Frequently asked questions

How to read RegCheck output, and answers to common questions.

  • Deviation: the model judged that the paper diverges from the registered plan for that dimension.
  • No deviation: the model judged that the paper is consistent with the registration for that dimension.
  • Insufficient evidence: the model judged that there wasn’t enough information in the provided text to make a reliable call for that dimension.

We use "embeddings" models, a type of natural language processing model, to represent the meaning of extracted text. We compare the similarity of each extracted chunk of text (think, for example, paragraphs) against the definition of a specific dimension. We use a metric called 'cosine similarity' to estimate how similar each chunk is with that dimension (from a scale of 0 to 1). Higher values mean the chunk is more strongly semantically related to the dimension of comparison.

Direct quotes display direct quotes from the source material (i.e., registrations/papers) coupled with IDs and relevance scores. Summary displays LLM-based summaries of these excerpts.

Yes. Anyone with the report link can view it. Do not share links publicly if uploads contain sensitive information.

See the Privacy Policy for storage, retention, and deletion details.

For ChatGPT you can set how much ‘thinking’ the model does before answering. Higher effort is more thorough and can be more accurate on hard comparisons, but is slower and costs more; lower effort is faster and cheaper. It is not simply “higher is always better” — medium is a sensible default, and lower is often fine for straightforward documents. This setting applies to ChatGPT; other models ignore it.

The model menu lists families; the exact default models are ChatGPT = gpt-5.5, Claude Opus 4.8 = claude-opus-4-8, DeepSeek = deepseek-reasoner, and Qwen3.6 27B = qwen/qwen3.6-27b (open-weight, served via Groq). These are the defaults; a deployment can override them. The full API contract is in the API reference.