- Added .gitignore to exclude unnecessary files and directories.
- Introduced UPGRADE.md to outline architectural principles and ideal agent functionality.
- Updated run.py to include a persistent state path for the agent.
- Enhanced runner.py to integrate linguistic scores into the agent's posterior update and manage state across tasks.
Enhance README and scripts for cognitive architecture testing
- Updated README to clarify the functionality of `UnifiedField.predict()`, including the expected output format and common follow-up actions.
- Added a minimal example in the README demonstrating the use of `TopologyMapper` for graph projection.
- Improved error handling in `chat.py` by removing redundant traceback imports and ensuring consistent exception logging.
- Refactored `compare_iterative.py` to consolidate shared parameters for scoring bridges, enhancing maintainability.
- Enhanced `smoke_extract.py` with detailed logging and summary statistics for grounding checks, improving diagnostic capabilities.
- Updated `runner.py` and `tasks.py` for better error handling and logging during task evaluation and sample loading.
- Introduced new validation checks in `StructuralCausalModel` to ensure proper parent variable definitions and counterfactual limits.
Refactor README and add new scripts for cognitive architecture testing
- Updated README to reflect changes in the Tensegrity architecture, emphasizing the unified energy landscape and the new V2 `UnifiedField` stack.
- Introduced `chat.py` for free-chat mode, allowing users to interact with the cognitive layer and observe belief states.
- Added `compare_iterative.py` to benchmark single-shot versus iterative cognitive scoring without LLM involvement.
- Created `smoke_extract.py` to test LLM transduction capabilities, ensuring structured output from prompts.
- Enhanced `run.py` and `runner.py` for improved benchmark evaluation and task management.
- Updated `tasks.py` to include better error handling and logging for task sample loading.
Refactor logit bias processing and enhance CanonicalPipeline
- Added a call to `_ensure_torch()` in the `TensegrityLogitsProcessor` to ensure proper tensor handling.
- Removed redundant `_ensure_torch()` call to streamline the bias processing logic.
- Updated `CanonicalPipeline` to include an additional line for improved functionality.
Enhance cognitive controller and Broca interface with new features
- Added stricter matching for relation terms against hypotheses in the cognitive controller, improving accuracy in belief updates.
- Introduced a rolling window for confirmed facts in the cognitive controller to manage belief state more effectively.
- Updated Broca interface to truncate situation summaries for causal hypothesis prompts, ensuring concise input for model generation.
- Implemented a check for duplicate model names in the Broca interface to prevent conflicts during SCM proposals.
- Enhanced schemas with validation for causal edges to ensure acyclicity in proposed structural causal models.
- Improved error handling and logging throughout the causal energy framework for better diagnostics.
Add initial project configuration and dependencies
- Introduced `pyproject.toml` for project metadata, specifying the name, version, description, and required Python version.
- Listed essential dependencies including `numpy`, `scipy`, `torch`, and others for the project.
- Added `uv.lock` to manage package versions and ensure reproducibility.
- Updated evaluation runner to use semantic field scoring instead of the previous Tensegrity v2 scoring.
- Enhanced the cognitive controller with new hypothesis generation capabilities and session management.
- Implemented a new causal energy framework for model competition based on prediction errors.
- Introduced FHRR-RNS encoding for compositional observation encoding, enhancing semantic processing capabilities.
- Established a unified cognitive engine integrating various components for improved performance and flexibility.
Confidence-gated blending for local mode: graft scales by (1-LLM_confidence), gated when LLM is confident (>0.6), scores scaled to 0.3x of LLM score range
Add inference settings for model loading: prioritize CUDA, MPS, and CPU. Update runner and pipeline to utilize new settings for dtype and device placement.
