Decisions that become code.
Seven original recipes: prepare data, ask questions, check results, and leave a path for uncertainty.
RAG with citation checks
Select useful passages, reject embedded instructions, and check whether a citation actually supports the claim.
Up to 1 API requests per runExtract values without inventing them
Find the refund amount and scheduled date among several values in a document. Return exact source spans.
Up to 1 API requests per runClassification with a parent-category fallback
Choose a broad category first, then refine it. When uncertain, keep the request at the broader level.
Up to 2 API requests per runTool selection without arbitrary execution
Turn a request into an allowed read operation. Both the tool name and its argument must belong to a predefined list.
Up to 1 API requests per runSearch and reranking
Narrow a local corpus using query words, then assess which candidates actually answer the question.
Up to 1 API requests per runRepeatability and question batching
Compare three questions together and separately, preserving every answer, time, and usage value. Errors remain in the report.
Up to 12 API requests per runModeration with a review queue
Distinguish questions, respectful criticism, spam, and abuse. Leave uncertain messages for a moderator.
Up to 3 API requests per runInspect the plan. Then run it yourself.
Without flags, the runner shows a plan and author expectations without calling the model. --live makes real requests only with your API key and saves responses. --replay recomputes decisions from your saved report without network access.
Inputs are fictional and written in English. The bundle contains no invented model answers. These examples are not customer case studies or quality measurements; validate thresholds for your own task.