ADR-0012: "AI summarisation" means selection and structure, never rewriting

Status: Accepted — clarifies ADR-0002, which stands Date: 2026-09-02

Decisions

D-16 Summarisation of Safety-Critical Content is extractive: the model chooses which spans matter and how to lay them out, and never rewrites their words

Selected spans are displayed verbatim with their SDS Section reference. Choosing and structuring is permitted; generating replacement prose is not.

D-17 Explaining an unfamiliar term is Conversational Content, and may be generated

The value itself stays verbatim. A model may explain what "vapour density" means; it may not restate what the vapour density is.

Context

R-27: the revised poster makes AI สรุปข้อมูลสำคัญจาก SDS ให้เข้าใจง่าย a headline differentiator against apps whose ข้อมูลยาวอ่านเข้าใจยาก. R-32 sharpens the underlying complaint — SDS data is scattered, hard to search and hard to understand. Read literally this contradicts D-02.

The contradiction is narrower than it appears, because "summarise" carries two jobs in that sentence. The user's real complaint about an SDS Section 8 is not that the words are difficult but that finding which glove requires reading forty lines. That is solved by selection and structure — pick the three spans that matter, lay them out as labelled fields, each still verbatim with its reference. Long becomes short and buried becomes immediate, with no new safety prose written. The genuinely hard-to-understand residue is a vocabulary problem, which D-02 already routes to the conversational layer.

Rejected options

• Generative summary displayed beside the verbatim span — rejected. Best readability and closest to the poster's wording, but the summary is what people would actually read, so a paraphrase error still reaches the user with the verbatim text serving as decoration.
• Generative for non-critical SDS sections only — rejected: it needs a per-section policy that someone must own and that will drift as sections are added.
• Drop the summarisation claim — rejected: it surrenders the product's main advantage over a plain SDS search, and the claim is satisfiable honestly.

This ruling may not be re-decided

If a change contradicts this ADR: stop and raise it. Do not implement over it. "We do AI summarisation" must never be read as permission to generate safety text — that reading is precisely why this ADR exists as a separate document rather than a footnote to ADR-0002.

Coverage

Upstream	Landed in	Evidence	Note
R-27	D-16	the summarisation claim satisfied extractively
R-32	D-16, D-17	"scattered/hard to find" vs "hard to understand" split across the two