@ejentum

Ejentum is an open-source reasoning harness tool designed to enhance LLM agents by injecting structured cognitive operations directly into their context during inference. It addresses the reasoning gap where models often stop at plausible explanations or commit logical errors without explicit feedback. The tool uses RA²R (Reasoning Ability-Augmented Retrieval) to retrieve engineered cognitive procedures, not documents, from a library of 679 operations across four harnesses. The API returns a six-block injection: PROCEDURE (natural language steps), REASONING TOPOLOGY (executable DAG), COGNITIVE PAYLOAD (amplify/suppress signals and style), FALSIFICATION TEST, NEGATIVE GATE (failure pattern to avoid), and TARGET PATTERN (correct reasoning). Suppression signals constrain output space multiplicatively. Two delivery modes are offered: Dynamic (single highest-scoring operation) and Adaptive (top-k retrieval with an adapter LLM rewriting the operation for task-specific identifiers).