Concept relations

Every relation on this page makes one claim and no other: these public sources address related concepts. It is not a statement that two sources are the same, that one stands in for the other, or that either settles anything the other leaves open. This record reports what documents say and when they changed; a relation here is a shared word, evidenced from each document in turn, and nothing more.

contextual mapping — not a regulatory applicability determination The policy, stated once

A concept is a controlled identifier in the topic. namespace published at /conventions, minted from a string the corpus already states. No list of concepts is written anywhere: a relation exists where two rows carry one identifier and each retained reading supplies a span containing the term that row binds it by.

What the rosters named

AI — AI (RMF · 42001 · 23894 · 42005 · FDA/EMA principles · GAMP AI)

Cyber and OT — cyber/OT (62443 · 800-82 · CSF · NIS2 · CRA · FDA cybersecurity)

Each name resolves to whatever rows carry its tokens in their own title, short title or aliases — whole-token and case-insensitive. Where a name resolves to more than one row, every matched row is admitted and the count is shown; nothing here decides which row was meant. The roster itself is a declaration and is kept beside this page at docs/referents/concept-roster.json.

2 of 12 names carry a token more than one row carries. A probe proves that a row’s own name holds the token; it cannot prove no other row does, so the count is stated and the rows are all admitted.

What may supply evidence

• 8 of 15 resolved rows hold a retained reading this record may quote from. The other 7 are listed below with the reason each was refused, because an absent relation and a refused one are different facts.
• 0 pairing(s): a pair of eligible rows taken together with one controlled concept identifier both of them carry. The figure counts pairings rather than pairs, because two rows carrying two identifiers in common stand in two relations, each with its own two spans, and a single count of pairs would hide the second. 0 of those 0 carry a span from each end; 0 are withheld, each with its reason stated beside the concept it would have joined.
• 0 of 8 eligible rows carry at least one relation, over 0 concept(s).
• By roster: 0 relation(s) join two rows the AI roster named, 0 join two the cyber and OT roster named, and 0 join rows the two rosters named separately. A roster is a grouping the order made for the purpose of looking; sharing one is never a reason for a relation, and no relation here was drawn from it.

Rows that supply no evidence, and why

The licence position is the row’s own recorded field and the rights limb is evaluated before any retained text of that row is opened, so no passage of a restricted document is read here at any point. The referent limb admits only a reading this record classifies as the document the row names: a reading of a page about a document, or of a document that superseded it, states nothing that can be attributed to the document itself.

The concepts

No two eligible rows carry a concept identifier in common, so there is nothing to relate. The address stays and the zero is the fact.

Every relation

None over the 0 pairing(s) the shared identifiers formed. That is a measurement over the whole population and not a sample of it.

How a span was selected, and what was refused

• The term must survive the cap. A window of 150 characters either side of the match is expanded to whole-word boundaries, then cut to at most 14 words centred on the matched term. A span that would not carry the term that selected it is not offered; a quote whose own reason sits outside it is evidence of nothing. Nothing is trimmed to fit, because a truncated quote is a fabricated measurement of what a publisher wrote.
• Page furniture is not evidence. A span matching a term in state/furniture-registry.json — chrome measured per host and cited to a retained snapshot — is passed over and the next occurrence is tried, and where every occurrence is furniture the relation is withheld. 35 of 35 registered terms were readable to this page.
• A digest is over normalised text, not over the bytes served. Tags, scripts, navigation and volatile strings are removed before hashing, so hashing a source as served will not reproduce the digest shown beside a passage. Where a reading is marked as prose extracted from a PDF, the words are an extraction of a page image and may run together in ways the published document does not.

What this page does not say

• Nothing about what either document requires. A shared concept identifier is a shared word with a span from each source behind it. It carries no direction, no strength, no ordering and no coverage.
• Nothing about a reader’s position. There is no lifecycle, product class or sector tagging here, and a relation is never a statement that either source reaches any particular thing.
• A withheld relation is not an absent one. Where a pair is withheld, the reason names this record’s reach or a row’s own recorded terms. It is never a measurement of what a publisher wrote.
• A restricted source is cited and linked, never excerpted. 5 of 15 resolved rows record their own terms as restricted. No retained text of those rows is opened by this page at any point, so nothing here is a measurement of what any of them says — including the absence of a relation.

Limits

• The quotation surface. Read from scripts/quotefidelity.py at build, the state of this route is declared: the quotation gate holds this route in its declared set, so the census that refuses a quotation on a surface it does not know does not refuse this page. Sitting in that declared set is what lets a passage render here; it is not itself a comparison of a rendered passage against the bytes it names. That gate keeps two lists in one file — the directories its census knows, and the surfaces it names a check for — and a route can sit in the first without sitting in the second. The directories that gate declares are r, body, concordance, graph, crawlability, crosswalk; the ones it excludes with a stated reason are papers, compare, status-change.
• One reading per source. Each end quotes the most recently first-observed retained reading that passed every limb, which is the record’s own ordering. A rule that picked the longest text, or the best-looking span, would be this record choosing its own evidence.
• One rule has two readers, and that is a gap. Whether a retained snapshot holds prose or the file itself is decided by src/lib/observations.ts per row; src/lib/conceptRelations.ts asks the same question of one reading, because a row can hold one reading of a file and a later reading of text extracted from it. The structural repair is a per-reading answer at the first home, which the second would then call.
• The furniture verdict is a port. scripts/content_region.py is where the definition lives and this page cannot call it, so the registry limb and the two shapes that are not language are restated here. The structural limb of that definition is not restated: no reading in this population carries a structure artifact, and nothing here pretends to have consulted one.

Last updated: 18 September 2026