Show HN: Veritas OS – Local OS for LLM Governance (github.com)

Veritas OS is an open-source, local “Decision OS” that wraps an LLM (e.g., OpenAI) as a low-level reasoning engine and builds a full agent stack — safety gates, memory, value-function, world model, debate/critique stages and hash‑chained trust logs — so the system acts as a safe, consistent, and auditable decision-maker rather than a chatbot. Exposed as a FastAPI service, a single POST /v1/decide runs a staged pipeline (Options → Evidence → Critique → Debate → Planner → ValueCore → FUJI → TrustLog) and returns a structured JSON: chosen action with rationale/uncertainty, alternatives, evidence, self-critique, debate views, telos_score, FUJI safety verdict (allow/modify/block/abstain) and an immutable SHA‑256 chained trust_log for auditability. System prompts are in Japanese but localizable. Technically, VERITAS OS centers on kernel/pipeline orchestration, Planner/Reason for chain-of-thought planning, WorldModel/MemoryOS with embeddings + nearest-neighbor search, ValueCore (EMA-based goodness metric), FUJI safety gate, and reflection/adaptation modules (RSI prototypes). It includes logging/dataset export, a Doctor dashboard for metrics, and OpenAPI 3.1 integration for reproducible experiments. Targeted at AGI research, safety/alignment work and auditable decision systems, it’s designed to run locally with Python 3.11, uvicorn and an LLM API, enabling repeatable, inspectable decision loops for safety testing and training data generation.