EaseBrainACADEMYInvestigate a dataset, compare a simple rule with an AI approach and produce an honest evidence card.
If one item feels unfamiliar, revisit these foundation missions:
day,litres,attendance,period 1,1000,100,before 2,950,95,before 3,1080,108,before 4,750,60,after 5,800,64,after 6,875,70,after 7,820,,after 8,900,0,after
Copy this into your chosen tool or trace it on paper. Do not enter passwords, private learner data or API keys.
A dataset needs a unit of observation. Here each row is a day, not an individual. A data dictionary records what a column means, its unit and possible missing values.
Copy the CSV starter into a spreadsheet or a paper table. Define every column and write one question the table can answer. Label the dataset fictional.
Your question fits the data. You do not claim to know any individual’s water use.
Missing and zero are different. A zero denominator makes litres per learner undefined. Keep raw data intact and record each cleaning decision so another person can reproduce it.
Make a working copy. Flag rows 7 and 8, then exclude their ratios with a reason. Calculate litres/attendance for rows 1–6. Use the foundation data lab to check.
The six valid ratios are 10,10,10,12.5,12.5,12.5. Excluded rows remain visible in a cleaning log.
An axis, unit and scale influence how a chart is read. A before/after difference is not proof that an intervention caused it. Confounders could change both timing and outcome.
Draw total litres and litres/learner as separate labelled charts. Calculate the 25% increase in the latter. Name one confounder and revise the original ‘water saving’ claim.
Your chart is traceable to the table; your conclusion does not assert unsupported causation.
A baseline is a simple method against which complexity is compared. A fixed FAQ lookup may outperform a generative answer on a small factual pack. Unknown should remain unknown.
Use the fictional River Club pack in mission 7. Write a lookup rule for time, cost and joining. Test known, missing and conflicting questions. Compare it with supplied AI outputs; live AI is optional.
Your comparison scores correctness, source support and uncertainty separately, including a case where no answer is available.
If you tune a rule using every test question, the score can overstate performance on unfamiliar questions. Reserve fresh cases before revising. Small samples cannot prove fairness or universal accuracy.
Prepare ten fictional questions. Use six to improve your method and reserve four. Report the two scores separately. Include varied phrasing and a question that the source cannot answer.
You disclose which questions influenced your changes, report errors, and avoid claiming a small sample proves fairness.
An evidence card describes intended use, data, method, evaluation and limitations. It helps another person decide whether the result fits a new context.
Choose a new public aggregate or fictional dataset. Produce the question, dictionary, cleaning log, chart, rule baseline, fresh-case test and human-review boundary. Python is an optional extension after you can explain the table method.
A reviewer can reproduce your analysis without a paid model, and can identify at least one situation where your method should not be used.
Choose a twist. Use paper, a sketch or a conversation.
Optional practice. Copy useful observations into your workbook; choosing a twist does not record completion.
Compare ideas with someone you know. Sharing is optional.