Evidence before pitch.
Frontier buyers dismiss unbacked claims on sight, so proof comes first. Orion, our Cognitive OS, leaves a tamper-evident receipt for every operation it runs — so "verified" is a literal claim, not a marketing word. The public surface shows the data structure, the schema, and sanitized samples; the graded records and receipts are available to qualified partners under diligence.
The benchmark is the proving environment.
Aurevus uses ARC-AGI-3 because it provides hard, novel, multi-step agentic tasks with objective outcome grading. ARC is not the product — it is the scored forge that produces verified trajectories.
View the evidence surface.
These public files are sanitized. They demonstrate the shape of the commercial dataset: linked records, environment-graded outcomes, and a point-in-time snapshot of the trajectory pipeline.
Sanitized public preview.
Production data is provided under partner terms. Public samples are redacted to show schema and utility without turning the website into a data dump.
{"record_type":"decision","problem_id":"redacted-problem","step":14,"state_hash":"sha256:redacted","available_actions":["redacted"],"value_estimates":{"action_a":0.71,"action_b":0.18,"action_c":0.11},"selected_action":"action_a","links":{"outcome_record":"redacted-problem:14:outcome","debate_record":"redacted-problem:14:debate"}}
{"record_type":"outcome","problem_id":"redacted-problem","step":14,"state_transition":"redacted","grade":"pass","verified_by":"environment","links":{"decision_record":"redacted-problem:14:decision","debate_record":"redacted-problem:14:debate"}}
{"record_type":"debate","problem_id":"redacted-problem","step":14,"hypothesis":"redacted","counterexample":"redacted","falsification_test":"redacted","verdict":"survived","chosen_action":"action_a","linked_grade":"pass"}Need deeper evidence?
Qualified labs can request fuller sanitized samples, the dataset card, schema review, graded records, and partnership terms under appropriate diligence controls.