Validating hypotheses with Tests

A Test checks a specific belief against the evidence Synthight already has — no new research required, answers in about a minute.

Writing a hypothesis

Sidebar → Research → Tests → New Test. In the "What should we test?" dialog, describe a belief in plain language, for example:

"Customers churn when they do not connect a data source."

Click generate. Synthight works through four steps: reframing your hypothesis, generating around ten differently-worded variations of it, using AI to match those variations against your finding groups (a finding worded completely differently can still match if it means the same thing — this isn't keyword search), then scoring confidence.

How confidence is scored

Two things drive the score, together:

• How many people are affected. A match against findings that touch a large share of your affected customers counts for more than one touching just a couple of people.
• How strong the match is. Each matched finding scores based on how closely it aligns with your hypothesis, not just whether it cleared the bar to count at all.

The combined score is then dampened when very few people are involved overall — the same idea as churn risk's confidence dampener: a hypothesis that only matches evidence from two or three customers can't reach a high confidence score, however strong the wording match, because there simply isn't enough evidence yet to be sure.

This produces two separate readouts:

• Status — Validated (strong evidence), Inconclusive, or Weak, based on the confidence score.
• Confidence level (High / Medium / Low) — factors in evidence breadth on top of the score. A test can land on "Validated" and still show Low confidence if it's backed by only a handful of people — worth noticing before you treat it as settled.

The test detail page

Three tabs:

• Evidence — matched findings with their match strength, plus a confidence gauge
• People — who's behind the matched evidence
• Variations — the different phrasings Synthight generated and matched against

From here you can Re-run analysis (useful after new data comes in), Duplicate a test to try a variation, remove a specific finding from the evidence set (which re-scores the test), or Archive/Delete it.

Next steps

• Simulations overview
• How churn risk is calculated — the same "not enough data yet" dampening logic, applied to a different score
• Understanding findings

Frequently asked questions