The original Nova Triangle (May 2026) described triangulated inference in theory — three models, an agreement score, "disagreement is the signal." This is that idea tested for real, months later, with actual results.
One question goes to three genuinely different architectures at once: Transformer (autoregressive, fine-tuned, a real consistent voice) — Diffusion (LLaDA-8B, denoises the whole answer at once instead of word-by-word) — RWKV (recurrent, no attention — included on purpose even though it's the weakest of the three; watching it struggle live is more honest than hiding it). This calls real hardware over a live tunnel, not a cached response.
—
—
—