CIPHER: Prospective Early Warning for Hidden Objectives

CIPHER studies whether hidden objectives can be detected prospectively—before they become reliably visible in model behavior. The work introduces Counterfactual Gradient Alignment (CGA), a prospective signal for comparing early training directions with counterfactual semantic objective directions, and integrates it into a structured-blind auditing pipeline for objective discovery, trigger identification, causal validation, and targeted remediation. In a locked blind evaluation on controlled model organisms, CIPHER identified 17 of 18 objective-bearing runs with 0 observed alerts among 14 controls, recovered the exact objective-and-trigger pair in 15 of 18 cases, and a frozen CGA forecaster identified 14 of 18 future objective-bearing runs. Among 11 future-positive runs whose behavior was still subthreshold at an early training checkpoint, 8 were already flagged prospectively. Targeted semantic remediation also substantially outperformed generic clean continuation in this controlled setting. The central scientific claim is intentionally scoped. The main new benchmark uses small controlled neural model organisms, and hidden-objective discovery is blind within a prespecified semantic hypothesis space rather than fully open-world. A Qwen2.5-0.5B experiment is included as a language-model anchor, not as evidence of frontier-scale generalization. The broader proposal is that hidden-objective auditing should be treated as a causal identification problem under distribution shift rather than as representation separability alone. The accompanying reproducibility package contains the manuscript source, vector figures, experimental code, protocol locks, blind predictions, result tables, and scripts required to inspect and reproduce the reported analyses.