A pull request that says "I picked this LLM" shows a decision, not an experiment. An R&D assessor needs to see what you tried, what you measured, and why you chose what you chose. Without measured results, you cannot prove the work was experimental.
Decide how you will judge success before running the experiment. This is your benchmark, and it comes from your core activity. See how to define your core activities for R&D.
For each experiment, define:
Test several options against the same input and expected result, then record the results side by side.
Experiment: Which LLM identifies tasks best from a call transcript?
✅ Figure: Good example - Several options are tested against the same input, with measured results
PR title: "Use Model X for task extraction"
No other tests or results are recorded.
❌ Figure: Bad example - A decision with no evidence of what was tested or why
You need to be able to prove what happened. For every experiment:
Summarize the measured comparison above in four short statements so readers see the result and its boundaries together:
✅ Figure: Good example - A report opening that gives the result and its boundaries together
Follow the opening with the hypothesis, success criteria, method, full results and reproducible evidence. Before sharing, read only the opening and check it still says what was tested, what remains uncertain and who acts next.
Failed experiments still count as R&D. A clearly recorded "this did not work, and here is the data" is valuable evidence.