About Plot2Data
Don't just extract a chart. Verify what it claims.
What is Plot2Data?
Plot2Data is an evidence-first workbench for reconstructing numerical values from supported chart images and checking whether those reconstructed values support a specific quantitative claim. It keeps the source figure, axis calibration, extracted traces, review decisions, calculation, and verdict connected in one visible workflow.
The current product is intentionally narrower than the original broad Plot2Data extractor. It prioritizes reviewability and reproducibility over claiming automatic support for every chart type.
From Extraction to Evidence
A CSV alone does not show whether the axes were calibrated correctly, where two traces were confused, or how a conclusion was calculated. Plot2Data places the reconstructed trace over the original figure, highlights uncertain regions, and asks for human review when the image does not justify an automatic choice.
After review, deterministic code converts pixel positions into data coordinates, interpolates supported values, propagates digitization uncertainty, and applies explicit verdict rules. The numerical result is not written or estimated by a language model.
How the Workflow Works
- Load one figure. Use a supported line-chart image or open the packaged reproducible demo.
- Inspect its structure. Gemini may assist with semantic inspection of the plot region, tick labels, scale candidates, legend, and visible series. It does not produce the final values or verdict.
- Calibrate and trace. Deterministic browser code maps pixels to axis coordinates and follows the selected traces.
- Review uncertainty. The overlay exposes deviations and ambiguous regions so a person can correct a specific area instead of redigitizing the entire figure.
- Evaluate one supported claim. The claim is converted into a bounded structure and evaluated with explicit numerical rules.
- Export the evidence. The browser creates a downloadable bundle containing the source, hash, overlay, calibrated data, claim, verdict, provenance, and report.
Current Supported Scope
Plot2Data currently supports one 2D line chart with up to three series; linear or log10 axes; solid or dashed traces; and circle, square, or no markers. Supported claim forms cover a ratio at a specified X value, an absolute or relative difference at X, and a single-series threshold crossing.
Multi-panel figures, bar charts, pie charts, heatmaps, candlesticks, statistical-significance analysis, error-bar semantics, full-paper interpretation, and arbitrary scientific claims are outside the current scope. See the benchmark protocol for the conditions and metrics used to evaluate the workflow.
Three Distinct Roles
- Gemini semantic inspection proposes chart structure, labels, scale candidates, legend entries, and potential ambiguity from the uploaded image.
- Human review resolves visual ambiguity that cannot be justified reliably from the pixels alone.
- Deterministic code performs calibration, tracing, interpolation, uncertainty propagation, arithmetic, and verdict classification.
This separation is central to Plot2Data: model output is a proposal for inspection, not an authoritative measurement.
What a Verdict Means
Plot2Data may return SUPPORTED, NOT SUPPORTED, INCONCLUSIVE, or NOT VERIFIABLE under its published calculation rules. A verdict means only that the values reconstructed from the provided figure do or do not support the entered quantitative claim within the selected tolerance and measured digitization uncertainty.
Digitization uncertainty describes uncertainty introduced while reading pixels, calibrating axes, following traces, and interpolating values. It is not a statistical confidence interval. Plot2Data does not determine whether a paper is correct, whether a result is statistically significant, or whether the underlying source data is authentic.
Our Mission
Data locked inside figures should be inspectable, challengeable, and reproducible. Plot2Data is designed to help researchers, engineers, analysts, journalists, educators, and students move from an opaque chart image to an evidence trail they can review and share.
Get in Touch
Have questions, benchmark cases, or feedback? Join our Discord community or visit the contact page.