The Descendant Standing Problem: Governing Learned Policies After Retrospective Warrant Failure
Year: 2026
Type: preprint
Abstract: This paper introduces and formalizes the Descendant Standing Problem: how to assess the current governing basis of a learned policy or adaptive AI system when a governance determination within its learning lineage is retrospectively found to have lacked warrant under the historical rules applicable when it was issued. Governance outputs—including approvals, denials, reviewer judgments, verifier decisions, and eligibility determinations—can become supervisory signals, training labels, rewards, preferences, or memory entries. When their original warrant subsequently fails, correcting the source determination does not automatically resolve the standing of learned descendants. The framework distinguishes three structures that must not be conflated: recorded provenance, causal learning descent, and justificatory support. It develops a formal model using finite AND-OR/Horn support structures, set-relative warrant-withdrawal semantics, and three-valued clean re-derivation to distinguish established independence, certified failure of independent support, and unresolved support. The paper establishes conditional semantic correctness results, examines the relationship between causal influence and independent governing standing, and demonstrates why influence removal alone neither guarantees nor determines justificatory independence. Formally evaluated synthetic lineage cases illustrate these distinctions, including situations in which a fresh evaluator continues to rely on lineage-shaped evidence. The resulting remediation architecture separates source correction, causal repair, current admission, operative disposition, and downstream decision review. It provides a claim-relative framework for evaluating whether a learned descendant possesses a governing basis that can stand independently of a retrospectively defective ancestor. The contribution is not a new unlearning algorithm or provenance mechanism, but a distinct governance problem and formal architecture for assessing descendant standing after retrospective warrant failure in adaptive and agentic AI systems.
Author: Škrinjar, Sandra
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7221829256
Cites:
Cited by:
Related to: