This paper introduces the Future Access Without Presence (FAWP) framework and the Triple Horizon model for detecting alignment failure and control collapse in modern AI systems. We formalize a critical asymmetry: predictive performance can remain strong even as the ability to steer, align, or constrain system behavior degrades. The framework separates three operational boundaries: (1) the Readout Horizon (loss of reliable observation), (2) the Steering Horizon (loss of effective intervention), and (3) the Functional Horizon (loss of task performance). Alignment failure corresponds to crossing the post-zero steering horizon (τₕ⁺) while predictive coupling persists—creating an operational "Ghost Zone" where systems remain legible but uncontrollable. We extend the framework in five directions: (i) distinguishing genuine from deceptive readout via Mirage Systems, where supervisor-visible telemetry remains coherent while true steerability fails; (ii) proposing horizon-scaling hypotheses (Sₙ, Gₙ) for model size and optimization depth; (iii) incorporating stochastic human response latency and motivating a Minimum Viable Safety UI; (iv) formalizing jurisdiction-dependent steerability via a Regulatory Divergence score; and (v) elevating circuit-level steering via sparse autoencoders, identifying a Last-Chance Intervention Band where latent interventions remain effective after surface prompting fails. Empirical grounding is provided via a layered benchmark suite (THB-5): synthetic delayed-control systems, LLM prompt-horizon tests, layer-wise internal Ghost Zone probes, multi-agent cascading FAWP simulations, post-hoc retrofit on log-structured traces, and synthetic Mirage benchmarks. Representative runs confirm nonempty FAWP windows, internal horizon separation across depth, cascading failure in supervisory stacks, and two-stage deceptive-readout dynamics. The resulting framework treats alignment failure as a layered transition—not a binary event—and yields practical early-warning signals, design constraints for horizon-locked architectures, mechanistic intervention targets, and governance metrics for systems that remain predictable after they have become difficult or impossible to control. Associated Resources: The reference implementation of the FAWP detection pipeline, including calibrated thresholds (αₐ ≈ 0.007297, αₐ² ≈ 5.325×10⁻⁵), persistence gating, Mirage Window detection, and the full THB-5 benchmark suite, is available via the fawp-index Python package and web scanner.
Clayton Ralph (Sun,) studied this question.