliably reports that it checked.

The measured failure mode

Two independent runs on the same repository, one building a skill library, one authoring an implementation design, both reviewed by a different model:

• 58 of 62 citations exactly right. Trigger names, ERRCODEs, migration numbers, statutory message prefixes, function signatures: all correct. Technical grounding is not the weakness.
• The misses were all about WHERE and WHEN, not WHAT:
  • A path invented around a borrowed line number. A spec named a file with no directory and cited :47-49. The author reproduced the line numbers correctly and fabricated a plausible directory to complete the reference. The file was real; the path was not.
  • A line number drifted onto nothing. :239 was cited as a function; 239 was a blank line between two functions.
  • Ranges that were loose rather than wrong. A cited :82-118 contained the claimed content and also a comment block and a grants preamble, presented as one coherent thing.
  • Provenance misattributed. "Landed 2026-07-01 under audit C4" when git log --follow showed 2026-06-15 under C5.
• Every run self-reported full verification. The error is invisible from the inside. That is the single most important fact in this file.

The rule that does the work

Every file:line citation carries the quoted text of the line it points at.

This is not a style preference. It is the only rule here that cannot be satisfied by intending to be careful, because you cannot quote a line without having opened the file at that path and looked at that line. It kills the two commonest failures outright: a fabricated path has no line to quote, and a drifted line number quotes a blank string.

It also makes review mechanical. A reviewer runs sed -n 'Np' FILE and compares, instead of re-deriving your argument to decide whether they believe you.

If quoting every line would bloat the document, quote in the citations table at the end (below) and keep bare file:line inline. The table is the evidence; the inline reference is the pointer.

Four supporting rules, one per observed failure

1. Never complete a partial reference from another document. If a source names a file without a directory, or a symbol without a file, resolve it yourself with a glob or find and cite what the search returned. Inheriting :47-49 from someone else's citation does not license inventing the path it hangs off. Correct line numbers are not evidence of a correct path.

2. Cite the narrowest range that contains the claim, and check both boundaries. A range is a claim that everything inside it is the thing you are describing. Read the first and last line of any range you cite. If they belong to different constructs, the range is wrong even though the content is in there somewhere.

3. Provenance is the highest-risk class and needs its own command. Any claim of the form "added on DATE", "landed in PR N", "shipped under audit item X", "introduced by commit SHA" requires git log --follow, git show, or the forge's API, run now. Never carry a date, PR number or SHA from memory or from a summary. In a squash-merge repository, note that "merged" is a PR fact and not a git-ancestry fact, so git branch --merged will lie to you in both directions.

4. Write "unverified" out loud. An explicit gap costs a reader nothing. A confident wrong citation costs a reviewer their entire pass and can outlive the document in things that quote it. If a tool was unavailable or the check was skipped, say which claim is affected and why.

The output contract

End any document that makes more than a handful of citations with a table the reviewer can check mechanically:

"How verified" is the command you actually ran (grep -n, sed -n, git log --follow, git show). If a row cannot name a command, it is unverified and must say so in that column.

"I verified everything" is not admissible. The table is the evidence. Do not write a summary sentence claiming the citations were checked, because that sentence has been true-sounding and false in every recorded instance of this failure.

Scope, and what this does not replace

This applies to whoever is writing, not to one model. The same session that caught a design agent inventing a directory also produced two first-hand errors of the same genus, both asserted confidently from structural inference, both one command away from being settled: "widening this trigger alone would break production" (the caller made it irrelevant) and "the CLI is not installed on this machine" (it was simply not on PATH). A reference you derived is not safer than one you were handed. It only feels safer.

Adjacent rule, same root: a failure message is a symptom, not a diagnosis. "Not recognized" means not on PATH. Establish that a blocker is real before designing around it, because a workaround built on a phantom blocker is usually worse than the path it replaced, and it will not be reviewed as carefully because it looks like a solved problem.

This skill lowers the error rate. It does not remove the need for an independent verification pass. The 58/62 figure above was achieved without this skill, and the reviewer still earned its keep on the two that mattered. If a workflow adopts this and quietly drops the adversarial review because "the skill handles it", that trades a control that works for one that merely feels diligent.

When you do run that review, brief it at provenance specifically. A reviewer asked to critique a document returns opinions. A reviewer asked to verify every path, line and date against the repository returns findings.

Statutory, legal, and domain facts (added 2026-07-23)

The same discipline extends beyond code and history to any encoded real-world obligation:

• Verify Act, year, AND Part against the primary source (legislation.gov.uk for UK statutes); model memory produced two confident wrong "facts" headed for a statutory register (a non-existent statutory interval; a mis-attributed legal basis).
• Statutory day-counting is inclusive ("beginning with the day X" includes day X).
• Data that encodes an obligation (statutory frequencies, legal bases, deadlines) cites its primary source in the artifact itself, so the next reader verifies rather than trusts.
• A specialist reviewer contradicting the brief on law earns a primary-source check, not an override — the brief handed to a verifier is itself a claim.

Single-homing note: "a failure message is a symptom, not a diagnosis" lives canonically in findings-are-evidence; this skill defers to it.

Routes

• Generated client documents (rewrites, owner-only facts, machine-parsed output) → load deliverable-integrity.
• Customer-facing certification/statistic claims → load substantiate-outward-claims.
• Producing a handover → load handover.