The Harness Thesis: Reliability is the next Frontier

This white paper reframes the trajectory of artificial intelligence capability, arguing that frontier progress now depends on harness sophistication rather than model scale alone. The paper introduces a structural diagnosis of why enterprise AI adoption stalls despite benchmark advances: capability scales along two independent axes—model scale (raw parameters) and harness sophistication (systems that wrap the model)—yet investment has skewed disproportionately toward the former. Through analysis of the 2025 ARC-AGI-2 results and other frontier benchmarks, the author demonstrates that sophisticated harnesses (verification loops, preference learning, and orchestration) can extract more value from frozen foundation models than additional training runs, while avoiding catastrophic forgetting and enabling modular, auditable enterprise deployment. Key Contributions: The Two Scaling Axes Framework: Distinguishes between model capability (what the generator can propose) and harness intelligence (what the system can verify), showing how the latter now offers higher ROI for enterprise reliability. The Reliability Frontier: Identifies the "accuracy ceiling" (compound probability degradation over sequential steps) and the "preference gap" (failure to capture tacit expert judgment) as the dual barriers to deployment. The Harness Architecture: Proposes a three-component system—(1) an orchestrator governing iteration and test-time compute allocation, (2) a preference learning system capturing expertise from implicit signals without modifying model weights, and (3) a verification layer employing formal methods (Lean, etc.) in closed domains and parameterized simulation in complex ones. The Convergence Thesis: Argues that robust verification of open-world claims eventually requires the same capabilities as general intelligence, suggesting that building harness infrastructure constitutes a complementary path to AGI. Structure: The paper progresses from diagnosis (Part I: Why Impact Stalls) to architecture (Part II: What Scales Next) to strategic forecasting (Part III: How Value Accumulates).