Research & case studies
The research program behind Manifest: RL environments, benchmark audits, and fine-tuning runs, each written up with the methods and the numbers.
Benchception
Which model builds the better RL environment: Claude Opus 4.8 or GPT-5.5? We gave both the same plain-language spec, had each build an environment and train a student model on it, and let a held-out supply chain benchmark pick the winner.
SupChain-Bench Verified
326 benchmark items went in, 288 came out. We ran every SupChain-Bench item through its own graders, cut the ones that were wrong or could not be graded fairly, and shipped the survivors as two deployable HUD v6 environments.
Build
Snap blocks together and get a working RL environment. Build is our Scratch-style builder for HUD: describe each block in plain language, then check, run, and train. No code, no JSON, no CLI.
Classhopper Set
100 bug-fixing tasks lifted GPT OSS 120B by 13% on best@10. We fine-tuned with GRPO on the Classhopper Set and also measured +8% on pass@1, with the model solving each task in 4 fewer steps.