Poisson Labs Log in Contact

Poisson Labs · Research

Published with what the run could not detect.

The studies, in the order we would read them. Each one that reports a change says how it was counted: seeds shared between runs, a correction across every condition, and the smallest effect the run could have seen. The methods are one page.

  1. 2026.08
    Map the Failure Boundary A Go1 locomotion policy across 400 conditions of floor friction and lateral push. Retrained on the two gaps the map exposed: 60 conditions significantly safer, none worse. Go1locomotion read reportreceiptcode
  2. 2026.09
    145 of 783 Jev (jev-1.13.0), sent the identical request twice on items chosen to sit beside a cut: 159 of 858 near-cut decisions flipped at 0.50, 145 of 783 at 0.60 and 48 of 900 at 0.90, while the action rates moved a point or two and the flips nearly cancelled. Jevdecision models read reportreceiptcode and data
  3. 2026.09
    45.0% and 44.0% Two runs of τ²-bench airline with identical config and seeds scored 45.0% and 44.0%, while 44 of 200 task outcomes flipped and every pair took a different path. Divergence does not accumulate with trajectory length, it varies by task, and matching seeds did not stop it. τ²-benchagent evals read
  4. 2026.09
    Where a 14-DoF Biped Falls Two checkpoints of a Microduck walking policy under a lateral push: a cliff and a slope, with a crossover between 0.600 and 0.625 m/s. A retrain meant to widen the boundary was worse in 14 of 14 cells. Microducklocomotion read reportreceiptretrain report
  5. 2026.05
    A Decorative Channel and a Reward Exploit in One of Our Benchmarks A communication channel in a MARL benchmark that looked load-bearing and was not, and a reward function that let trivial policies win. MARLcommunication read code
  6. 2026.05
    Where V-JEPA 2.1's Dense Features Hold Up (and Where They Don't) Across four model sizes, dense features predict failure under temporal corruption and not under image noise, and the 2B model is less robust than the 1B on three of five perturbations. V-JEPA 2.1representation read codecable-insertion code
  7. 2026.01
    10,924x: The Instability Bomb at 1.7B Scale Scaled the reproduction from 10M to 1.7B parameters: unconstrained Hyper-Connections reached 10,924× signal amplification; the Sinkhorn-constrained version held at 1.0. mHCscaling read codedatastress data
  8. 2026.01
    DeepSeek's mHC: When Residual Connections Explode Reproduced the 9× signal amplification in DeepSeek's Hyper-Connections at 10M parameters, and the fix: Sinkhorn-Knopp normalisation of the mixing matrices. mHCresiduals read code