castform

a headless roblox studio for rl rollouts, three judge rewards mirroring lemonade's evals, and what a large moe model scored after training

train on harbor tasks with castform

sft is now first-class on castform: launch a run in under 10 lines of code, plus a pii-masking finetune of qwen3.5-4b that beats an untrained frontier model on 3,000 examples.

A 4B open-source model post-trained with Castform retrieved search results as accurately as GPT-5.6 Sol, while costing 100x less.

training a 4b model with reinforcement learning to rewrite ai-sounding text so it reads human again - using ai detectors as the reward, and what happened when the model learned to hack them.

at castform, we

A new leak claims that Pokémon Winds and Waves will give monsters weather forms, so let

how the shift to agent-driven development reshaped our onboarding: why we scrapped a config wizard for an agent-first flow, and what actually convinces a first-time user.

why llms always output 4 when asked to roll a dice, and how exploration techniques in reinforcement learning fix this.

how we used rl to train an attacker and defender in an iterative loop, what broke along the way, and what we learned about reward design for adversarial safety.