Source check

Most citation checking stops at "the work exists". That catches typos in a page range and nothing else. The errors that damage an argument are of a different kind: the work exists, is cited correctly, and does not say what it was cited for. Those survive every reference manager and every reviewer who recognises the name.

So this skill runs two passes that stay separate, and writes them to two files, because collapsing them lets the weaker pass borrow the authority of the stronger one.

A bibliography that has been through pass 1 only is a verified bibliography. A checked one has been through pass 2 as well. The file must say which of the two it is, in those words. Overclaiming here is the one failure mode that makes the whole exercise worthless.

The rules that make this worth doing

Never confirm a citation you have not opened. That rules out "this is the standard reference for X", a page number reconstructed from memory, and a characterisation of an author's argument based on how they are usually summarised. Secondhand citation is precisely what produces the errors this skill exists to catch: an argument described as logical when the author is explicit that it is physical, a theorem described as proving randomness when it proves a martingale property.

Open the block around every search hit. A pattern search tells you where the tool matched. Before stating what a file says, read the enclosing block around the match: two lines adjacent in the search output can belong to different constants in the file, and one audit lifted a regulation citation from the constant below the one it described. Before stating that something is absent, search for the concept under every spelling it could take (a check date can live in a constant, a comment and a rendered string under three names), and say how you looked. A claim of absence is the weakest claim available: ask what the strongest evidence would look like if the thing were present, and whether your search would have found that.

Quote verbatim, with a page number. A pass-2 finding is only checkable by the next reader if it carries the words. Paraphrase is where the original error crept in; repeating the paraphrase does not test it. Quote whole sentences or none: an internal ellipsis is an edit, and the clause that looks redundant is usually the one carrying the reading. If length is a constraint, quote a shorter continuous span and put the omitted material in a bracketed gloss that is visibly yours. Have the person who retrieved the text check the transcription; the assembler can no longer see the clause they cut.

Say what you did not check. The credibility of the output comes from the scope statement ("six of the twenty works were checked at primary-text level; §4 lists the fourteen that were not"), far more than from the checks themselves. An honest gap is an asset. A hidden one is what sinks the document later, and it sinks the checked claims along with it. The counts in the scope statement are claims too: derive every one from a command whose output you saw, and state the command. A listing eyeballed at 42 files counted 85 when piped to wc -l. If duplication inflates a count, say so in the same sentence.

The same rule covers your own verification. "Grepped committed history for X" names a method; a scope statement adds the set and what a negative result excludes: "grepped committed history for string X; does not exclude other names or other identifying content". A negative result is a fact about the query. Reporting it without its coverage turns a narrow check into a broad assurance, and the assurance stops the next person running the query that would have found it.

Mark how each source was reached. Full text, a scanned page image, an abstract, a secondary source quoting it, a publisher landing page. These are not equivalent and the entry should not present them as if they were. Record the route that produced the bytes, with URL, date and byte count.

Flag load-bearing failures. A failed claim in a passing aside and a failed claim the argument rests on are different events. Say which one you are looking at, in the summary line, before the reader has to work it out.

Workflow

1. Extract every claim that rests on a source

Read the document and list, separately:

• Citations: every work referenced, with what it is being used to support.
• Claims about what a source says: the sentences that characterise an author's argument, result or position. These are pass-2 material and they are easy to miss, because they often appear without a citation directly attached.
• Numbers with provenance: factors, thresholds, rates, prices, GWP values. These need a source, a retrieval date and a data vintage, and they follow the same two passes: does the source exist, and does it give this number for this scope and this year.

2. Pass 1: existence

For each work, confirm author, year, title, venue, volume, issue and page range against a publisher record, a DOI landing page, or the publisher's own catalogue. Where the publisher page is paywalled, RePEc / EconPapers / PhilPapers / PubMed / the national library catalogue are acceptable and should be named.

Use deep-research for the fan-out when the list is long. It saves each source to its own file with verbatim quotes, which is exactly the raw material pass 2 needs. Its summaries are a map of the territory; the reading is still yours.

Write REFERENCES.md (see references/templates.md for the full structure). State the date of the pass and, in plain words, what "verified" does and does not mean here: specifically that it does not verify that any work says what the document claims.

3. Pass 2: substance

Take the claims from step 1 and check each against primary text. For each:

• Locate the passage. If you cannot reach the primary text, that claim is unchecked. Record it in the unchecked list and move on; settling it from a summary is the one move the skill exists to prevent.
• A fetch succeeds when the body holds the text. EUR-Lex answered every URL form with HTTP 200 and a 13.7 KB navigation shell that carried the act number in its menu and none of the regulation text. Count a fetch as successful only when the body contains the row or sentence you are checking. Record the URL, the date and the byte count; a body that is small or the same size on every request is a failure whatever the status says. If two routes exist, record the one that produced the bytes. For EU acts the scripted route is the CELLAR resource http://publications.europa.eu/resource/celex/<CELEX> with Accept: application/xhtml+xml; OJ tables use a decimal comma (3,251), so parse them as such.
• A tool that cannot parse a PDF may still have saved it. Before recording a claim as unchecked, look for the local path in the tool's output and run a local text extractor against the file; one call rendered a statutory information sheet the fetch tool had refused, version date included. A source counts as unreachable only after the local extraction also fails, and the entry then names the methods tried.
• Quote the relevant lines verbatim with a page reference.
• Decide: does the claim hold, hold with qualification, or fail?
• If it fails, write the fix: the sentence the document should say instead, and why the corrected version is or is not weaker for the argument. Often it is stronger; say so when it is.

Write SOURCE-CHECK.md. Order matters: failures first, then qualified holds, then clean holds, then the explicit unchecked list. A reader who stops after the first screen should have seen the bad news.

4. Feed the fixes back

A check that does not change the document was a waste of a pass. Apply each fix to the source document, and where a fix removes a support the argument was leaning on, say so in that document's own limitations section; a quiet patch to the sentence hides the loss.

Then update the summary line at the top of SOURCE-CHECK.md (how many claims checked, how many held, how many failed, how many of the failures were load-bearing) so the state is legible without reading the file.

Numeric figures

Same two passes, different evidence. For an emission factor, a regulatory threshold, a carbon price or a GWP value:

• Primary source only. The implementing act, the standard, the agency's own dataset. A consultancy summary, a blog or a previous project is a pointer; follow it to the act, the standard or the dataset and cite that.
• Retrieval date and data vintage are different things and both are needed. "Retrieved 12 August 2026" says when you looked; "2024 vintage" says which edition of the number you took. A figure without the second cannot be defended when the source restates it.
• Scope has to match. GWP values differ by IPCC assessment report; emission factors differ by boundary, by region and by whether they are location- or market-based. A correct number for the wrong scope is a wrong number.
• When sources disagree, say so and pick with a stated reason. Taking the first one silently is the failure.
• Every arithmetic step owes its own check. A derived figure inherits the source and date of its inputs only when each input carries the unit the arithmetic needs. "$499 per project plus DAU-based rates" has no billing period, so multiplying it by 12 invents one; "$6,500 a year" was written from it anyway. Before any per-month or per-year arithmetic, re-read the source page for the period. If it publishes none, say the figure cannot be derived and anchor on a source that does publish its period.

If the figure is going into code, it belongs in an explicit argument with the citation next to it. A hardcoded default is a number nobody will re-check.

Carrying provenance through a tool

A citation kept next to the number in the source file is lost the moment the number is exported. Provenance survives only as a field of the artefact, so make it one:

• Name the factor set with a version string that encodes vintage and status, such as 2026.07-indicative.
• Stamp that string on every result object and every exported artefact: the workbook's summary sheet, the tool's factor page, and each line of the audit trail next to the per-figure source string.
• Repeat the per-figure source on the line that uses it, so a reader holding only the export can tell where each number came from and whether it is fileable.

Two designs at the ends of the range are defensible: ship no regulatory constants and require them as arguments, documenting why, or ship a named, versioned, explicitly marked set. The failure case sits between them: unmarked defaults that look authoritative after they go stale.

Reference

references/templates.md has the exact structure of REFERENCES.md and SOURCE-CHECK.md, with the summary-line format and the scope-statement wording. Read it before writing either file.