Development Investigations
Every step of the investigations behind the training example, run in the digital twin before and after training, in the agent’s own words.
Each case is one development lot investigated twice in the digital twin by DeepSeek V4.1 Flash, once from the starting checkpoint and once after twelve reinforcement-learning updates on whole investigations. Both runs of a case share the lot, its batch noise and the allowance. The cases are every matched run on the degassing-decides development lots where exactly one of the two agents reached a correct decision, 12 where only the trained agent did and 2 where only the starting checkpoint did. Case 1 is the example in the release.
Each step shows what the agent did and the results that came back from the twin, with its full reply one click away. The agent starts from the task contract described on the task and environment page.
Correct only after training
Correct only before training
Lot 1 · run 2
Loading the investigation
What these records include
Only what the agent saw and wrote. Each step holds the model’s reasoning, its message, each tool call and the simulated results it received. Lot identities, seeds and evaluator fields are removed, and lots are numbered in order of appearance, so a lot run twice appears in two cases. These are development lots on the lean interface used for the training evaluation, and they informed development. The 29 SDL-2 test worlds stay sealed and are not included. The same records are available as a dataset on Hugging Face.