- Updated
- Version
- v1.1
- Sources
- 2 sources
- Status
- Published
fabricated answer · false claim
Hallucination
Generated information presented as factual when it is false or inconsistent with supplied material.
Everyday phrases
- verify a doubtful answer
- my agent invents sources
Key difference
A confident answer or citation is not evidence; open the source and check the passage supporting each useful claim.
Example
An answer attributes a date to a contract that contains no such date; you mark the claim unverified and reread the contract.
Counter-example
A story explicitly requested and presented as fictional is not a factual claim to verify.
Continue with a guide
Related concepts
Sources
- https://ai.google.dev/gemini-api/docs/google-search
- https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
Revision glossary-v1.1:hallucination · 2026-09-08
- Sourced content digest
- sha256:9bdea94f2745342d56ba680d13e96e4ef710eb31b162082c012382859d124fc0
Understand → Recognize → Choose → Compare
Pack for your agent
Pre-written instruction by SkillCodex — your request is neither sent nor used to adapt this text; no content is generated, and copying executes nothing.
Implement correctly
Use Hallucination deliberately
# Ground Hallucination in your agent ## Objective Use “Hallucination” precisely in code, documentation, and conversations: Generated information presented as factual when it is false or inconsistent with supplied material. ## Prerequisites - Inspect how the term is used today in the repository, documentation, and prompts. - Identify the neighboring notions already named: Agent, Chatbot. - Preserve existing correct usages. ## Working definition Generated information presented as factual when it is false or inconsistent with supplied material. ## Decisive difference A confident answer or citation is not evidence; open the source and check the passage supporting each useful claim. ## Example that counts as Hallucination An answer attributes a date to a contract that contains no such date; you mark the claim unverified and reread the contract. ## Counter-example A story explicitly requested and presented as fictional is not a factual claim to verify. ## Acceptance criteria - Every use of the term matches the working definition. - Edge cases are settled with the decisive difference, not by intuition. - Neighboring notions (Agent, Chatbot) are named with their own term. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
- Requires · The agentic vocabulary already used in the repository and documentation
Why it works
- The notion is framed by a sourced working definition, not ambient usage.
- The decisive difference settles edge cases reproducibly.
- The example and counter-example bound the term on both sides.
Try next
Map the neighboring notions
# Map the neighboring notions ## Objective Extend the same rigor to the related notions (Agent, Chatbot) so the project vocabulary stays coherent. ## Checks - List every neighboring notion used without a definition. - Apply its documented decisive difference. - Record the still-ambiguous cases for the next entry. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
sha256:fe40c0fb56e7976395dc288418c59a16e264d5ccec0a879c4cd360c2b6405f89
Diagnose a problem
Fix a Hallucination mix-up
# Fix a mix-up around Hallucination ## Observed symptom [DESCRIBE THE SYMPTOM HERE] ## Observable checks - Compare the disputed usage with the working definition: Generated information presented as factual when it is false or inconsistent with supplied material. - Apply the decisive difference: A confident answer or citation is not evidence; open the source and check the passage supporting each useful claim. - Confront the case with the documented counter-example: A story explicitly requested and presented as fictional is not a factual claim to verify. ## Possible causes - The term is used for a neighboring notion (Agent, Chatbot). - The observed form (interface, script, service) is confused with the actual behavior. - An alias (fabricated answer, false claim) carries a different meaning in the project. ## Bounded fixes - Rename or reclassify only the usages that contradict the working definition. - Document each choice by citing the decisive difference. ## Final verification - Reread every corrected occurrence against the documented example and counter-example. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
- Requires · The agentic vocabulary already used in the repository and documentation
Why it works
- The diagnosis compares usage with documented criteria instead of an opinion.
- Fixes stay bounded to genuinely contradictory usages.
- The final verification runs back through the sourced example and counter-example.
Try next
Anchor the vocabulary in the project
# Anchor the vocabulary in the project ## Objective Turn the clarification into a reusable rule so the mix-up does not return. ## Checks - Add the working definition to the project glossary. - Link every neighboring term to its own entry. - Check new documents against the decisive difference. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
sha256:6f5e1ad82cbe059c42b91b43da56e1637a92dd6344533d8ed91babc240607ee9
Pack digest: sha256:5e15eb6514f91c6bf7cb85a18e7edbc2a119ccbdc77a66ab329d29df21c0ebe1