Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Vizly computes statistics and drafts comparisons; the statistician recomputes figures, checks deviations and signs the readout.
Collect results, plan and any run log
Draft statistics and the deviation table
Check figures and decide deviation impacts
Draft the readout from checked tables
Check every claim
Use: You
Have ready: variant results, written plan and any run log; retain planned sample size, allocation, primary metric and stopping rule.
Confidential experiment data goes into an external tool only if your organisation's rules allow it.
Use: Vizly
Ask Vizly for rates, confidence intervals, between-variant effect, sample-ratio check and code, then planned-versus-observed differences with missing evidence marked.
What to paste into the tool
Task: Draft results table (CSV: variant, n, conversions, rate, interval) including the between-variant effect and its interval, plus the code behind the analysis. From this table compute per variant: n, conversions, conversion rate and a 95% confidence interval; then the difference in conversion rates between variants with a 95% confidence interval for the difference; and a sample-ratio mismatch check against the planned allocation. Show the code behind the analysis and return the result as a CSV table. Task: Draft deviation table (planned, observed, difference) with still-to-confirm marks where evidence is missing. Compare the planned values in this written test plan with the observed values in the results table, row by row. List each difference. Where no run log is supplied, mark the peeking and stopping-rule rows as still to confirm; do not infer them from the totals. No new numbers beyond the differences.
Use: The statistician
Recompute every figure independently; check deviation rows against the plan and checked results; decide which deviations invalidate the result.
Recompute every figure with your own tool; if any number differs, the Vizly output for that row is not used. Decide which deviations invalidate the result. Where no run log exists, peeking and stopping-rule deviations stay marked still to confirm and are not inferred from the totals. Check each row against the written plan and the results table before using it.
Use: Vizly
Ask Vizly for effect, interval, deviations, sample-ratio check and a statistician's review list using only checked results and deviations.
What to paste into the tool
Task: Draft readout: effect, interval, plan deviations, sample-ratio check and a statistician's review list. Draft a readout using only these two tables. State the effect and interval, list each plan deviation and the sample-ratio check, keep every still-to-confirm mark, and end with a review list for a statistician. No new numbers.
Use: The statistician
Check the readout line by line against checked tables; remove untraceable numbers, sign the review list and store the audit files.
Numbers reproduce from the table; each deviation from the plan is listed. A readout is not a launch decision. Remove any number that does not trace back to the checked tables.
You import results and plan; the statistician computes the statistics; Gemini drafts the comparison table, then you write the readout.
Collect the plan and import the results
Compute rates and intervals
Draft the planned-versus-observed table
Check comparisons, write readout
Use: You
Have ready: variant results, written plan and any run log. Give Gemini the variant table and the plan as two pasted inputs.
Confidential experiment data goes into an external tool only if your organisation's rules allow it; this record does not establish that upload is permitted.
Use: The statistician
Compute the conversion rate and confidence interval per variant, the difference between variants with its confidence interval, and the sample-ratio mismatch check with spreadsheet formulas.
Recompute with a second method or tool; the spreadsheet formulas are inspected cell by cell before the table is used.
Use: Gemini
Ask Gemini to merge the results with the plan sheet: planned sample size, allocation, primary metric and stopping rule.
What to paste into the tool
Task: Draft deviation table (Table: planned, observed). Join the plan sheet to the results table on variant and metric. Return one row per planned value with the observed value beside it. Do not change any numbers.
Use: You
Check every merged row; add impact decisions. Write effect, interval, deviations, sample-ratio check and review list from the checked tables.
Check every merged row against the written plan and the results table; you add the impact column and mark peeking and stopping-rule checks as still to confirm where no run log was supplied. Approve each claim in the readout against the results table and the deviation list.
The statistician computes statistics and deviations; Hex drafts text from the checked tables, then the statistician rewrites and signs it.
Compute statistics
Draft readout text
Check claims
Use: The statistician
Have ready: variant results, written plan and any run log. Compute rates, confidence intervals, effect, sample-ratio check and planned-versus-observed impacts.
Recompute with a second method. Peeking and stopping-rule checks are marked still to confirm where no run log was supplied, not inferred from the totals. Confidential experiment data goes into an external tool only if your organisation's rules allow it; this record does not establish that upload is permitted.
Use: Hex
Paste only the checked results and deviation tables into a Hex notebook; ask for readout text preserving every still-to-confirm mark.
What to paste into the tool
Task: Draft readout: effect, interval, plan deviations, sample-ratio check and a review list. Write a readout using only the two tables in this notebook: effect and interval, each plan deviation, the sample-ratio check and a review list for a statistician. Keep every value marked still to confirm as still to confirm. Add no new numbers.
Use: The statistician
Rewrite the draft where needed; check each sentence against the two tables, remove untraceable claims and sign the review list.
Remove anything that does not trace back to them.
Best when: You want visible analysis code and a draft readout.
What it does: Vizly computes and drafts; you recompute figures and sign
Strong at: Computes statistics with code and drafts planned-versus-observed comparisons.
Limit: Every figure needs recomputation; missing run logs leave stopping checks unresolved.
Free plan: Yes, limited to 10 messages a month
Best when: You want human-owned arithmetic and a spreadsheet comparison table.
What it does: Gemini compares tables; you compute statistics and write readout
Strong at: Merges planned and observed values into a draft deviation table.
Limit: You compute statistics and add every impact decision yourself.
Free plan: Yes
Best when: You want all computation and plan comparison kept with people.
What it does: Hex drafts text; you compute, compare, rewrite and sign
Strong at: Turns two supplied checked tables into readable readout text.
Limit: You compute, compare, rewrite and sign; Hex does no statistical checking.
Free plan: Yes, limited AI credits
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Numbers reproduce from the variant result table.
Each deviation from the written plan is listed with its impact.
Peeking and stopping-rule checks are marked still to confirm where no run log was supplied.
A statistician recomputes the figures and signs the review list.
Check data-handling rules before uploading experiment data to any tool.
Check the result against your original inputs. Approve it before sharing or acting.