nx_hybrid_moe_cascade_gate.nx

nx_hybrid_moe_cascade_gate.nx -- THE SYSTEM-LEVEL MoE the operator actually wants (2026-07-15): NOT a monster model, but a two-expert CASCADE that spends the FEWEST tokens. Expert 1 = the local sovereign no-float model (free, deterministic, and VERIFIABLE by the LCF kernel). Expert 2 = the lightest Claude (Haiku) -- invoked ONLY for the residual the local model can't self-certify. The router: task -> LOCAL PROPOSE (no-float model) -> SOVEREIGN VERIFY (af_decide, kernel-checked) -> certified => RETURN local answer, 0 Claude tokens (the local expert handled it, PROVABLY correct) refuted => ESCALATE to Haiku (local was WRONG -- caught, never shipped) unsupported=> ESCALATE to Haiku (out of local scope) The KEY SOTA property = SOUND token-minimization: a local answer is returned ONLY when kernel-certified, so it is NEVER wrong; every uncertain case escalates to the cheapest tier. Better local model => more certified => fewer escalations => fewer tokens (this is why the MoE/MLA arch work pays off HERE). Teeth: T1 SOUNDNESS: every LOCAL-routed answer is kernel-certified (0 wrong answers shipped locally) T2 ROUTING CORRECT: verifiable-and-right => LOCAL; wrong-or-out-of-scope => ESCALATE T3 TOKEN SAVINGS MEASURED: zero-Claude-token fraction reported; escalations go to the LIGHTEST tier T4 ESCALATION CONTRACT: each escalation emits a real Haiku-bound prompt (wired, not hand-waved) T5 DETERMINISM: the whole routing decision is byte-stable on repeat Requires /home/elderwesto/nx_stage/nx_real_model.gguf. Heavy (~1 min: local model + per-task gen). Return from main. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL