HDD-ISA - AI Architectures for Causal Discriminations

Abstract

This paper introduces the Intervention-Separable Architecture based on History-Dependent Dynamics (HDD-ISA), an architectural interface specification for designing or instrumenting AI architectures to turn functional claims into testable causal hypotheses. It provides: A five-stage causal chain: access → validity → engagement → effect → discrimination Construct-specific protocols for testing history-dependent prediction, causal trajectory dependence, feedback recurrence, functional self-reference, and self-modeling Design rules for architectures built to support causal testing Implementation guides for Transformers, RNN/LSTMs, and black-box LLMs Clear reporting categories: Supported, Negative Evidence, Uninterpretable, Non-Identifiable The framework serves two purposes: retrofit (testing existing architectures) and design (building new architectures with causal-discrimination interfaces from the outset). Core claim: HDD-ISA does not determine whether an architecture possesses a functional construct. It specifies the interfaces and conditions under which competing hypotheses about that construct become causally distinguishable.