Short-Termism

Near-term rewards and deadlines displace long-horizon truth obligations

MARKO ŽIVKOVIĆ In review

Short-termism is the organisational and individual habit of optimising for immediate visible payoff — this quarter, this grant cycle, this release, this publication deadline — at the expense of obligations whose cost is deferred and whose benefit is long-horizon: reproducible artifacts, honest tier, maintained provenance, and corrections that propagate.

It is a persistence driver for many epistemic failures: provenance is “later,” documentation is “after acceptance,” replication is “someone else’s problem.” The ML reproducibility crisis is often misread as culture alone; short-termism is one reason culture and architecture reinforce each other — when pinning datasets and publishing code does not pay off before the next review, ruptured provenance chain becomes the default unless structure makes deferral impossible.

Where it shows up

Domain	Example
Research	Ship the paper; skip the reproducibility package (publication over provenance)
Product ML	Deploy the model; defer lineage and monitoring commits
Enterprise	Close the audit; provenance deferred until the next incident

Relation to other drivers

Local optimization — each actor hits their local deadline; the system loses reproducibility. Misaligned incentives — reward functions that pay for headlines, not chains. Together they explain why “try harder” fails without experiment architecture and pinned epistemic harness.

Corpus stance

A1 — adopted on this site: Structural factor, not moral excuse; architecture must make honest capture cheaper at commit time than deferral.