Workspace
Blog

Research with Beeblio, Part 3: Clean, Analyze, and Visualize

1 Nov 2026·

Trace cleaning choices, reproduce analysis outputs, and check figures against the data in Beeblio’s synthetic demo project.

Part 3 of How to Finish Your Research with Beeblio. Part 2 established the method, instrument, codebook, and raw-data snapshot. Here we turn data into results that another person can trace.

A clean table and a persuasive chart are easy to produce. The harder task is explaining every decision between the original data and the final claim. The Everyday Habits & Well-Being demo puts raw and clean files, analysis code, machine-readable outputs, and figures in one project so you can inspect that chain.

The codebook first: the definition every cleaning rule must answer to.

Compare raw and clean before calculating anything

Open 2-Data/daily-habits-survey-raw.csv and daily-habits-survey-clean.csv. Look for blank sleep values, NA screen-time entries, and category capitalization differences. Then read 2-Data/codebook.json to check expected types, units, and missing-value codes. A blank cell is a question for the method, not an invitation to fill in a convenient number.

In your own project, keep the raw file unchanged and create a cleaning log or executable transformation. For each rule, record the affected variable, condition, action, number of rows affected, and reason. If a response is excluded, the log should explain why it fails a rule set before analysis. If a missing value is imputed, the method should explain how and why. Revisit the study plan when a decision affects the sample or estimand.

You can mention the raw file in Beeblio chat with @ and ask for an inventory before asking for a transformation: “List missing-value codes and inconsistent categories in this CSV. Propose rules and counts, but do not overwrite the source.” Inspect the proposal, choose the rules, and save a new cleaned file plus the steps that made it. Challenge a questionable AI suggestion before accepting it.

The clean table, where every cell is a decision already made.

Match the analysis to the plan

Open 3-Analysis/analysis-notes.md. It calls for descriptive statistics, Pearson correlations, and a model of well-being using sleep, weekly activity, leisure screen time, and social connection. Open analyze.js to see the actual input path and outputs. The script reads daily-habits-survey-clean.csv and writes analysis-results.json and summary-statistics.csv.

Read those files together. Check that the analyzed count is 240, that activity is expressed per 100 minutes in the model, and that each reported number corresponds to the intended variable. Ask Beeblio to explain a coefficient or trace a table value to its calculation, then verify the explanation against the script. For a real analysis, also inspect assumptions, diagnostics, uncertainty, and the effect of missing-data choices; a single correlation or model fit value is not a complete result.

If you change the model after seeing data, record what changed and why. Do not quietly present an exploratory analysis as if it had been planned from the start. Keep intermediate outputs when they help reproduce a decision.

Use the analysis computer for the method you need

The demo’s analyze.js is a small, inspectable example. Beeblio’s research agent can also run analysis in an isolated cloud computer with a scientific Python stack and save changed outputs back to the project. That supports more than the demo’s correlations and linear model: for a suitable question, you can work with statistical tests and diagnostics, larger files or SQL over data files, notebooks, and publication figures. Choose the method from the research design rather than from a list of available tools.

For example, ask for a Python diagnostic of the planned model using the clean CSV, with the command or script, assumptions checked, and a saved results file. Then inspect the code and output in Analysis. If the agent proposes a different model, ask what problem it solves, whether the data support it, and how the change should be recorded in the analysis notes. The cloud run produces workspace artifacts; the researcher still judges whether the method and interpretation are sound.

Try asking Beeblio: “Trace the well-being model from its input file through the script to analysis-results.json. State each predictor’s unit and identify any interpretation that would overstate this design.”

The analysis, held to account: the project’s reproducibility checklist.

Build a figure that can be checked

Open Figures. The demo provides 3-Analysis/sleep-and-wellbeing.svg and activity-and-wellbeing.png. Check the axes, units, chart title, and description. Ask which rows and transformation produced the marks and whether the line represents an adjusted or unadjusted trend. The report’s sleep figure calls its line an unadjusted fitted trend; that is narrower than the multivariable model.

When creating a new figure, tell Beeblio which cleaned file to use, which variables to plot, the units, accessible labels, and the output format. Compare the generated figure against the saved analysis, then request a specific revision if the visual encoding is misleading or incomplete. Save a caption that explains the population, sample, measure, and uncertainty needed to read it.

The demo’s interactive-dashboard.html offers a way to explore a fitted scenario. Its slider holds other predictors at fixed illustrative values. That is useful for exploration, but moving the slider does not show that changing someone’s sleep would change their well-being. For a manuscript, keep a static figure that stands on its own; use an interactive view as an optional companion.

The demo also saves 3-Analysis/research-workflow.mermaid. Open it to see the question-to-report process as a diagram, then compare it with the rendered version in the report. A workflow diagram can explain provenance or a method; it should use the same file names and stages as the actual analysis.

A figure that can answer for itself: title, units, and sample in plain sight.

End with a results memo

Write a short memo in 3-Analysis that states what was analyzed, the sample and units, the main results, what remains uncertain, and where the supporting files live. The memo should distinguish a descriptive pattern from a causal explanation. The demo’s analysis-notes.md, JSON results, summary CSV, and figures give you material to make that check.

The handoff to Part 4: Write and review the manuscript is a set of traceable results, figures, and limitations. The manuscript should be built from those outputs, rather than from a remembered impression of the analysis.

Beeblio Team

researchbeeblioanalysis

The Premise

Why we're building a full-stack research workspace — and finishing the job the IDE started.