Local Optimization

Local optimization (local optimisation) is rational behaviour at team, lab, or service scope that damages system-wide truth obligations: each unit maximises its own velocity, citation count, benchmark score, or sprint goal while externalising reproducibility, lineage, and cross-team consistency.

It is distinct from incompetence. The PhD student ships results before the Dockerfile exists because the lab’s metric is papers accepted, not rerunnable artifacts. The platform team publishes a score API because downstream teams need a number today, not a provenance graph tomorrow. Each choice is locally optimal; the global causal chain ruptures.

Mechanism

Local metric ↑  →  global provenance ↓  →  reproduction heroic  →  refutation cannot propagate

Metric substitution is local optimization with a named KPI standing in for the aim. Publication over provenance is local optimization in academic reward geometry.

Counter direction

Single writer principle and clear domain ownership reduce cross-team negotiation cost — but reproducibility also requires shared L1 registries (datasets, environments, run IDs) that no local optimizer can skip without visible breakage. See experiment architecture and Event Sourced Science.

Corpus stance

A1 — adopted on this site: Explains persistence of ruptured provenance chain alongside architecture gaps.