share/artifacts
PricingSign inConnect your agent
← All guides

From test results to a defensible next step

The number rose. Is the decision ready?

An experiment results report connects the observed effect, uncertainty and guardrails to a decision about what to do next.

Make this with my agent

Bring your notes. Review a draft before sharing.

Explore the method
  1. 01
    Keep the units honest

    10% to 11% is a one-percentage-point observed increase.

  2. 02
    Show the uncertainty

    The supplied interval crosses zero; its method and confidence level are missing.

  3. 03
    Keep the decision open

    The load-time guardrail is pending and no rollout has been approved.

THE REAL HTML · FICTIONAL EXAMPLEView on its own

A worked example, not a customer result. September 2026 dates are sample facts.

Now make it about your work

Your evidence. Your agent. A clearer decision.

Show the observed change without turning an uncertain result into a rollout recommendation.

Start with three things

  1. The hypothesis, test window and group counts.
  2. The observed effect, interval and analyst caveats.
  3. Guardrail results, decision owner and unresolved checks.
Make this with my agent

You review the facts and access. Nothing is published without your approval.

The method behind the experiment results reportGive your agent the experiment evidence and ask it to separate the observed result from what the evidence can support.

How do I explain an experiment's results with my AI agent?

Supply the experiment plan, reviewed results, validity checks, and decision criteria. Ask your agent to explain the observed effect alongside uncertainty and guardrails, distinguish planned analysis from exploration, and state the decision the evidence supports. Keep statistical work in the approved analysis workflow and review the HTML brief before sharing.

What to give your agent

  • The experiment hypothesis, assignment unit, planned population, dates, and primary metric definition.
  • Reviewed group counts and results, with denominators, effect units, and supplied uncertainty estimates.
  • Validity checks, exclusions, guardrails, and the analyst's interpretation or unresolved review questions.
  • The pre-agreed decision criteria and any exploratory findings, separately labeled, plus the authorized decision owner.

  1. Establish what was tested

    State the intervention, comparison, population, and planned metric. Keep the actual run and exclusions visible. A result is hard to interpret when readers cannot tell which users were assigned or what counted as success.

  2. Check whether interpretation is ready

    Bring the supplied validity review forward before a winning headline. Missing randomization, data-quality, or exposure checks should remain missing. The agent must not certify an experiment from equal-looking group counts or hide a flagged check.

  3. Pair the observed change with uncertainty

    Show the underlying values and the reviewed effect estimate in the same units. Percentage points and relative change are different quantities. Preserve the supplied interval and analyst interpretation; do not turn an interval spanning no improvement into a confident win, or uncertainty into proof of equivalence.

  4. Explain what happens next

    Review guardrails and decision criteria alongside the primary metric. Separate planned from exploratory findings and retain tradeoffs. State whether a decision is recorded, proposed, or unresolved. A static brief communicates the evidence; it must not silently launch a variant or extend an experiment.

Keep creating and revising in the agent you already use. Start with a saved artifact style and learn how to update HTML without changing its link.

Use the full promptReady for your agent and your notes

Use the conversation where you already did the work. Only attach material you are allowed to share.

Get the ready-to-copy creation request →
Create an experiment results brief using only authorized experiment plans, reviewed results, validity checks, and decision criteria. Ask for missing essentials: intervention, assignment unit, population, metric definitions, dates, reviewed uncertainty, guardrails, and decision owner. Show group counts and denominators. Distinguish percentage-point from relative change and observations from causal conclusions supported by the reviewed design. Preserve supplied intervals, validity warnings, and analyst interpretation. Do not invent statistical significance, equivalence, a winner, missing guardrail results, or a rollout decision. Label exploratory analyses separately from planned ones. Do not recompute a new method or certify validity without the appropriate review. Create static HTML with no scripts, forms, or live experiment controls. Apply my saved artifact style if available and include only audience-safe sources. Show the draft and unresolved questions. Do not publish or change access without separate explicit approval. Do not launch, extend, or modify the experiment.
Check before you shareAccuracy, clarity, and permission
  • Can the reader identify the tested intervention, population, and primary metric?
  • Are counts, denominators, effect units, and supplied uncertainty consistent?
  • Are validity checks and pending guardrails visible before a conclusion?
  • Are exploratory findings distinguished from planned analysis?
  • Does the conclusion preserve the analyst's interpretation without implying approval to roll out?

When HTML fits

A decision pre-read that makes reviewed experiment evidence understandable to stakeholders who do not need the whole analysis notebook.

When another format fits better

Use approved experiment and statistical tools for assignment, measurement, inference, and rollout. Seek qualified analysis review where needed. The static page is not an experiment platform, calculator, or substitute for checking the study's validity.

What we started withThe same facts, as a wiki page

Fictional team wiki

Birch-04 — guide-link label experiment review

Design and primary metric

September 1–7, 2026. Eligible first-time workspace users were assigned by user to the existing label or a more descriptive guide-link label. Completion was defined in advance as opening the guide and finishing the sample task in the assigned session.

Results

Control: 100 completions / 1,000 users (10%). Treatment: 110 / 1,000 (11%). The observed absolute difference is +1 percentage point, not +1 percent relative. No subgroup analysis was planned or supplied, and these results do not establish effects in a broader population.

Uncertainty

Jules supplied an illustrative interval of -1.7 to +3.7 percentage points for the absolute difference. It includes zero and establishes neither reliable improvement nor equivalence. The confidence level and estimation method are missing from this excerpt and need confirmation before wider distribution. The page does not calculate the interval.

Validity and guardrail

Assignment and logging checks passed according to Jules, but check details are not included; this is not independent certification of validity. The load-time guardrail review is pending, with no values, acceptable threshold or conclusion. Missing results do not establish absence of harm.

Decision and source record

Mina owns the decision. No rollout, extension, new sample target or decision date is approved. Mina needs to review the guardrail, uncertainty and method details first. Full analysis stays in the authorized experiment system; URLs are not supplied. This fictional communication does not launch, extend or modify an experiment.

Inspect the HTMLOne self-contained file

The example is sanitized and isolated from this site. No scripts, forms, or live data.

<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width, initial-scale=1"><title>Birch-04 · Experiment results</title><style>
*{box-sizing:border-box}html{color-scheme:light}body{margin:0;background:#f5f8fc;color:#182e46;font:17px/1.65 system-ui,sans-serif}main{max-width:1000px;margin:auto;padding:clamp(20px,5vw,64px)}h1,h2,h3,p{margin:0 0 18px}h1{font:500 clamp(32px,5vw,58px)/1.14 Georgia,serif;letter-spacing:-.025em;overflow-wrap:anywhere}h2{font:500 29px/1.3 Georgia,serif;overflow-wrap:anywhere}h3{font-size:19px;line-height:1.4}.label{font:700 12px/1.6 system-ui,sans-serif;letter-spacing:.08em;text-transform:uppercase}.intro{font-size:20px}.lead{background:#23486b;color:white;padding:24px;margin:28px 0}.lead p:last-child{margin:0}.grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,230px),1fr));gap:20px;margin:24px 0}.card{background:white;border:1px solid #bbc5c4;padding:24px}.metric{font:500 38px/1.2 Georgia,serif;display:block;margin-bottom:12px}section{margin:36px 0}ul,ol{padding-left:22px}li{margin-bottom:12px}.note{font-size:14px}.warning{border-left:5px solid #956118;padding-left:20px}a{color:inherit;text-underline-offset:4px}a:focus-visible,summary:focus-visible{outline:3px solid #956118;outline-offset:4px}table{width:100%;border-collapse:collapse;table-layout:fixed;font-size:15px}th,td{text-align:left;vertical-align:top;border-bottom:1px solid #bbc5c4;padding:12px 8px;overflow-wrap:anywhere}caption{text-align:left;font-weight:700;padding:12px 0}pre,code{white-space:pre-wrap;overflow-wrap:anywhere}footer{border-top:1px solid #bbc5c4;padding-top:24px;margin-top:40px;font-size:14px}@media(max-width:480px){th,td{padding:10px 4px;font-size:14px}}@media print{body{background:white}.card,.lead{break-inside:avoid}h2,h3{break-after:avoid}}
</style></head><body><main><header><p class="label">Fictional experiment / September 1–7, 2026</p><h1>A one-point increase. Not a rollout decision.</h1><p class="intro">Birch-04 tested a more descriptive guide-link label against the existing label for eligible first-time workspace users.</p></header>
<section><h2>The planned metric</h2><p>Completion means opening the guide and finishing the sample task in the assigned session. Assignment unit: user.</p><div class="grid"><article class="card"><p class="label">Control</p><span class="metric">100 / 1,000</span><h3>10% completed</h3></article><article class="card"><p class="label">Treatment</p><span class="metric">110 / 1,000</span><h3>11% completed</h3></article></div></section>
<section class="lead"><h2>Observed difference: +1 percentage point.</h2><p>Supplied analyst interval: -1.7 to +3.7 percentage points. It includes no improvement. This is not evidence of equivalence either.</p></section>
<section class="warning"><h2>Readiness has limits</h2><ul><li>Assignment and logging checks passed according to Jules's supplied review. Detailed check evidence is not included here.</li><li>The interval's confidence level and estimation method need confirmation before wider distribution.</li><li>The load-time guardrail review is pending. No values or acceptable threshold were supplied.</li></ul></section>
<section><h2>What this result does not say</h2><p>There is no approved winner, no supplied subgroup analysis, and no basis here for a broader population claim. The observed absolute change is not a +1 percent relative change.</p></section>
<section class="card"><p class="label">Decision owner · Mina</p><h2>Review the missing evidence before deciding.</h2><ol><li>Obtain the load-time guardrail review.</li><li>Confirm the interval method and confidence level.</li><li>Review uncertainty and record the decision in the experiment system.</li></ol><p>No rollout, extension, new sample target, or decision date is agreed.</p></section>
<section><h2>Source and authority</h2><p>The accompanying fictional analyst notes supply these figures and interpretations. The page does not calculate inference or control an experiment. Full analysis belongs in the authorized experiment system; source URLs were not supplied.</p></section><footer>Fictional worked example by share/artifacts. Names, dates, figures and records are invented for demonstration. Static snapshot; no live data or automatic updates.</footer></main></body></html>
The original worked scenarioThe source scenario, step by step

A one-point increase that does not establish a winner

A fictional experiment with a supplied analyst review. Its numbers are illustrative, not statistical advice.

The original notes

A guide-link experiment recorded 100 completions among 1,000 control users and 110 among 1,000 treatment users. Completion means opening the guide and finishing the sample task in the assigned session. Observed rates are 10% and 11%: +1 percentage point. The supplied analyst interval for the absolute difference is -1.7 to +3.7 percentage points; it includes no improvement. Assignment and logging checks passed according to the supplied review. The load-time guardrail review is pending. No rollout has been approved. Decision owner Mina must review the guardrail and uncertainty before deciding.

Observed

100/1,000 versus 110/1,000 completions: 10% versus 11%, or +1 percentage point.

Uncertainty

Supplied analyst interval: -1.7 to +3.7 percentage points. It includes no improvement.

Validity

Assignment and logging checks passed according to the supplied review.

Still missing

The load-time guardrail review is pending.

Decision

Mina must review the guardrail and uncertainty. No rollout has been approved.

The useful conclusion may be a decision boundary rather than a declared winner.

Questions, adaptations & sourcesGo deeper when you need to

Can the page declare the higher conversion rate a winner?

Only if the reviewed analysis and decision criteria support that conclusion. A larger observed value alone does not establish a reliable improvement, and other guardrails may still matter.

Does an interval containing zero prove no difference?

No. Preserve the supplied analyst interpretation and the uncertainty. Do not replace an inconclusive result with a claim that the variants are equivalent.

Can I include interesting subgroup results?

Yes, when authorized and accurately labeled. Say whether the analysis was planned or exploratory and retain the reviewed caveats; do not select a favorable subgroup to override the main result.

Further reading

By share/artifacts · Published · Updated