Distinguishability, Recoverability, and Latent-State Identifiability: An Adversarial Audit of Discrete, Continuous, and Retrieval-Based Models

Technical research note and timestamped research disclosure documenting an adversarial model-identifiability program across discrete, continuous, dynamic, and retrieval-based latent-state models. The release consolidates empirical results from the E01–E10 computational program together with subsequent mathematical and causal-identifiability audits. The main findings are: (i) an apparent static advantage of a discrete latent-state model does not persist as a clear advantage against a complexity-matched continuous rival; (ii) low-order temporal similarity does not imply equality of full sequential path distributions; (iii) path-law distinguishability does not guarantee reliable finite-sample model recovery; and (iv) stochastic retrieval/access models are not uniquely identified by ordinary choice, recall/recognition, cueing, or consideration-set behavior when strong continuous-strength models with task-specific stochastic criteria/readouts are admitted. The archive includes the technical article, formal derivations, empirical and logical evidence ledgers, falsification tests, claim-status matrix, original E01–E10 audit materials, source provenance, and SHA-256 manifests. Importantly, this release preserves negative, corrected, and computationally inconclusive findings rather than presenting only positive results. It does not claim that a biological discrete latent state, binary access state, or any particular latent ontology has been established. Author: Alim ul Haq Khan ORCID: 0009-0001-4708-0365 Version: 1.0 Disclosure date: 5 October 2026