Hand-checking the output of a language model
A guided workspace for verifying what an LLM gives you: decompose it, check every claim and citation against a source, and leave with an evaluation record you can defend. Each step is backed by the published method it comes from.
The order is the method. Reading an LLM's output for plausibility in a single pass is exactly what lets fabricated quotes, invented citations, and confident wrong numbers through. Work the thread top to bottom; tie a knot wherever you find something.
Everything below is pre-filled so you can see how a finished pass reads: the claims table, the flagged citation, the findings on the thread, and the compiled record at the bottom. Clear it when you're ready to check your own.
Record what produced this output
Before judging anything, capture the conditions. Non-deterministic output can't be reproduced from a prompt alone, so the record is the artifact, and disclosure of model, version, and use is what reporting guidelines now require.
Provenance fields
Your compiled result
This updates as you work. Export it as JSON to archive or re-open later, or print it to PDF as a standalone evaluation record.
Provenance
Claims examined
Findings by stage
References
Every method in this tool traces to one of these. Links were checked to resolve; nothing here is generated. If a link ever breaks, the DOI or arXiv ID will still find it.