play icon for videos
← Academy
THE LOOP · REPORTING THAT WINS FUNDING

Get the Same Numbers Every Time

A general AI re-guesses every run, so the numbers move. Trust is the opposite: one scoring guide applied the same way to the same data, so the figure is the same on the second look as the first — and a number you can reproduce is a number you can defend.

Why does the same question give a different answer each time you ask?

Because a general AI re-guesses from scratch on every run — so ask it the same thing twice and the numbers move. Trust is the opposite: one scoring guide applied the same way to the same data, so the figure is the same on the second look as the first. A number you can reproduce is a number you can defend.

You’ve collected clean data against your definitions. This step is about making sure the analysis of it holds still — because a report whose numbers change when someone re-checks them loses a funder’s confidence fast.

Key takeaways

  • Consistency is what you’re selling — the same figure on the second look as the first.
  • The fix isn’t a cleverer model, it’s a fixed method: one scoring guide, one set of data.
  • A chat tool can sound sharper, but it won’t give you the same number twice.
  • Sopact Sense keeps the count steady, tied to source, even across thousands of responses.

Consistency is the thing you’re actually selling

Everyone has felt the drift: paste a set of responses into a chat tool, get a clean summary, run it again an hour later, and the counts have shifted. For an internal brainstorm that’s harmless. In a funder report it’s the difference between evidence and a guess.

The fix isn’t a cleverer model — it’s a fixed method. When the same scoring guide is applied the same way to the same stored answers, the result doesn’t wander between runs, and every number carries the words that earned it.

Try it: run the same scoring twice

Take a handful of open-ended answers and a simple scoring guide. Run it, then run it again unchanged.

Sort each answer below into exactly one category — ACCESS, COST, TIME, CONFIDENCE, OTHER. Quote the words that justify the choice. If no barrier is stated, mark NOT STATED and don’t guess. Return: answer · category · quote. Then repeat the exact same task again. The two results must be identical, line for line. Answers: [paste 5–10]

If they match, you’ve felt what a trustworthy number is: reproducible, and tied to the words behind it.

Where a chat tool stops

On a handful of answers, a general model can look consistent. Push it and the cracks show: on a few hundred responses it starts to drift and drop things, the counts change between runs, and it can’t tell you which answers produced a given number. It’s honest to say a chat tool can give a sharper-sounding summary — but it will occasionally invent detail, and it won’t give you the same figure twice. For a report, same-every-time beats clever.

How the number stays put

In Sopact Sense the scoring guide you wrote is applied the same way to every response, held in a durable record you can re-query. The count is the same next week unless the data changed, it holds on thousands of responses, and each figure stays tied to its source. One nonprofit leader described the switch in exactly those terms:

“The reports I’m getting are much more consistent and accurate — the same prompt used to give me very different results. Now I can see where the numbers came from, and they’re consistent from report to report.” — Marco, Open Play Foundation

When this helps — and when it doesn’t

Reach for it whenever a figure will be quoted, compared across groups, or put in front of a funder. For a rough first pass nobody will check, drift doesn’t matter — use whatever’s quick.

Frequently asked questions

Why does ChatGPT give a different answer to the same prompt?

It samples a fresh line of reasoning each run and re-reads the input from scratch, so counts and wording shift. A fixed scoring guide over stored data returns the same result every time.

Isn’t a consistent method less insightful?

It trades a little improvised flair for accuracy you can defend. For a funder report, that’s the right trade.

How do I show a number is trustworthy?

Reproduce it, and point to the responses behind it. Consistency plus a source is what earns trust.

Does this hold on large datasets?

Yes — large sets are where general tools drift most. One fixed rule applied to a durable record avoids that.

Next: Show Funders Where Every Number Came From → · or see consistent analysis in Sopact Sense →

Related from the Academy

Ready to try it for yourself?

Open Sopact Sense, paste your program description, and put it to work.

Try in Sopact