Verifiable agent environments

It jointly learns to generate executable Gym environments and solve them, outperforming fixed-environment self-play by +5.3 average across eight held-out benchmarks.

Unlike teacher- or label-dependent distillation, it uses clean-versus-degraded views to self-distill, lifting Qwen3.5-4B from 70.7% to 77.4% and surpassing Qwen3-VL-235B.

Targets stateful agents by evolving shared environment state, reaching 85.0% ASR across 75 setups.

Unlike scalar-only video benchmarks, it uses model-agnostic skill routing with auditable evidence trees.

Adds explicit camera-ray token coordinates to a pretrained video DiT with under 0.1% new parameters.

Unlike recent camera-control video models, it sustains long interactive rollouts with bounded memory and mid-rollout re-prompting in three CFG-free steps.

It links 1,440 de novo binders to raw two-vendor kinetics, structures, and stepwise design provenance.

Unlike typical 3D corpora, it pairs 9.3K text-conditioned meshes with four safety tiers and reusable views, SDF, point-cloud, and embedding artifacts.

Unlike recent agent datasets, each of 21,207 multi-turn traces includes deterministic verifiers and exact reward contracts.