ARC-AGI-3

A dated reading of what is claimed, reported, and independently verified in the current evidence.

Current reading

Readiness band and full integer triple

AI-assisted assembly · derived results

Readiness band

Watch

Primary summary from verified readiness

Evidence strength Critical

Decision

Agent-assisted assessment · independently reviewed · 2026-10-02

Use ARC-AGI-3 for controlled agent research when the task set, harness and scoring version are fixed.

Why: Repeated experiments demonstrate practical research utility. Harness sensitivity, public-set contamination risks and incomplete technical validation limit broader conclusions.
Next: Record model and harness versions, repeat trials, distinguish public from held-out results, and add evaluations drawn from the intended deployment tasks.

Constraints

Standard and Provider Adapter results require separate interpretation.
Public-game saturation does not establish performance on unfamiliar held-out tasks.
The audit found discrepancies in pinned public implementations; their relevance to current official scoring remains unresolved.
Closed, deterministic games provide limited evidence about real-world deployment reliability.
High-performing evaluations can require substantial inference expenditure.

Evidence summary

AI-assisted assembly

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Revision 4