ADR-0027: The Thai-SDS preference stands; its text layer is never trusted

Status: Accepted — amends ADR-0009 Date: 2026-09-02

Decisions

D-56 A Thai-published SDS remains the preferred source, and its extracted text is verified against the rendered page image

Verification is a named Curation step with its own acknowledged cost.

D-57 D-11's rationale is corrected: preferring a Thai SDS does not avoid Curation cost, it changes its shape

Translation review is replaced by transcription verification, not eliminated.

Context

R-40: the Thai spike documents' text layers are corrupt — Doc B extracts 0 correct ำ against 87 broken, Doc C 12 against 643. This is Thai SARA AM decomposition failure, and it means extracted Thai text is wrong in ways that do not look wrong on a skim: a dropped vowel changes a word silently.

The tempting conclusion is to reverse D-11 and prefer English SDS with a curated translation, since English layers extract cleanly and translation is already a controlled, reviewed step (D-11, D-12). Corrupted Thai is by contrast an unreviewed corruption nobody may notice.

That reasoning is rejected on principle: a tooling defect must not decide a content-authority question. The Thai SDS is the document The Plant holds and that Thai law requires for that product. An English sheet from a different supplier is not a like-for-like substitute — R-39 demonstrates that two suppliers genuinely disagree, so swapping documents to obtain cleaner bytes trades a solvable extraction problem for an unsolvable provenance one.

Rejected options

• Reverse D-11, prefer clean English plus curated translation — cleaner pipeline, rejected per the principle above.
• Per-document, whichever extracts cleanly — pragmatic, rejected because Corpus provenance would vary for no principled reason.
• Invest in correct Thai text normalisation — the best long-term answer and it would remove the manual cost, rejected for prototype scope (D-01) as real engineering that may not fully succeed. A good candidate for production work.

This ruling may not be re-decided

If a change contradicts this ADR: stop and raise it. Do not implement over it. Never admit extracted Thai text to the Corpus without verification against the page image.

Coverage

Upstream	Landed in	Evidence	Note
R-40	D-56, D-57	the character counts are quoted in Context