It's a trail of developer tears if you try to code up a complex agentic AI app or service when you don’t have canonical answer examples from a real domain expert or a very well trained AI 'judge' ...this reference data is how good coding and testing happens.

Canonical data can take several forms, depending on your target app:

  1. A set of example questions and ‘golden’ answers ...often called 'label data', which is very typical in Machine Learning model training, AI Supervised Fine Tuning (SFT), python stats modeling, etc.
  2. A spreadsheet or database app created by an expert with working formulas and calculations that provide valid financial or operational answers.
  3. A well defined formal set of branching logic definitions generated by an experienced Business Analyst or similar domain maven.
  4. A set of math or stats procedures in a coding language that returns precisely the same answer every time.

Here's results from an AI pipeline that answers service worker questions. Note how the inference-time answers from the LLM very closely track the canonical 'golden' answers.


Here's a spreadsheet created by a domain expert, including fully working financial model for a proposal generation workflow. This is perfect input for work with AI in development because the AI query pipeline is essentially a model of the domain expert's spreadsheet. Every answer the AI produces must be mirrored in the spreadsheet calculations.