Articles

The Inference-Time Cognitive Configuration Series

An investigation into why frontier AI models produce mediocre output by default, the systematic failure modes responsible, and the interaction design framework that closes the gap.

0118 min

Why Frontier AI Models Are Architecturally Underutilized

The gap between what models can do and what they actually do in default operation, and why that gap is a design problem, not a scaling problem.

0224 min

The Eight Failure Modes of Default AI Reasoning

A diagnostic taxonomy of the systematic reasoning failures that emerge when frontier models operate without cognitive configuration.

0320 min

Cognitive Leverage: How Semantic Architecture Outperforms Brute-Force Compute

The evidence that interaction design produces measurably superior outputs to model scaling, and what that means for the future of AI deployment.

Mechanism and foundation

0417 min

Why AI So Often Gives You the Wrong Kind of Smart

Frontier models default to autopilot, pulling responses toward familiar attractors instead of deploying their deeper reasoning. The fix is changing the inference regime, not the question.

0520 min

The Model Knows Whether It Knows

A May 2026 paper from MIT and UT Austin measured the latent geometric substrate that meta-cognitive priors have been operating over and showed it predicts hallucination at AUROC 0.993. Output entropy scores 0.968 and degrades with scale. The mathematical foundation chapter of the series.

0618 min

The Map Is Not the Territory, But It Still Changes the Route

Behaviorally framed meta-cognitive priors outperform mechanistically precise instructions across three model families. The empirical evidence shifts the debate from philosophy to results.

Part of the Inference-Time Cognitive Configuration Series.

Standalone essays

Essay26 min

The Missing Copilot: Why Every Claude Code User Needs a Claude Chat Instance Running Alongside It

Cognitive stacking, the practice of running a second, more capable model at deliberate cognitive distance from the one doing the work. The four classes of error a second AI catches that the executing agent is structurally blind to, with real examples.