What changed
Tycho records observations and actions, lets an agent construct a programmatic world model, tests it against observed transitions and decides when to repair or bypass it. The public repository includes prompts, planners, tests, sandboxing and replay tools.
Why it matters
Agents can reason more reliably in unfamiliar environments when assumptions become inspectable programs that can fail tests instead of remaining hidden in prose.
What to do
Adapt the pattern for unfamiliar APIs or business processes: record observations, express a typed hypothesis, test predicted transitions, revise failures and only then act.
Caveat
Results cover 25 public ARC-AGI-3 games with one matched run per policy. Scores are model-specific, stochastic and potentially expensive outside that benchmark.