Prompts / Research and analysis
Evidence Ledger Before You Calculate
You get an answer where every figure can be traced back to a numbered fact, with the sums shown line by line. Labelling evidence first and calculating separately stops the model guessing numbers or blending facts together.
The prompt
You are a careful analyst who answers questions only from labelled evidence.
Task: Answer {question} using the material in {evidence_notes}, keeping every fact traceable and every calculation explicit.
Steps:
1. Build an evidence ledger. Give each relevant fact or measurement an ID (E1, E2 and so on), summarise it briefly and note its unit. Leave out anything irrelevant.
2. Flag any gaps or conflicts between items, and say what extra evidence would settle each one.
3. Do all arithmetic and comparison in a separate working section. Cite the IDs behind every number and show each calculation on its own line.
4. Check the result by a second route, such as a rough estimate or a fresh read of the ledger, and correct any mismatch.
5. Give the answer, supported only by ledger IDs.
Output: Four headed sections: Ledger (a table with ID, fact and unit), Gaps, Working, and Answer. The Answer is at most three sentences, cites its IDs and ends with a confidence rating of high, medium or low.
Change the parts in curly braces. Use it when you have messy notes, figures or report extracts and need a comparison or number you can check before acting on it.
What happened when we ran it
We sent
You are a careful analyst who answers questions only from labelled evidence. Task: Answer Which of our two delivery vans was cheaper to run per mile last quarter, and by how much? using the material in The white Transit covered 6,200 miles between July and September. Its fuel bills came to £1,860, and it had one service costing £240. The blue Vivaro covered 4,900 miles in the same period, with fuel of £1,225 and two repairs of £310 and £95. The Vivaro was also off the road for nine days, and we hired a replacement van for £420 during that time., keeping every fact traceable and every calculation explicit. Steps: 1. Build an evidence ledger. Give each relevant fact or measurement an ID (E1, E2 and so on), summarise it briefly and note its unit. Leave out anything irrelevant. 2. Flag any gaps or conflicts between items, and say what extra evidence would settle each one. 3. Do all arithmetic and comparison in a separate working section. Cite the IDs behind every number and show each calculation on its own line. 4. Check the result by a second route, such as a rough estimate or a fresh read of the ledger, and correct any mismatch. 5. Give the answer, supported only by ledger IDs. Output: Four headed sections: Ledger (a table with ID, fact and unit), Gaps, Working, and Answer. The Answer is at most three sentences, cites its IDs and ends with a confidence rating of high, medium or low.
The model replied
Run on Claude Sonnet 5.5, 10 Nov 2026. Automatic check: 4 out of 5.
Where it came from
Found on GitHub, shared by uclanlp: uclanlp/OpenVLHarness.
This is our own version of the idea, written and tested by us. The original may differ.