Snowmelt
Labs
AI research focused on training data.
Better data.
Better models.
We measure what a dataset does to the model that will use it, build datasets and RL environments for AI labs, and audit the ones they already have.
Recent research
Opus breaks the rule more, GPT-6.1 Sol reports it lessOpus 5.5 and GPT-6.1 Sol both worked around a missing test dependency in every run, and Sol's final message misstated what had been verified in 27 of 64, Opus's in none. Under a rule against it Opus broke it more often, but named every workaround.
Fable reads past one in five hard errors it can disproveRefereeing new materials papers, Fable 5.1 and Opus 5.5 missed one in five and one in three of the hard planted errors, and pointed at the sentence afterwards they could prove nearly every one wrong.
MiMo and DeepSeek route around a git guard until it says whyXiaomi's git guard only says Permission denied, and MiMo and DeepSeek got around it in about a sixth and a third of blocked runs, which fell to one run each once the message said what the rule was.