Does a higher score mean the better prompt?
It means that version's answers have confidence better backed by their reasoning. That is one signal, not the whole decision.
Prompt A/B Scorer
Pick the prompt whose answers stay grounded, with a table instead of a hunch.
Upload your test inputs, choose two prompts or two models, and see side by side which version answers with confidence its reasoning supports, row by row, with an export you can drop into the decision doc.
Bring your own OpenAI, Anthropic, or Google key. Model costs stay on your account. Packs from $19 per 1,000 rows are planned. Early access only. No checkout.
Before and after
Before
Two prompt versions, a dozen hand-read samples, and a meeting where the decision comes down to whoever argued last.
After
A table. In one example, not a result, 500 inputs showed version B scoring higher on 61 percent of rows, and version A had 14 answers flagged as overconfident versus 5 for B. The decision has a record.
How you get there
You test on your own inputs.
Upload a CSV of inputs.
You compare like for like.
Pick two configs, prompt and model, across OpenAI, Anthropic, or Google. Both run on the same rows with your keys.
You see where they differ.
A side-by-side score and grade per row, with named problems, and a summary on top.
You keep the evidence.
CSV export of every row and score.
Honest limits
- ClearRun checks whether an answer's confidence is backed by its reasoning. It does not check facts. A high score means the confidence matches the reasoning, not that the answer is true. English only. Early access: independent validation of the score is in progress.
- It compares one dimension: whether each answer's confidence is backed by its reasoning. It does not measure correctness, tone, style, or cost. Use it alongside your own evals.
- Your model keys are used to generate both sets of answers. How those keys are stored is not settled yet, so do not assume they stay only in the browser.
Planned pricing
There is no checkout on this page. Joining the list does not start a subscription.
- Pack. Planned at $19 for 1,000 rows. Both configs on one input count as one row.
- Monthly. A plan at $29 a month is planned. The monthly row allowance is not set yet.
Questions
Do you pay for model calls?
No. Your keys, your model bill. The pack covers scoring.
Which languages?
English only.
Who is behind this?
Ross Shaw at Adepoco. Built on ClearRun. Tell us where the score got an answer wrong: contact@adepoco.com.
Join the early-access list
Tell us where to write when this opens. If you arrived with a referral tag, we keep the first one.