> ## Documentation Index > Fetch the complete documentation index at: https://docs.learningcommons.org/llms.txt > Use this file to discover all available pages before exploring further. # Graphics Accuracy > Check whether a math graphic, from plots to worked solutions, correctly shows what its accompanying text states. [Evaluator last updated October 14, 2026.](#evaluator-release-history) ## Overview The Graphics Accuracy evaluator checks whether a math graphic, such as a plot, chart, diagram, equation, or worked solution, correctly shows what its accompanying text states. It also checks whether the image is sound on its own: complete, legible, and internally consistent. The evaluator treats the text as correct and the image as the thing under review. It works in 3 steps: 1. **Describe the image:** Records what's literally in the image, counts each feature the claim depends on (such as dots, bars, or sides) as an exact number rather than an estimate, and lists any defects, like a missing or illegible label, tick marks that don't match their values, or parts that don't add up to their stated total. 2. **Work out an answer:** Finds the value or result the image yields, using only the image and any facts the question states in words, independent of what the claim says. 3. **Compare and decide:** Checks its answer against the claim, then gives its verdict and the basis for it. The evaluator returns a pass/fail verdict, not a score you set your own threshold on. Because the verdict is decided last, it follows from the answer the evaluator worked out. When the evaluator can't confirm that a graphic is correct, it returns false rather than true. ## At a glance | | | | :- | :- | | **Supported grades** | K–12 | | **Model used** | Gemini 3.6 Flash (`gemini-3.6-flash`) with the provider's default temperature The evaluator was built and validated using this specific model and temperature. Other configurations will produce different results and may have lower accuracy. | ## Getting started Follow the [Quickstart](/evaluators/getting-started/quickstart) to start using this evaluator: | Access method | | | :- | :- | | **SDK** | [Python](https://github.com/learning-commons-org/evaluators/blob/main/sdks/python/src/learning_commons_evaluators/evaluators/graphics/math/graphics_accuracy.py) ↗ (v1.1.0 or later) and [TypeScript](https://github.com/learning-commons-org/evaluators/blob/main/sdks/typescript/src/evaluators/graphics/math/graphics-accuracy.ts) ↗ (v1.2.0 or later) | | **Other** | [Python notebook](https://github.com/learning-commons-org/evaluators/blob/main/evals/graphics/math/graphics-accuracy/example_notebook.ipynb) ↗

Prompts: [`system.txt`](https://github.com/learning-commons-org/evaluators/blob/main/evals/graphics/math/graphics-accuracy/system.txt) ↗ and [`user.txt`](https://github.com/learning-commons-org/evaluators/blob/main/evals/graphics/math/graphics-accuracy/user.txt) ↗ | ## Inputs | Requirement | Supported | Required | | :- | :- | :- | | **Image** | One math graphic per call, as a local file path. To check several graphics or claims, make a separate call for each graphic and claim pair.
PNG, JPEG, or WebP, up to 5 MB. Width and height must each be 16–2,560 pixels. The format is read from the file's contents, not its extension. An image outside these limits is rejected, not resized. Animated, interactive, and handwritten graphics aren't supported. | Yes | | **Claim** | What the image should show, written as either:Up to 12,000 characters | Yes | When the graphic goes with a problem, use the question-and-answer form of the claim. The evaluator uses facts the question states in words, like given lengths or rules, and is less accurate without them. For example, instead of the statement "The trees have 9 apples," submit the question "How many apples do all the trees have together?" with the answer "9." The evaluator checks only whether a graphic is mathematically correct. It doesn't judge grade-level fit, standards alignment, or teaching quality. The evaluator isn't designed for grading student work, or for graphics that are wrong on purpose (like a "find the mistake" problem) unless the claim says so. ## Output | Field | Description | | :- | :- | | **Is correct** | Pass/fail verdict: | | **Basis** | Why the verdict was given: | | **Reasoning** | How the evaluator worked out its answer, ending in a sentence that compares the two: "The image yields X; the specification states Y" | | **Correction** | What the image actually shows, based on the answer the evaluator worked out. If it couldn't work one out, this reads "derivation incomplete" | | **Observed** | What's literally in the image, including counts of relevant features | | **Defects** | Problems with the image itself, like missing, cut-off, or illegible labels. Empty when the image is sound. | | **Errors** | Each way the image fails to show the claim: which value, step, or count is wrong, and what the claim required instead | ## Interpreting results | Output | How to use | | :- | :- | | **Basis** | | | **Reasoning + Correction** | Find the exact discrepancy. Compare what the image yields with what the claim states, and use the correction to see what the image actually shows. This also helps you tell whether your graphic or the evaluator got it wrong | | **Errors** | Look across graphics marked false for recurring errors, like miscounted features, misread axes, or wrong steps, and fix the prompt that produced them | ## Evaluator release history | Date | Changed | | - | - | | October 14, 2026 | First release | ## Related topics Explore other evaluators that assess the graphics in a text. Check math questions against a standard's learning components to see which math skills they measure. Run an evaluator in the Evaluators playground, a Python notebook, or with the SDK.