Residual Boundary Theory: An Integrated Computational Framework for Source-Sensitive Self-World Revision, Latent Release, and Perspectival Access

Abstract

Residual Boundary Theory (RBT) studies when an unresolved discrepancy changes a self-world relation and when the revised relation is subsequently used. The proposed mechanism organizes permitted experience by source uncertainty, action contingency, event identity and failed obligation, then uses that organization to select histories for revision under a fixed resource budget. Causal inference, case retrieval, representation learning and flexible access are established components. The remaining claim is a task-bounded hypothesis about their specified dependency and its incremental effect under matched alternatives. The framework separates predictive evidence, provisional revision, validated repair, content-specific use and report. Conditional results cover source identifiability, retrieval under a total budget, the propagation of retrieval gains through downstream stages, event-pair retrieval, censored first use and pre-probe completion under bounded probe-induced change. These results identify both useful regimes and counterexamples: equally accurate predictors need not identify a causal source; higher relevant-hit probability can lose after indexing cost or failed validation; and later success need not imply earlier computation. The mathematical results are design and measurement constraints, not evidence that an RBT-specific mechanism operates in humans. A staged experimental programme couples a frozen episode bank, strong case-based and replay controls, consistent event renaming, selective misbinding and restoration, and independently assessed downstream use. Optional branches address new feature construction, latent release and governed agentic adoption. The two causal roles are assumed rather than bootstrapped, and phenomenal sufficiency is not claimed. The package contains executable finite diagnostics and prospective predictions; it reports no new participant observations or trained-agent efficacy result. Scope and status. This theoretical manuscript (about 15,200 words, dated 19 September 2026) integrates six technical modules; "Paper 1" through "Paper 6" refer to those modules. It reports no new participant data or trained-agent efficacy study. This is the flagship item of a four-part Version 1.0 submission: the integrated flagship, Companion A (representation revision and counterexample-driven conceptual expansion, Modules 2 and 4), Companion B (governed binding and agentic reorganization, Module 5), and a prospective empirical programme protocol. The accompanying supplement archive holds the module-level formal models, external evidence ledgers, finite diagnostics and their reference outputs for all six technical modules, with a README, MANIFEST.json and SHA-256 checksums (80 files). The diagnostics check the stated mathematical models; they are not independent empirical confirmations.