Sep 24, 2026145 of 783— I measured whether Jev, TypeSafe's decision model, gives the same yes or no twice on the same request near its thresholds. Roughly one near-threshold decision in five flipped at 0.50 and 0.60 while the headline rate barely moved.
Sep 15, 202645.0% and 44.0%— I ran the same agent benchmark twice with the same seeds. The two scores were 45.0% and 44.0%. Compared task by task, 44 of the 200 matched pairs came out differently, and all 200 took a different path.
Sep 9, 2026Where a 14-DoF Biped Falls— Frozen checkpoints of a 14-DoF walking policy, measured under lateral pushes with Proofload. The first answer was wrong, the map showed a cliff and a slope, and a retrain meant to widen the boundary made it worse in 14 of 14 cells.
Aug 22, 2026Price the Reward Function— I measured what a code-execution reward costs a GRPO training loop in GPU-seconds. The median service time turns out to be the wrong number to look at.
Aug 19, 2026Map the Failure Boundary— I froze a Go1 locomotion policy and measured where it fails across 400 conditions of floor friction and lateral push, then retrained on the two gaps the map exposed and measured again.
May 22, 2026A Decorative Channel and a Reward Exploit in One of Our Benchmarks— Adversarial training exposed two failure modes in a MARL benchmark: a communication channel that looked load-bearing but wasn't, and a reward function that incentivized trivial policies.
Jan 16, 202610,924x: The Instability Bomb at 1.7B Scale— Part 2 of the mHC reproduction series. I scaled from 10M to 1.7B parameters and watched Hyper-Connections hit 10,924x signal amplification. That's not a typo.
Jan 11, 2026DeepSeek's mHC: When Residual Connections Explode— DeepSeek replaced standard residuals with massive Hyper-Connections, and then watched them explode. I reproduced the 9x signal amplification to understand their fix: the Sinkhorn-Knopp algorithm. Part 1 of 2.