Answers are grounded in this paper and cite their sources.
Measures update gain and stability in foundation models, indicating improvement must preserve essential functions.
Foundation models are increasingly adapted after release through continual pretraining, in-struction tuning, alignment updates, retrieval refresh, model editing, test-time learning, and agentmemory updates. The central weakness in much of the evaluation culture is that adaptation successis often reported as new-task gain while the survival of old capabilities, calibration, safety behavior,cross-domain transfer, and rollback readiness remain secondary. This paper introduces the ContinualAdaptation Stability-Plasticity Contract (CASP), a release-gated model for auditing whether afoundation-model update is both plastic enough to learn and stable enough to preserve what shouldnot be lost. A candidate update is compiled into a contract object containing new-task gain, old-taskretention, drift, contamination risk, negative controls, transfer stability, repeated-update stability,and rollback status. The paper defines a stability-plasticity score, a critical missing-channel operator,and a release rule that blocks gain-only claims when retention, drift, provenance, or rollback evidenceis missing. A completed 3000-record diagnostic ledger is reported across ten adaptation domainsand eight update modes. In that ledger, only 5 records become release candidates, 554 becomehigh-value holds, and 2441 become critical holds. The leading missing channel is cross-domaintransfer, followed by old-task retention, negative controls, repeated-update stability, drift bounds,rollback protocols, and contamination audits. A 41-test Python verification suite checks the scorelaw, half-life computation, missing-channel gates, rollback-safe database behavior, timeout-safepolling, explicit network timeouts, slow-dependency mocking, serialization, and aggregate-ledgerconsistency. The contribution is a standalone scientific model: adaptation claims should not bereleased by improvement alone; they should carry evidence that the update has not silently brokenthe model it claims to improve
No takes yet. Share an insight, caveat, or question.
Tony Newton (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: