Sufficient Statistics
B_H values horizon progress, B_F values reliability in unforgiving environments, and C measures the value mass where both constraints bind.
Main Intuition
AI progress is not one scalar quality ladder. Tasks fail for different reasons. Some tasks require long chains of coordinated action. Other tasks can be short but unforgiving because a local error is costly or irreversible. The model turns that distinction into sufficient statistics for valuing frontier movement and firm races.
Race Inversion
The same economy can race on runtime in one task region and on base reliability in another. The regional ratio B_H^R/B_F^R determines the direction.
Data Flywheels
Deployment scale creates productive tipping only when served tasks produce usable traces that move the scarce boundary. Otherwise scale can be pure lock-in.
Task Regions
High forgiveness
Low forgiveness
Low horizon
Product recommendation, ad copy, search help. Low-cost systems and imitation can be strong.
Driving maneuvers, medical dosage, payments control. Reliability and verification dominate.
High horizon
Coding, debugging, data analysis, research assistance. Runtime, tools, retries, and workflow integration matter most.
Drug discovery, critical infrastructure, autonomous organizations. Full-stack frontier advantages can persist.
Paper
The compiled PDF, LaTeX source, and math checker in this folder are synchronized with the current revision.