Active Traversal and Load-Bearing Dependency II (ATLD 2): Residual Coordinate Identification and Self-Auditing Matched-Control Evaluation for Long-Horizon AI Systems

Abstract

Active Traversal and Load-Bearing Dependency (ATLD) v1.0 introduced a matched-control protocol for testing whether meaningful typed dependency architecture and active recursive traversal contribute causal functional gain beyond the same content presented through flat, isolated, semantically scrambled, or non-recursive controls. Its primary outcome vector contained seven readouts: task accuracy, multi-entry convergence, contradiction localisation, correction propagation, calibration, branch-failure localisation/repair, and efficiency plus evidence-path validity. The present paper asks a stricter successor question: can materially important long-horizon failures remain outside that seven-coordinate measurement body even when the original readouts are individually strong? ATLD 2 preserves the causal grammar of v1.0 and treats measurement refinement as the object of study. A structural seven-on-seven self-application of the frozen v1.0 vector yields five candidate residual coordinates: exact object/address custody (O), permission/action-state integrity (U), future-sufficient continuity/re-entry fidelity (R), receiver-side closure (V), and parent fixed-point closure (F). The paper formalises each candidate as an observable over a matched task world, defines direct and mirror deformations, and gives selective-ablation, counterfactual-substitution, coalition, causal-shadow, and held-out diagnostic tests. The resulting candidate measurement body is then subjected to a twelve-on-twelve self-audit, but self-application is explicitly non-certifying: retention requires prospective selective deformation, unique diagnostic gain under a no-smuggling rule, neutral-domain transfer, and cross-model replication. A derived route-support readout separates endpoint correctness from support for the declared traversal route using SUPPORTED, FAILED, and UNRESOLVED transition states and first-demonstrated-break localisation. A neutral synthetic “Release Relay” world demonstrates how the five candidate coordinates can be deformed one at a time without disclosing the private reasoning architecture that motivated their discovery. Causal-shadow accounting prevents one upstream fault from being counted as multiple independent failures, while bidirectional mirror tests prevent the new coordinates from becoming one-way optimisation targets. ATLD 2 therefore remains the same portable causal instrument as ATLD v1.0, but with a sharper measurement surface and a stricter burden for adding new dimensions. No empirical superiority or twelve-dimensional ontology is claimed.