ADR-0003: The corpus is curated manufacturer-published SDS PDFs

Status: Superseded by ADR-0011 — do not follow this ruling Date: 2026-09-02

Decisions

D-04 The SDS corpus is 20-40 manufacturer-published PDFs, curated, with per-document provenance

Each document is stored with its manufacturer, revision date and retrieval URL, and is parsed against the GHS 16-section structure that regulation already obliges manufacturers to publish. The set covers the five chemicals named in the source (R-16) plus common Thai industrial reagents.

Context

R-06: every capability in the source reads from an "SDS database", but the source never says who supplies it or vouches for it. Under D-02 every safety string must cite a source span, so the corpus is the product's entire authority — it cannot be left implicit.

Rejected options

• Hand-entered records from public sources — rejected because there would be no parsing to study, weakening study objective R-04.2 and hollowing out the production core required by D-01.
• Licensed SDS feed (Verisk 3E, Chemwatch, SiteHawk) — the only route with real coverage and update guarantees, rejected as out of reach: it needs a contract and budget this effort does not have.
• Thai regulatory source (Department of Industrial Works, Hazardous Substance Act lists) — best jurisdictional fit, rejected for now because what these publish and in what format is unknown. Revisit if investigation shows a usable machine-readable inventory; that would be a new ADR superseding this one, not an edit.

This ruling may not be re-decided

If a change contradicts this ADR: stop and raise it. Do not implement over it. Do not substitute hand-typed data for a parse because a PDF is awkward — that is the rejected option, and it silently removes the thing being studied.

Consequences

Two gaps must be declared openly in the spec and the report: no automatic updates (a revised SDS is not noticed until someone re-curates) and no coverage guarantee (a chemical outside the set has no answer, and by D-03 the app must say so).

Coverage

Upstream	Landed in	Evidence	Note
R-06	D-04	provenance gap named in Context
R-16	D-04	the five named chemicals seed the set