The Amaya Axioms: A Structural Latent-Dynamics Audit Framework for Evaluating Joint Resistance to Goodhart Optimization

This repository presents the Amaya Framework, a reproducible structural auditing suite for evaluating the internal latent-state dynamics of simulated and computational systems. The framework evaluates whether a system’s latent trajectory satisfies seven preregistered structural constraints: self-model existence, identity persistence, metacognitive error decoupling, cross-partition verifiability, perturbation propagation, bounded recursive depth, and predictive advantage over baseline forecasting models. The implementation uses ridge-regularized vector-autoregressive transition estimation, chronological train/test evaluation, singular-value diagnostics, spectral-radius analysis, and empirical perturbation-propagation testing. The repository includes a multi-target adversarial red-teaming matrix consisting of 20 base seeds across seven axiom targets, yielding 140 independent hill-climbing optimization trajectories and up to 42,000 candidate evaluations. Targeted attacks reliably spoofed Axioms 4 and 7 in isolation, demonstrating that the optimizer was mechanically capable of satisfying individual constraints. However, no trajectory satisfied all seven axioms simultaneously, yielding a joint pass rate of 0/140 under the tested attack budget and configuration. These results provide preliminary evidence for joint resistance: individual structural constraints may be manipulated in isolation, but isolated optimization did not produce full-framework compliance in the tested setting. The findings do not establish strict axiom orthogonality, universal Goodhart resistance, consciousness, agency, or phenomenal experience. The framework is released as a falsifiable research baseline requiring broader model families, stronger baselines, larger adversarial budgets, real-world time-series validation, and independent replication. Methods Summary: The audit evaluates latent trajectories using seven structural constraints: self-model existence, identity persistence, metacognitive error decoupling, cross-partition verifiability, perturbation propagation, recursive depth, and predictive advantage. Transition matrices are estimated with ridge-regularized vector-autoregressive regression using a 70/30 chronological train/test split. Perturbation propagation is measured by applying normalized perturbations to latent states and computing their amplification under the fitted transition operator across 20 propagation lags. The adversarial evaluation consists of 20 base seeds across seven axiom targets, producing 140 independent hill-climbing trajectories and up to 42,000 candidate evaluations. A structured positive control and Gaussian-noise negative control are included. Key Results: Across 140 targeted adversarial trajectories, Axioms 4 and 7 were individually spoofed in 20/20 trials each. No trajectory satisfied all seven axioms simultaneously. The joint pass rate was 0/140. The structured positive control passed all axioms, while the Gaussian-noise negative control failed identity persistence, cross-partition verifiability, perturbation propagation, and recursive depth. Interpretation Boundary: These results provide preliminary evidence for joint resistance under the tested hill-climbing adversary, seed set, optimization budget, and audit thresholds. They do not establish strict axiom orthogonality, universal Goodhart resistance, consciousness, agency, or phenomenal experience. A computational representation is not the phenomenal state it describes. Code Availability: The complete audit and adversarial-optimization code, including the seven-axiom harness, multi-target hill-climbing adversary, and reproduction routines, is included in the PDF version of this manuscript. No external repository is required to access the code.