What is it?

HALc implements a 'constructive confrontation' algorithm between multiple LLM agents.

For each request, agents independently frame the problem, propose solutions, challenge competing views, provide evidence, steelman alternatives, and revise their positions. HALc then scores the proposals, selects a decision, and records unresolved dissent and reasoning in a ledger. The goal is to turn disagreement between agents into more robust, transparent, and evidence-driven decisions.

Motivation

Building robust architectures or solving complex problems using single LLM prompts often overlooks edge cases, security hazards, and operational risks. Constructive confrontation introduces a dialectic process where multiple distinct agent personas (e.g. architect, product, security) evaluate, challenge, and score each other's proposals. By aggregating these critiques and performing structured Copeland ranking, HALc filters out fragile concepts and presents the most robust solution as an OpenAI-compatible API response.