Proposal-Veto Balance for Observable-Only Autonomous Intelligence: Stability Thresholds, Identifiability Limits, and Commit-Window Effects

This repository contains the source files and compiled manuscript for a theoretical study of long-horizon autonomous self-modifying intelligence under observable-only, no-meta conditions. The central problem is how an autonomous system can decide which self-modifications to commit when true proposal quality is latent, decisions must be made from internal state and current observables alone, and no external meta-controller can be assumed to intervene reliably. The paper develops a stochastic proposal-veto framework with verification signals, commit windows, accumulated error debt, and delayed correction. It derives explicit mathematical conditions for bounded expected debt, positive Harris recurrence, and geometric divergence in the absence of debt-proportional correction. It also introduces observable-only progress-safety frontiers based on total variation and ROC analysis, posterior-threshold acceptance rules, progress ceilings, delusion and ossification horizons, compute-budget and thermal constraints, trust-chain amplification on dependency graphs, rollback and committee-veto effects, and finite-horizon overflow bounds under a clipped-affine reduction. The aim of the work is to provide a scientifically explicit and machine-readable foundation for analyzing when autonomous systems can remain both productive and stable under finite observability, finite resources, nontrivial latency, and self-induced error accumulation. The manuscript includes theorem-level dependency metadata in YAML form to support automated parsing, formal inspection, and AI-assisted verification workflows.