Project Eidolon explains how LLMs work by examining how and why they fail.
It is a systematic study of "hallucination" not as a bug, but as an architectural feature of coherence-optimized systems. It provides a standardized terminology stack, technical analysis of failure modes in agentic workflows, and a framework for understanding "confidence theater."
- The Terminology Gap: "Hallucination" is anthropomorphic; "Error" is too vague. We lack precise language for why a model invents a fact.
- The Agentic Risk: As LLMs move into IDEs and autonomous roles, "plausible but ungrounded" outputs (confabulations) become critical system risks.
- The Confidence Illusion: Users mistake stylistic fluency for epistemic certainty.
- Developers & Architects: Building agentic systems (IDE agents, MCP servers, RAG pipelines).
- Researchers: Looking for a comparative analysis of model failure modes.
- Educators: seeking to explain LLM limitations without mysticism.
- Not about Sentience: We do not discuss "thinking" or "feeling."
- Not about "AI Lying": There is no intent to deceive, only an optimization for pattern completion.
- Not a Prompt Library: This is an architectural analysis, not a collection of "jailbreaks."
EIDOLON_PRIMER.md: The core thesis and terminology stack. Start here./docs/: Detailed analysis of mechanisms, failure modes, and mitigation.