ADR-0015: Extractive selection must never drop a restriction

Status: Superseded by ADR-0022 — premise upheld, mechanism replaced Date: 2026-09-02

Decisions

D-26 Omission is a safety failure mode of selection, and is bounded explicitly

Within any SDS Section a selection draws from, every span expressing a restriction, exclusion, incompatibility or negation must be carried through whenever any span from that section is shown. Restrictions are never optional detail.

Context

ADR-0012 granted the model the power to choose which spans are displayed. Every rejected option there, and its whole Context, addresses paraphrase. None addresses omission.

The failure this leaves open is worse than the one it prevents, because it is invisible. A model selecting three verbatim spans from Section 8 and silently dropping a fourth reading "not compatible with nitrile" produces output that is 100% verbatim, fully cited, correctly sourced — and lethally wrong. Every safeguard in ADR-0002 and ADR-0012 passes. The user sees a citation and trusts it.

This gap was found by an adversarial review of ADR-0012, not by writing it. It is a defect D-16 introduced; ADR-0012 is otherwise sound and stands.

Rejected options

• Show the whole SDS section, never select — the safest possible answer, and it is exactly the "long, hard to read" problem R-32 and R-27 exist to solve. Rejected because it reverts the product to the thing it defines itself against.
• Trust the model to include restrictions — rejected: it is the same class of trust that ADR-0002 already declined to extend to paraphrase, and the failure is silent.
• Flag sections containing restrictions and show them whole — a reasonable middle path, rejected for now as more mechanism than a prototype needs when the narrower rule above covers the same cases.

This ruling may not be re-decided

If a change contradicts this ADR: stop and raise it. Do not implement over it. Do not drop a restriction span because it makes a card too long, and do not treat "the selection is all verbatim" as evidence that the selection is safe.

Consequences

Selection needs a restriction detector, and it needs to fail loudly rather than quietly — if it cannot tell whether a span is a restriction, the section is shown whole. Every selection is testable against this rule, which makes it a natural verification gate later.

Coverage

Upstream	Landed in	Evidence	Note
R-12	D-26	the nitrile-incompatibility case is worked in Context