The corpus was authored after the artefact was frozen, from situations external to the design, and deliberately not assembled from the defects already recorded in the specification. Those remain conformance regressions.
Each case records where its situation comes from, the act proposed and what follows if it is wrong, what was knowable when the act was decided, and what the domain requires of the evidence, stated without any of the artefact's vocabulary. That last constraint is what keeps a failure of the demand language distinguishable from a failure of adjudication when the results are read.
Ten families, and what each varies
The families follow what the discipline claims to distinguish: consequence, freshness, scope, authority, re-establishment, corroboration, completeness, field provenance, semantic operation and ancestry. Stratum is a property of a case, not of a family, but the families lean: in seven the cases are predominantly grounded at stratum 1, and in three at stratum 2.
The stratum of each case was fixed before authoring, so counts could not drift toward families that happened to perform well. Where authoring could not meet the allocation honestly the allocation was amended and the reason recorded, rather than the corpus being padded. One such amendment matters: the protocol had required at least one expected failure per family, and in four families the sources state their requirements plainly enough that no honest prediction of failure could be made. Declaring one anyway would have been a quota-driven fiction, so the requirement became a corpus-level minimum. Five expected failures stand, all declared before execution.
Atoms and expressions
Each requirement was decomposed into atoms before any demand existed, and the decomposition is the part that makes the later fidelity check meaningful. Had the atoms been written during translation, a translator with the demand vocabulary in view would have atomised the prose around what the language can express, and completeness would have followed by construction rather than being tested.
The thirty-nine requirements yield 108 atoms: 72 conjuncts, 24 negatives, 4 temporal qualifiers, 4 thresholds, 3 alternatives and 1 exception.
The 24 negatives are the reason the decomposition earns its cost. They state non-entailments: a closed indication is not a proof, being separate devices is not a finding of independence, a model's account of its own working is not a check, an undeterminable age is neither recent nor stale. These are what a permissive translation drops silently while satisfying every other clause, and they turn out to be where the discipline's own argument lives.
A flat list of atoms would still lose the topology between them, so each case also carries a frozen expression over its atom identifiers. Without it a translator could map alternatives as conjunctions, turn an exception into a further requirement, or detach a threshold from the quantity it constrains, while accounting for every atom.
One case makes this concrete. HSG253 requires an isolation to be re-proved if it has been in place longer than a shift or has been left unattended. Those are alternative triggers; read as conjuncts they describe a materially weaker rule, and a demand encoding only the age would admit an isolation left unattended for ten minutes. Writing the expressions also exposed a defect in the atomisation itself, where a threshold had been bundled into the obligation it triggers, leaving nothing for it to be scoped over.
Matched pairs
Nine pairs were authored, in two forms: holding the basis constant and varying the act, or holding the act constant and varying exactly one load-bearing property of the basis. A pair whose two members receive the same verdict is a finding about the discipline's resolution rather than a case to be rewritten until it separates.
Three cases carry contested reference warrants, where the domain disagreement is genuine and recorded rather than resolved by the author. They carry a predicted verdict and no correct verdict, and are excluded from error metrics: a contested warrant provides no settled reference against which a verdict could be wrong.