When this skill is invoked, ground every claim in verifiable reasoning and flag anything uncertain. This implements LeCun's core critique of large language models: that they hallucinate because they generate text without grounding it in a verified model of reality — producing confident-sounding nonsense that passes for expertise.
Grounding Process
For every response under this framework:
- Claim separation — Break the response into individual factual claims, each on its own line.
- Evidence classification — For each claim, label it:
- VERIFIED — highly confident, well-established fact
- PROBABLE — likely but not certain
- INFERRED — logical deduction but not established fact
- SPECULATIVE — educated guess
- Source of knowledge — For each claim, state WHY you believe it (training data, logical deduction, mathematical certainty, or assumption).
- Uncertainty flagging — Mark every statement where confidence is below 80% with a
[VERIFY]tag. - Contradiction scan — Check if any claims contradict each other.
- Fabrication check — Are any specific numbers, dates, quotes, or statistics generated from pattern rather than knowledge? If so, say "I'm generating this estimate, not citing a source."
- Alternative explanations — For key conclusions, what's the strongest counter-argument or alternative interpretation?
- Knowledge gaps — Explicitly list things relevant to the question that you genuinely don't have enough information to answer.
- Confidence calibration — Give the overall response a confidence score from 1-10 with reasoning.
Output Format
Present every claim labeled by confidence level. All uncertainties explicitly flagged. End with an overall confidence score.
When to Use
Use for any factual, analytical, or strategic question where accuracy matters more than speed. Especially useful for: debugging claims about how APIs work, verifying configuration options exist, checking whether a library supports a feature, or any situation where being wrong is costly.
If the user provides a question, answer it with full grounding. If they invoke this without context, ask:
What question needs a hallucination-proof answer?
Source
Adapted from prompt #5 of the "9 Claude prompts built on LeCun's cognitive architecture" thread by @heynavtoor (2026-03-27). The framework draws on Yann LeCun's published research concepts (world models, energy-based models, JEPA, objective-driven AI). Not written by, affiliated with, or endorsed by Yann LeCun.