ADR-0018: Corrections found by adversarial review of stages 0 and 1

Status: Accepted Date: 2026-09-02

Decisions

D-27 A prior inference later confirmed by the source is recorded as confirmed, and its ADR stands unchanged

R-21, R-28, R-30 and R-34 each confirm a decision previously reached by argument (D-01, D-06, D-10, D-13). Confirmation is not disagreement, so D-20 ("where the two posters disagree, poster.pdf wins") does not cover them — they land here. R-31 lands here too: the screen-7 gap is reproduced in both revisions rather than resolved, which is evidence about the source's production rather than about the product.

D-33 Five ADRs cite a deleted objective; their rulings stand and their rationales are void in that clause only

R-24 removed "study the benefits of using AI" (formerly R-04.2). It is load-bearing in the stated reasoning of ADR-0001, ADR-0002, ADR-0003, ADR-0006 and ADR-0009. ADR-0014 named only ADR-0001. Every one of those five conclusions survives on other evidence; no ruling changes. The affected clauses are simply no longer support.

D-34 ADR-0005's rationale rests on two retracted premises and is restated here

Its Context cites R-03's four settings (superseded by R-22/R-23) and D-04's curated corpus (superseded by D-14). The offline-first ruling stands, now on R-22 (one plant, whose floor and store are the setting) and D-14 (The Plant's binder, still a few megabytes on device).

D-35 D-05's "chemical name" fallback means product name

ADR-0013 said "ADR-0004's preference order is unchanged", which left the wrong noun standing: a Formulated Product has no chemical name, and per D-18 those are the common case. The order is unchanged; the term is corrected.

Context

A fork-based adversarial review of both research notes, ADRs 0001-0014 and CONTEXT.md returned twelve defects the coverage gate cannot see. The gate proves an id appears in some downstream Coverage table; it cannot tell whether the mapping is truthful, whether a superseded document still claims to be binding, or whether a rationale has been hollowed out by a later finding.

Eight defects were bookkeeping and are fixed in place — ADR-0003's status line, four nominal Coverage rows in ADR-0014, orphaned rows for R-06/R-11/R-16/R-17, and two glossary items (a missing Supervisor entry, and Corpus using "case-study plant" where the canonical term is The Plant). The four judgment defects are the decisions above. A ninth, the most serious, was a genuine safety gap and has its own document: ADR-0015.

Recorded because the review was a fork, sharing the Proposer's priors — weaker evidence than a cross-model grill. It still found twelve defects, four of which no automated check could have caught. That is an argument for doing the cross-model pass at the spec, not for treating this one as sufficient.

This ruling may not be re-decided

If a change contradicts this ADR: stop and raise it. Do not implement over it. Do not "tidy" a superseded ADR by deleting it — its ids are referenced by living documents.

Coverage

Upstream	Landed in	Evidence	Note
R-24	D-33	the five ADRs citing the deleted objective are enumerated	corrects ADR-0014's incomplete bookkeeping