GalacticGazer_

was messing with the OpenAI base URL in Cursor and caught this accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast so composer 2 is just Kimi K2.5 with RL at least rename the model ID

We reasoning-architected Mercury-2: 80% accuracy at 503ms TTFB. Two-phase variant: 62% accuracy, 376ms latency. The only variable: whether domain reasoning is pre-structured before generation begins. Raw Mercury-2 scores 10%. Architected scores 80%. +70pp from architecture alone. On the emotional intelligence test (10 episodes), both raw models score 0% regardless of model size. The failure is not lack of emotional awareness. Both open with empathetic language. The failure compounds: they pivot to health content on life moments (substantive failure) and produce 30+ sentence responses where 2 to 3 are appropriate (structural failure). The raw system prompt ("You are a health coach. Answer the patient.") provides no instruction to withhold advice. The 0% EQ reflects the absence of constraint architecture. A patient says "my dad passed away last week." The raw model pivots to dry health content and actionable suggestions. In healthcare coaching, that is a clinical misstep, not a communication style. The correct response is acknowledgment as is taught. The architected version produces: "I can only imagine how hard this must be. I'm here if you want to talk." It scores 90% on the same EQ test. This is emergent: the architecture detects the emotional signal through a single focus directive and sequences the concerns across turns. Empathy first communication, health coaching when the patient signals readiness. While it does not drop health coaching role.