Frontier Agentic Trajectories.
A dataset of verified, step-by-step reasoning from an agent solving ARC-AGI-3 — joined by problem + step and tied to environment-confirmed outcomes. Built for training and evaluating agentic models. The first datasets are being compiled now.
Ground-truth-graded agentic data.
Scraped text can teach style. Unverified logs can imitate process. Ground-truth-graded trajectories teach decision quality, because every action is linked to a result the environment itself confirmed.
Verified, not scraped
Outcome records are environment-graded, making the data usable for reward modeling, RL, RLAIF, and evaluation.
Genuinely agentic
Multi-step, tool-using problem-solving on a benchmark built to stress autonomous reasoning — not static chat completion.
Reasoning, not just the answer
The debate layer captures adversarial hypothesis testing and verdict formation before the chosen action.
Three linked record types.
For every problem the agent works, the dataset captures three layers joined by problem_id + step, so a buyer can follow the chain: situation → reasoning → move → result.
Moment-to-moment decision signal
Board state, candidate actions, and a learned value estimate for every possible action.
one per step
Verified state-transition grade
Ground-truth pass/fail confirmation from the environment. This is the RL-usable layer.
one per grade
Adversarial reasoning trace
Claim → counterexample → falsification test → verdict → chosen action, tied to the outcome grade.
one per hard step
The first datasets are being compiled now.
The public sample is intentionally compact and sanitized for review — it shows the schema and the linked record types, not production volume. The dataset is built to compound as the Cognitive OS solves more problems, and early partners can arrange access as it comes online.
Built for teams improving reasoning models.
Frontier labs can use the feed for supervised fine-tuning, reinforcement learning, reward modeling, capability evaluation, and failure-mode research.
SFT
Train on structured problem-solving traces rather than generic reasoning prose.
RL / RLAIF
Use verified outcome records as reward and preference signal.
Evaluations
Measure agentic reasoning and self-correction against held-out probes.
Failure analysis
Study where base systems fail and what corrections change the outcome.