Retrieval-augmented "context layers" -- curated documentation an analytics agent fetches at query time -- produce large accuracy gains on text-to-SQL benchmarks. A with/without comparison cannot say which part of the layer does the work: the semantic content, the retrieval scaffolding that delivers it, or the pre-computed views that usually accompany it.

We report a four-arm ablation on DABStep across four model families. The fourth arm empties every field of prose in a frozen data contract while holding the tool surface, retrieval instruction, table allow-list and operation rules byte-for-byte fixed. Contract-delivered semantics raise hard-task accuracy from 13.9% to 55.1%, 22.6% to 56.6%, 22.9% to 68.4% and 37.0% to 77.4%, beating the same knowledge pasted verbatim into the prompt on every model (p<=0.0067, paired McNemar); on one model the pasted manual is indistinguishable from a bare schema. Content dominates scaffolding everywhere, but scaffolding is not zero: not significant on the two flash-tier models, worth +14 to +15 points on the two frontier-class ones. A null across two adjacent models is weak evidence of absence.

The contract also compiles: its declared SQL expressions reproduce gold on all 176 tasks they cover, so each arm's accuracy reads as efficiency against a proven ceiling. The contract arm recovers 61--95% of what it was given, with the shortfall smallest on the frontier-class models; on tasks that compose those semantics further, capability buys much less. Ungoverned arms submitted 166 mutating statements and corrupted the warehouse twice; governed arms attempted none. The gain is confined to the domain the contract describes. Golds are reconstructed from leaderboard artifacts; a submission audits them: 95.3% concordance, a one-sided 5.7-point leniency, contrasts moved under 3 points. Harness, artifact, 6,416 transcripts and raw results are released.
