Atomic Skills

Distinguishing between isolated capabilities and integrated professional competency.

cognitive_depth_matrix.sh

Level 01

ATOMIC

Single-step, isolated prediction. No dependencies.

"Is this hearsay?"

Level 02

COMPOSITE

Multi-step reasoning with limited context linkage.

"Extract holding -> Apply"

Level 03

AGENTIC

Autonomous workflow across 10+ dependent steps.

"Complete Research Chain"

The Reasoning Cliff

The LegalChain framework categorizes AI legal capabilities across three levels of cognitive depth. This taxonomy ensures that performance measured at simple levels (e.g., "finding a case") is not mistaken for performance at professional levels (e.g., "completing a research memo").

Research into AI agents has identified a "Reasoning Cliff": models that perform exceptionally well at atomic tasks often experience 10-15% accuracy drops when those same skills are embedded in agentic workflows.

Hidden Failure Modes

By testing at Agentic depth, we reveal reasoning collapse and error propagation that invisible to standard atomic testing.

Why Depth Matters for Law

In legal practice, an "Atomic" success (extracting the right quote) is worthless if the "Agentic" workflow fails (e.g., the case was overruled, or the citation was fabricated). Professional legal work is inherently agentic; therefore, its evaluation must be as well.