hallucination-reduction

Reduce hallucinations through first-principles reasoning, verification, self-critique, and explicit uncertainty for AI agents.

Anti-hallucination research skill for Claude Code — admits uncertainty, extracts direct quotes before analysis, cites every claim, retracts unverifiable statements. Based on Anthropic's official guardrail techniques. By TheGEOLab.net

About Essays, and Poc on using runtime evidence to build cleaner code context for AI to reduce hallucinations.

Genesis Governance OS The Operating System for Multi-Agent AI Inspired by Political Science, built to coordinate intelligent agents at scale.

Self-healing RAG system that retrieves, verifies, and grades its own answers. Automatically rewrites queries and retries when outputs are weak, ensuring accurate, hallucination-free responses.

Hallucination-prune multiagent RAG for pharmaceutical knowledge bases

Why Pure Vector Search is a "False Proposition" for RAG?

Dependency-free evidence core for AI agents: observation envelopes, provenance, memory continuity and claim gates to reduce hallucination drift.

Autonomous AI research agent using LangGraph to eliminate LLM hallucinations via a Generate-Critique-Refine self-reflection loop.

Allow coding agents to copy large blocks of files without reading them

System prompt that enforces strict compliance, self-auditing, and hallucination reduction in any LLM. Time-anchored, evidence-declared, confidence-scored, release-gated.

Developer-first prompt engineering patterns for grounded, testable, and reliable AI outputs.

Structured memory system and behavioral guardrails for AI agents. Reduce hallucinations, preserve technical decisions, and enforce Explore→Execution workflow boundaries during vibe-coding with OpenCode / Claude Code.

An agentic, self-correcting RAG pipeline that extracts claims from generated answers, verifies them using Qdrant and NLI, and automatically repairs hallucinated facts.

An RLHF-inspired DPO framework that explicitly teaches LLMs when to refuse, significantly reducing hallucinations.

Prompt engineering framework + evaluation harness for LLM workflows (classification, summarization, extraction).

Code and data for the INLG 2026 paper "Do LLMs Make More Mistakes If They Do Not Believe the Input Data?"

Label every claim in AI output as (u) given, (m) checked or (g) generated, in plain text, so a guess can't quietly become a "fact" when text passes between people and AI agents. Includes a swarm test (the spoke and wheel test), a parser and a gate. Early findings.

Differential Attention Transformer — training & inference framework (reduces hallucination, improves long-context retrieval).

Professional cross-agent answer quality gate for improving AI responses: intent match, evidence, assumptions, verification, brevity, and usefulness.