How GradeLamp Scores an Essay

Step 1: checks in code, before any AI

When you press grade, our own code looks at your essay first. Nothing here involves a language model.

• Length. Each rubric has a minimum length, and no essay can be longer than 2,500 words. You see the problem next to the box, and it doesn't use a grade.
• Language. GradeLamp grades essays written in English. If most of the words aren't English, we say so and don't grade; if it's mixed, we ask before grading.
• Is it an essay? Lists, code or a few notes get a gentle “this doesn't look like an essay yet” instead of a meaningless score.
• Text stats. Word count against the limit, sentences, paragraphs and your most repeated phrases appear straight away. They're shown for your information and are not part of your score.

Step 2: Claude scores each rubric row

We send the rubric and your essay to Anthropic's Claude. The rubric and our grading instructions go in first; your essay goes in separately, marked as text to assess. If an essay contains instructions to the grader, they're treated as part of the writing and the result tells you we ignored them.

For every row of the rubric, Claude must return:

• the level the essay reaches and the points for that level;
• one or two sentences quoted word for word from your essay as evidence;
• a short reason that ties the quote to the rubric's description of that level;
• one fix: an instruction about what to change and where, never replacement wording.

The instructions tell the grader to say plainly when a draft isn't ready yet, without padding the feedback with praise, and never to comment on whether writing “sounds like AI”.

Step 3: code checks the work and adds the points

Language models can slip on arithmetic, counting and rules, so the model never adds up your score. Our code does:

• Quote check. Every quote must appear in your essay, after allowing for spacing and curly or straight quote marks. Quotes that don't match are removed and flagged; if more than half fail, the essay is re-graded once automatically.
• Point caps. Each row's points are kept within that row's maximum.
• Fix check. Any fix that offers wording for you to copy, or quotes text that isn't in your essay, is rejected and asked for again once, then dropped.
• Rubric rules. Word limits (such as 350 words for each UC question and 650 for the Common App essay), required elements that cap other rows, and, once AP® grading opens, DBQ document counts worked out from verified quotes.
• Total and verdict. Code adds the rows, maps the total to a band and, for college and general essays, to a plain readiness verdict: “Not ready yet”, “Getting there” or “Ready after a polish”. Totals on college essays are GradeLamp's own practice scale; colleges don't score essays out of 30.

Consistency: the same draft gets the same result

AI models don't always give the same answer twice, and inconsistent scores are the biggest complaint about AI graders. We handle that in code rather than hoping the model agrees with itself:

• Each result is filed under a fingerprint of the exact text, prompt, rubric and grader version. If you are signed in and submit unchanged text again, you get the saved result back at no cost. Without an account, your browser keeps your recent results for 2 hours and shows the saved one.
• When you revise, the new draft is graded fresh: earlier scores are never shown to the model.
• A row only shows a change (“+1 since last draft”) if its level changed and the sentences behind it changed. Otherwise it reads “About the same (within normal re-run variation)”, so you don't chase noise.
• If we update the grader between your drafts, the result says so, because some of the change is ours.

The model, and what happens to your essay

Model. Every essay type currently uses Anthropic's Claude, model claude-opus-5-5 at “medium” effort, in one call per essay that scores every rubric row. Free and paid users get the same grader. If the model declines to grade an essay, we retry once on claude-sonnet-5-5; those results are labelled “Graded by our backup model”, and we'll publish how often that happens.

Connection. We call Anthropic's API directly. Your essays never pass through any other AI service or router.

Your essay. It's never used to train AI, by us or by Anthropic. Without an account we don't store it; only the finished result is held, encrypted, for up to 15 minutes so a dropped connection can collect it. Signed-in essays are deleted after 30 days, or when you delete them. Anthropic deletes what we send within 30 days, unless its safety systems flag it, in which case it can keep it for up to 2 years. We don't have a zero-data-retention agreement with Anthropic yet.

Where. Our servers and database run in the United States. Anthropic's processing isn't limited to one country. Details are in our privacy policy.

The rubrics

Every rubric is written in our own words. Where an exam or application has published guidance, we link to it and follow its structure (for example, how many points each row is worth), but we paraphrase the descriptions rather than copy them. We don't send the College Board's prompts, documents or sample essays to the model, and the AP rubrics open only after they pass their own test.

Each rubric is reviewed and signed off by a qualified reviewer before its grader opens. The grader uses the same rubric files as the tables on our rubric pages, so what you read is what you're scored against.

How accurate is it?

Honest answer: not measured yet. Every essay type that's open carries a Beta: accuracy still being measured tag, and each result says “Accuracy not yet measured for this essay type” until we have results to show. We publish a number only when the essays, raters, date and grader version behind it are stored and linked from this page. Below 30 essays, we'll show counts (“16 of 20”) labelled “preliminary” instead of percentages.

We test on essays we commission from paid adult writers, never on students' essays. Each test essay is scored independently by two qualified raters who don't see the intended score or the AI's. We then run each essay three times with saved results switched off, to measure how often a score changes on its own.

Calibration status

What each essay type must pass before it opens

These are the bars, not results. An essay type that misses them stays closed. The Beta tag comes off only after a larger test, when GradeLamp's agreement with raters is about as close as the raters' agreement with each other.

The accuracy study (January 2027)

We're running a larger study: 150 essays (50 AP Lang, 50 DBQ and 50 LEQ essays written for us by paid adult writers), scored by 2-3 experienced AP teachers, scoring blind and independently. At least half are written under timed conditions with no target score. We'll report GradeLamp's agreement with the teachers row by row, beside how well the teachers agree with each other, with confidence intervals, how often scores change between runs, and how a plain chat prompt does on the same essays.

We aim to publish it on 15 January 2027 (at the latest 5 February 2027), with every essay's scores downloadable under CC BY 4.0. It uses only our own essays and prompts: no College Board samples.

Known limits

• Holistic rows are the least reliable. Rows that reward judgment, such as sophistication, complexity and personal voice, carry a “less certain” tag on every result.
• Some writers may be scored differently. Essays by English-language learners or written in a dialect could be scored less fairly. We test for this with dedicated sets of essays; the results aren't in yet, and we'll publish them here.
• Text only. DBQ documents must be pasted as text. Images of documents aren't supported.
• Scores are estimates. They aren't official AP scores, admissions ratings or your teacher's grade. Use them as a second reader, not a final judge.
• We don't detect AI writing. Detectors wrongly flag honest writers, and that isn't a call software should make. Our academic integrity page explains more.

What we never do

• Rewrite, draft, paraphrase or “humanize” your essay.
• Run AI detection or label writing as AI-written.
• Give percentiles, “chances” or admissions predictions.
• Train AI on your essays, or use them as our test essays.
• Let the model add up points or count documents.
• Publish an accuracy number without the data behind it.

Grader changelog

Every result records the grader version that produced it: a fingerprint of the model, its settings, the rubric and our instructions, so any change creates a new version. Each change is tested against our test essays before release, and the grader for college essays stays frozen through the main deadline season.

Versions are listed here from launch. Spotted a score that looks wrong? Use “Question this score” on the result, or email support@gradelamp.com.