The Record, Not the Model: Cross-Architecture Evidence that Near-Domain Retrieval Confabulation Is Dominated by the Served Record

Cross-architecture matrix preprint, v0.7. A controlled behavioral matrix scoring abstention as a first-class success across eight contemporary architectures from seven organizations, over a frozen, hash-sealed set of 168 private-fact probes under matched context arms and two instruction regimes. Context-free failure-to-abstain ranges 3.8% to 95.8%; an abstain-permission instruction cures context-free and irrelevant-context arms to near zero, but on authentic near-domain art-provenance records instructed failure persists at 14.6-20.8% in every architecture. The central corrected finding is stimulus dominance: on identical near-domain records, failure rates vary by 34.4 percentage points across record triads while varying negligibly across architectures, with identical triad ordering in all eight models; architecture explains approximately zero variance by four estimators. A public value-rotated release of the corpus reproduces the triad signature in all six re-run models, showing the effect follows the structure of near-domain records rather than their content. Includes a five-way dialect taxonomy of confabulation strategies, an effort economy spanning 9.5 to 371 tokens per answer, and a Method-Integrity section recording the paper's own adversarial self-corrections. Companion to the single-model pilot (DOI 10.5281/zenodo.21367214).