human-cognition-layer

Human Cognition Layer (HCL) is a research project focused on modeling human mental states and calibrating decisions. Its core approach is to not modify base model weights but instead use explicit, reproducible, and transferable cognitive structures to more stably model human knowledge, beliefs, intentions, causal uncertainty, and social decisions, thereby improving final responses or actions. The project adopts a module-first principle, where the cognitive layer is always retained in the processing pipeline, and performance degradation is treated as a diagnostic signal rather than bypassing HCL for better benchmark scores. The current phase is PHASE-04, focusing on decision strategy repair and new holdout validation, with model training not yet started. The HCL v0.3 architecture consists of three main parts: 1. Frozen Cognitive State: Each input first constructs a structured cognition state, supporting modes such as SIMPLE, EPISTEMIC, and CAUSAL_AMBIGUITY, emphasizing semantic principles like separation of world truth and agent knowledge, separation of first-order and second-order beliefs, and the need for evidence bridges for information transfer. 2. Always-On Answer Loop: After input passes through the cognitive state, base model draft, consistency, and calibration checks, a final answer is generated. The checker focuses on factual consistency, information accessibility, belief hierarchy, error disambiguation, and answer granularity. 3. Interaction Decision Strategy: In environments like SOTOPIA, the cognitive state is further mapped to action strategies to avoid overly passive or low-information behavior due to lack of action strategies. The project provides two types of evaluations: CogToM and SOTOPIA-Hard. CogToM currently serves as a diagnostic and regression tool, while SOTOPIA-Hard is the primary method validation environment. The README documents experimental data on state fidelity, answer checker, and SOTOPIA-Hard, but explicitly states these are research evidence rather than product capability claims. Positive holdout signals still require validation through predefined repeated experiments and cross-base model transfer. The repository also includes historical CogToM runner scripts that can run baselines or diagnostic tests via the DeepSeek API. Real-time execution status is based solely on STATUS.md as the single source of truth.