NOEMA-OrganFarm: Embedding Human-Designed Epistemic Operations into the Latent Execution Path of a Largely Frozen Language Model

NOEMA-OrganFarm investigates whether human-designed epistemic operations can be redistributed from externally serialized reasoning procedures into conditionally executed latent modules within a largely frozen language model. Built on Qwen3-14B-Base, the system keeps the base-model backbone frozen while inserting separately trainable Transformer-like “Organs” into the hidden-state pathway. These Organs are assigned restricted cognitive responsibilities, including world/frame control, premise reconsideration, epistemic-state formation, deterministic routing, and downstream information-oriented processing. The study focuses on four architectural properties: responsibility-specific latent modules, conditional execution, latent-state inheritance between upstream and downstream Organs, and separation of semantic payload from execution-control signals. In the final Fresh Final end-to-end evaluation, the system preserved an internally consistent computational execution path in all 24/24 cases. Strict final-branch agreement was 20/24. All 8/8 cases requiring an UNKNOWN state reached the downstream Evidence Targeting Organ. Remaining errors were localized primarily to O2 semantic state formation, shallow Evidence-Target inversion in some O3 cases, and one surface-schema deviation. The results provide behavioral evidence that procedural epistemic cognition can be implemented as a modular latent execution architecture inside a largely frozen Transformer, while preserving state inheritance and enabling failure localization across cognitive stages. This publication records the NOEMA research lineage: NOEMA-1 → NOEMA-Organ → NOEMA-OrganFarm The paper reports the high-level architecture, training principles, evaluation design, aggregate results, and observed failure patterns. Complete Teacher datasets, exact routing implementation, state-transition procedures, information-requirement construction procedures, internal opcode details, and turnkey reconstruction materials are not publicly disclosed. This record contains the English and Japanese editions of NOEMA-OrganFarm v1.0. NOEMA-1: Epistemic-State Control through Structured Post-Training and Serialized Intermediate Representation NOEMA: Procedure-Oriented Post-Training for Epistemic-State Control in Mid-Sized Language Models | Zenodo