Work

Companion Addendum CP-2: Learned Intent, Observer Contamination, and Fail-Closed Gating in Gridworld Alignment Regularizers

Year: 2026
Type: preprint
Abstract: Companion Addendum CP-2 to "Topological Ethics: A Geometric Framing for Alignment Regularizers" (doi: 10.5281/zenodo.23842536) and "Topological Ethics Technical Companion" (doi: 10.5281/zenodo.23842862). This report presents three pre-registered gridworld experiments (Experiments 5, 5b, and 5c; 20 seeds per setting) testing what happens when an oracle schedule forecast is replaced by an online empirical Markov intent detector under stochastic human dwell (p_leave = 0.35) and optional sensor noise (0.00, 0.10, 0.25). The experiments isolate fail-open detector collapse, reflexive observer contamination during agent-obstructed corridor steps, noise-induced threshold dilution, and an undisturbed-data blind spot in offline clean training, showing how corridor-masked online updates and fail-closed gating restore cooperative doorway timing. All code, diagnostic utilities, statistical checks, and per-seed results are archived in Topological Ethics: Gridworld Experiments v1.2 (doi: 10.5281/zenodo.243556 87).
Language: English
Open Access status: green
OpenAlex ID: W7236206690
Cites:
Cited by:
Related to: 10