Cross-vendor heterogeneity as a guard against epistemic homogenization in AI-assisted scientific synthesis

AI-assisted synthesis is becoming consequential infrastructure for scientific research, and a new failure mode emerges when a single AI model is the sole synthesizer of a long-lived knowledge corpus: the corpus accumulates that model's blind spots and biases over time, with each individual edit plausible and internally consistent but the cumulative drift invisible without an outside reference. We call this drift epistemic homogenization and distinguish it from per-output hallucination. We argue that the right granularity of defense is cross-vendor heterogeneity, routing successive synthesis passes through frontier models trained by different companies, with different training corpora and different RLHF procedures. The pattern has a defensive value (catching errors and framing biases aligned with the dominant vendor's prior) and a generative value (surfacing connections the dominant vendor would not propose unprompted regardless of how rigorous its self-review pipeline is). We describe a production deployment as a two-layer pattern: an interactive hypothesis-development layer (author plus Anthropic Claude Opus) and an automated sweep daemon (DeepSeek V4-Pro at propagation and synthesis with Google Gemini 2.5 Pro as automatic Pass 2 fallback; OpenAI GPT-5.5 at review). The strongest evidence in the paper is reflexive: while drafting this manuscript, the cross-vendor review pass caught a magnitude error in the internal wiki documentation that the paper was citing as evidence (§5.0); the corrected figure appears in §5.3 and the catch is verifiable as a literal diff against the manuscript's revision history. Six additional operational case studies (§5.1–§5.6) extend the same pattern across approximately three weeks of operation on a ~95-page personal-research wiki on gout and engineered koji as a therapeutic-enzyme chassis, including one boundary-case limitation (§5.6: corpus-level contamination inherited identically across vendors from training data, defended by independent verification-of-verification chains rather than by cross-vendor heterogeneity). A pilot empirical study of inter-vendor agreement (§6) reports mean pairwise cross-vendor agreement of ~82% across 8 prompts × 4 vendors and a within-vendor replicate study at ~91%, partitioning the ~18% cross-vendor disagreement into roughly half within-vendor temperature noise and roughly half real cross-vendor heterogeneity (approximately 8 percentage points of real prior-distribution difference under the pilot's manual claim-coding scheme). The disagreement is concentrated in the task types the architecture is designed to exploit. The contribution is a methodology case study with N=1 operational deployment plus pilot empirical evidence; controlled ablations comparing cross-vendor against alternative review architectures remain future work.