Intuition Labs · φ research · 08 · projected · 22 may 2026
train without a grader
crystallization moments carry the reward signal, for free.
folded into the-lab-runs-itself · a training reward that needs no separate grader did ship, but it scores the run's own verified outcome rather than the confidence number; the lab running itself carries it now.
−50% training compute
reward_t = base · 1[Δφ_t > θ] · φ_t · 1[phase = answer] · phi-anchored reward