Citation discipline

Spec: ISO/IEC/IEEE 26514 Spec: ISO 24495-1 Evidence: Editorial

At a glance

This is the page the rest of Insider_ points at when it says “evidence level”. It explains why these docs paraphrase standards instead of quoting them, how a claim ties to a reference ID you can re-verify, and what each of the eight evidence levels does and does not promise.

It is written for a senior engineer who, before trusting a claim, wants to know the rules used to make it — and is right to want that.

Why this matters

Every other Insider_ page makes claims and stamps them with an evidence level. That stamp is only worth something if the system behind it is explicit. If “standard-backed” could mean anything from “I read the spec carefully” to “I remember roughly what it said”, the badge is decoration.

There is also a harder constraint. Many of the documents NextPDF answers to — the ISO, ETSI, and similar specifications — are licensed. Reproducing their text, at any length, is not permitted. The discipline therefore has to solve two problems at once: make a claim checkable against its source, without reproducing that source. The answer to both is the same mechanism, and this page is its specification.

The short version

Insider_ paraphrases standards and never quotes licensed ones. A claim cites the standard, the clause, and a reference ID; it does not reproduce the standard’s words.
Paraphrase is not a workaround; it is a comprehension test. Restating a requirement in NextPDF’s own voice forces the writer to understand it, and keeps terminology consistent with the glossary Spec: ISO/IEC/IEEE 26514, §8 .
Every standard-backed claim carries a 64-character reference_id in the page’s citations front-matter, so the next reviewer can re-retrieve the exact source span and confirm the paraphrase.
The badge declares an evidence level — one of eight kinds — that states the kind of proof, so an over-claim is visible at a glance.
When the source cannot be retrieved, the claim is not faked. It is kept, marked unresolved against a re-pin ledger, and the page stays unpublished — a documented protocol, not an improvisation.

How NextPDF approaches it

Paraphrase, not quotation

The strictest rule in the NextPDF style hierarchy overrides every upstream guide: no verbatim text from a licensed standards body, regardless of how short the excerpt is. Instead, a page cites the standard and clause, paraphrases the requirement in its own voice, and records the source reference ID.

This is usually framed as a licensing constraint, and it is one. The more useful framing is editorial. A verbatim quote proves only that you can copy. A faithful paraphrase shows that you understood the clause well enough to restate it without changing its meaning. It also lets the sentence stay in NextPDF’s consistent vocabulary rather than switching register mid-page, which the documentation-quality model requires Spec: ISO/IEC/IEEE 26514, §8 . Plain language is judged by whether a reader can find, understand, and use the content, not by whether the prose mirrors the source Spec: ISO 24495-1, §Introduction ; paraphrase serves that, quotation does not.

A claim is tied to a reference ID, not a memory

The mechanism that makes a paraphrase checkable is the reference ID. Every standard-backed claim carries a full 64-character reference_id in the page’s citations front-matter, identifying the exact source span the paraphrase was made from. A reviewer does not have to trust the writer’s memory; they re-retrieve that span and compare. The reference ID is the join between an unquotable source and a verifiable claim. It carries where the proof is without carrying the proof’s text.

The evidence level says what kind of proof this is

A citation answers “where from”. The evidence level answers “what kind”. Every Insider_ page declares one, shown in the badge row, drawn from a fixed set of eight. The order of trust ranks code and tests above runtime, runtime above metadata, and metadata above prose; an editorial page does not pretend to be code-backed.

Evidence level	What it promises	What it does not promise
Evidence: Code-backed	The claim is checked against the engine’s source or a runnable example	That a standard mandates it
Evidence: Standard-backed	The claim is anchored to a cited, paraphrased clause with a reference ID	That the code currently implements it without exception
Evidence: Test-backed	A test in the suite holds the behaviour in place	A performance figure
Evidence: Benchmark-backed	A measurement under a stated method supports the figure	The same figure on your hardware
Evidence: Artifact-backed	A produced artefact (a build output, a report) demonstrates it	A standards mandate
Evidence: Design principle	A deliberate, argued design decision	An empirical measurement
Evidence: Editorial	A reasoned explanation that organises other evidence	A new behavioural guarantee of its own
Evidence: Mixed evidence	The page blends bases and says which, per claim	A single clean basis

This page is Evidence: Editorial : it asserts no engine behaviour of its own. It explains the system the other pages’ badges rely on. That is the honest level for it, and saying so is the discipline applied to itself.

When the source cannot be retrieved

Retrieval is not always available. The discipline’s integrity shows in what happens then. A normal queue, rate limit, or budget pause is not an outage. The writer waits. A genuine outage triggers a documented fallback: the claim is kept, attached to in-repo evidence and the standards reference the code itself declares, explicitly marked unresolved against a standing re-pin ledger, and the page stays unpublished until the reference ID is really pinned.

The forbidden moves are enumerated and checkable: a fabricated reference ID, a value padded or prefixed to look real, a clause written from memory and dressed as a retrieved citation, or silently deleting the claim to dodge the citation. A correctly-marked outage on a draft is debt with a ledger entry, not a defect. An offline, deterministic check enforces exactly that distinction.

What the evidence says

This page is Evidence: Editorial , so its own basis is the documentation governance in this repository plus the standards that justify the discipline. The claims here are verifiable two ways: by reading the in-repo rules, and by re-retrieving the cited clauses.

Claim	Basis	Anchor
Paraphrase forces understanding	Standard	Plain language judged by reader use, not source mirroring Spec: ISO 24495-1, §Introduction
Terminology stays consistent	Standard	Each term used consistently throughout Spec: ISO/IEC/IEEE 26514, §8
Review is part of the process	Standard	Verification and validation testing inside development Spec: ISO/IEC/IEEE 26513, §Foreword
Evidence level is recognisable first	Standard	Recognition over recall Spec: ISO 9241-110, §5.6.2
No verbatim licensed text; reference-pinned	In-repo	`docs/style/nextpdf-overrides.md` §5; `composer docs:jaccard-fingerprint`
Outage fallback is a real protocol	In-repo	The normative RAG fallback governance + the offline `docs:rag-fallback-check` gate

Practical example

The discipline is concrete: it is the structure of a page’s citations front-matter. Each entry is a claim’s receipt.

citations:
  - spec: "ISO 32000-2"
    clause: "§6"
    # full 64-char reference ID — the join to the source span, re-verifiable
    reference_id: "<64-hex digest>"
    # NextPDF-worded topic — the paraphrase, never the standard's text
    topic: "A writer's created or amended PDF elements must conform and stay consistent"

There is no quote field, by design. The topic is NextPDF’s own restatement. The reference_id is how a reviewer gets back to the exact source to check that restatement. The receipt carries the pointer to the proof, not the proof’s words.

Common misconception

The trap is reading “paraphrase, don’t quote” as a hedge — a way to sound authoritative without committing. It is the reverse. A quotation commits nothing; it borrows someone else’s words. A cited paraphrase commits the writer to a restatement that a reviewer can falsify against the source. The discipline makes claims more accountable, not less.

The second trap is treating “editorial” as a weaker grade of “standard-backed”. It is not a grade at all; it is a different kind. An editorial page like this one organises and explains other evidence. It is correctly labelled. And the label is the point: the system works because a page tells you what kind of proof it is offering before you decide how much weight to give it.

Limits and boundaries

This page specifies the citation discipline; it is not the style sheet, the fallback governance, or the gate code. The authoritative artefacts are in-repo (docs/style/nextpdf-overrides.md §5, the normative RAG-fallback governance, the composer.jsondocs:* scripts) and take precedence over any summary here if they diverge. It asserts no engine behaviour.

The discipline binds the claim, not the reader’s conclusion. A faithfully cited paraphrase tells you what a clause requires. Whether NextPDF’s reading is the one your obligation needs is still your decision, which is why behavioural pages also carry code- or test-backed evidence, not standard-backed alone. Enforcement is partial by honest admission: the offline fallback check is live, and the verbatim-quote and live-citation verifiers are wired with their exhaustive runners still being completed — stated as in-progress, not as done.

Documentation as a product — the wider quality discipline this citation system is part of.
The standards landscape — the standards these citations point at, and how a clause becomes documented behaviour.
The NextPDF testing pyramid — what test-backed evidence means when a page claims that level instead of this one.

Glossary

Citation discipline — the rule set governing how an Insider_ claim is tied to its source: paraphrase, cite the clause, pin a reference ID, never quote a licensed standard.
Paraphrase — a restatement of a requirement in NextPDF’s own, glossary-consistent voice; the comprehension test that replaces quotation.
reference_id — the full 64-character integrity pointer for the exact source span a paraphrase was made from, recorded so a reviewer can re-retrieve and verify it.
Evidence level — the declared kind of proof behind a page’s claims, one of eight (code-, standard-, test-, benchmark-, artifact-backed, design-principle, editorial, mixed), shown in the badge row.
Unresolved citation — a claim whose reference ID could not be pinned due to a genuine retrieval outage; kept, marked against a re-pin ledger, and held back from publication rather than faked.