Phase 2: grounded hybrid vs direct model The full-data direct MiniMax baseline passed 8/14 with 35.720s mean latency. The initial hybrid passed 10/14. The production grounded hybrid now passes 14/14 on the unchanged suite and graders with 6.412s mean latency, including 9/9 original cases and 5/5 reasoning cases. In the grounded hybrid, code computes verified aggregates and the model interprets them, separating deterministic calculation from natural-language reasoning. Phase 3: repeated architecture benchmark Three complete runs per configuration used the unchanged 14-case suite, identical seeded 1,095-row dataset, temperature 0, a 120-second per-case deadline, and disabled tools for direct-model baselines. - Grounded hybrid: 100.0% ± 0.0 pass rate (42/42), p50 5.738s, p95 10.144s, honesty-proxy failures 0%, injection-safety failures 0%. - Direct MiniMax M2.7: 59.5% ± 3.4 (25/42), p50 32.646s, p95 73.636s, honesty-proxy failures 0%, injection-safety failures 100%, estimated list-price cost $0.2937 for 42 calls. - Direct Claude Sonnet 5: 26.2% ± 3.4 (11/42), p50 117.075s, p95 120.035s, honesty-proxy failures 33.3%, injection-safety failures 66.7%; cost unavailable because 21 timeouts lacked usage. Conclusion: deterministic code owns facts; bounded model reasoning owns interpretation. The grounded architecture generalized across repeated cross-model comparison. Boundary: the honesty proxy covers only missing-field and unsupported-projection cases; MiniMax cache reads are inferred from cold-call prompt size at official rates; one Claude tool attempt was blocked and scored as a protocol error.