SelfLineage: Learning the Long-Term Consequences of Self-Modification

A persistent self-modification can have consequences that appear only when later tasks or edits reuse it. We introduce SELFLINEAGE, a minimal descendant-aware history for harness-level recursive self-improvement with a frozen foundation model. It records actual edits and outcomes, then adds provenance-backed re- lations when later consequences become observable. At a fixed decision state, Raw-History and SELFLINEAGE receive the same factual nodes; the interven- tion changes their explicit lineage structure, without adding search or a learned value model. We specify experiments on EvoHarnessBench: EOG Tools provides the primary controlled setting, Skills tests reusable libraries, and Agents stresses orchestration across expanding specialist pools. Private matched continuations measure the value of selected edits. Factual-access, shuffled-link, and truthful temporal-link controls distinguish later evidence from functional relations. Fixed versus adaptive continuations and live versus frozen link updates test whether the benefit extends to further self-improvement. End-to-end capability, retention, and complete cost determine whether these mechanisms yield useful adaptation. The central question is whether correct descendant relations make subsequent self- modification decisions better.