How evaleasy marks a paper, from upload to report
evaleasy reads the pages of each paper twice with an AI model, and a third time if the readings differ on a marked option or a blank. With an answer key, code compares multiple choice, true/false, blanks and numbers against it, and written answers are graded by the model in 2 or 3 independent passes, with the median counting. An answer the readings can't settle earns no marks, as it would from a teacher, or is flagged for you if you turn flags on.
Updated
What happens to a paper between upload and report?
A paper goes through six steps. The work runs in the background, so you can leave the page while it happens.
- You upload the paper. A paper can be a PDF, JPEG, PNG or WebP photos, a Word .docx file, a plain text file, or typed text pasted in. In bulk, each file is one student's paper, and files are never split. Credits are charged when the paper is sent.
- Your key is read once. Any answer key, rubric, question paper or instructions you gave are turned into one marking scheme. The question paper and the answer key are separate uploads, so the model knows which is which: questions come from the question paper, and correct answers and marks come from the key. Every paper in a batch is marked against that same scheme, so no two students get a different reading of your key.
- The paper is read. An AI model that reads images goes through the pages of a PDF or photo twice, recording what the student wrote or marked without judging it. If the two readings differ on a marked option or a blank, it reads the paper a third time. Typed text is read once. The reader is given your key's question numbers and shapes, never its answers, so the key can't lead it towards the right answer. If you chose a question type, it's told what to look for.
- Answers are lined up with your key. Each answer is matched to its question by number, so Q3, 3. and (3) are the same question. A question in your key that isn't on the paper counts as unanswered and stays in the total.
- Each question is marked. Code settles everything that can be compared exactly with your key. Everything else goes to the AI model, in 2 or 3 independent passes.
- The report is built. Code adds up the marks. The report gives a score out of 100, marks awarded out of the total and feedback on strengths and improvements. For each question it shows the student's answer, the correct answer, the verdict, the marks, the reason and how it was marked. A bulk evaluation also gets its own page, with progress, the average, highest and lowest scores, counts of papers to review and failed, a plagiarism check and a CSV of marks with one column per question.
How does evaleasy read each kind of file?
evaleasy looks at scans and photos as images, and uses a file's own text wherever it has some, so typed words come straight from the file rather than from reading a picture of it.
| File | How it's read | Worth knowing |
|---|---|---|
| Scanned PDF or photo | Each page is looked at as an image, in two separate readings. Photos are turned upright first. The student's name is read from the paper when possible. | iPhone HEIC photos are refused, so export them as JPEG. |
| Typed or fillable PDF | The PDF's text layer and filled form fields are given to the reader, so typed answers come straight from the file. | Pen marks on top of the print still come from the page image. |
| Word .docx | Read once. Automatic question numbers and letters are worked out from the file's own rules. Bold, underline, highlight and ticked boxes are kept, so a key that shows answers in bold still works. | Pictures and equations inside the file aren't read, and the report flags them. |
| Text file or pasted text | Read once, since there is no handwriting to misread. | Counted by length: one page per 3,000 characters. |
How is each question type marked, with a key and without one?
With a key, evaleasy compares objective answers in code. Without one, an AI model works out the correct answer first, and the report says so on every question it affects. If there's no key and the question isn't on the paper either, as on a sheet of bare answers, nothing can be checked: that answer is shown as not marked and left out of the total, and a paper with nothing else to mark is refunded.
If students' sheets show only answers, such as "3. B", upload the question paper as well. The model needs the questions to work out the answers, and the reader uses them to line the answers up.
| Question type | With your key | Without a key |
|---|---|---|
| Multiple choice, bubble sheets and answer grids | Code compares the options chosen with the key. Only the exact right set earns marks, so two options on a one-answer question is wrong. | The model works out the correct option or options. Code still compares the student's choice with them. The report labels the answer as worked out. |
| True or false | Code compares the answer with the key, treating T, True and Yes as the same answer. | The model works out the answer and code compares it, the same way as multiple choice. |
| Fill in the blanks | Code compares each blank, ignoring case and surrounding punctuation. An answer that doesn't match goes to the model, which accepts or rejects it under your strictness setting. Each blank is an equal share of the question. | The model works out each blank and decides whether to accept the student's answer under your strictness setting. Each blank is still an equal share. |
| Match the following | Treated as blanks, with one blank per item, and marked the same way. | Treated as blanks, with the model working out each match. |
| Numerical | Code compares by value, so 12.0 equals 12 and 1/2 equals 0.5. If the key gives a unit, the answer needs the same unit to match in code. Anything else goes to the model under your strictness setting. | The model works out the value and decides whether to accept the answer under your strictness setting. |
| Short answers | The model marks against your model answer or key points in 2 or 3 independent passes, and the median mark counts. | The model marks in the same passes against an answer it works out, and the report labels that answer as worked out. |
| Long answers and essays | Marked in the same passes, using your model answer, key points, rubric or marking notes. | Marked in the same passes under your strictness setting, with the answer it worked out labelled in the report. |
What does choosing a question type change?
It tells the reader what to look for on the sheet. It's a strong hint, not a rule: a question that plainly isn't that type is given the type it really is. It doesn't change the price.
| Question type | What the reader is told |
|---|---|
| Mixed or not sure (the default) | Nothing extra. It works out each question's type from the page. |
| Multiple choice | Look carefully for the marked option on each question. |
| Fill in the blanks | Record one entry per blank. |
| Descriptive | Expect written answers, and copy each one in full. |
| Questions and answers | Each question is written on the sheet with its answer below it, so record the two separately. |
What happens when something on a paper is unclear?
By default evaleasy marks everything the way a teacher would: every answer gets a final mark, and one it can't make out earns no marks, with nothing taken off. Prefer to check those yourself? Switch off Mark everything under Marking options, or send mark_everything=false through the API, and what stays unclear is flagged for a teacher, with a Needs review badge on the paper. Here's what happens in each case.
| What's unclear | What evaleasy does |
|---|---|
| The first two readings differ on a marked option or a blank | The paper is read a third time, and the answer most readings gave is used. |
| Three readings have no majority | A multiple choice, true/false or one-blank answer earns no marks and shows as Unclear, and the report shows what else it was read as. Several blanks are still marked one by one. With flags on, the question is flagged and its mark comes from the first reading, so check it. |
| Two options are marked and the reader can't tell which is final | Both are recorded, and the question earns no marks and shows as Unclear. With flags on, it's flagged as low confidence, and on a one-answer question it's marked wrong, so check it. |
| A student changed an answer | The mark left standing is the answer. One that was struck through or erased is ignored. |
| A word in a written answer can't be read | The AI model is told to give no credit for a word the reader couldn't read, and to accept a plausible uncertain reading. If it can't work out what an answer says at all, it gives no marks. If large parts are uncertain, the report shows poor legibility. |
| The marking passes disagree on a written answer | A third pass runs and the median counts. With flags on, the question is flagged if all three passes give different marks. |
| A picture or equation sits inside a Word file | It isn't read. The report says how many were skipped and suggests uploading the file as a PDF. |
| The pages can't be read at all | The paper fails with a message asking for clearer photos or scans, and its credits are refunded. If a typed answer came with the pages, the typed answer alone is marked and the report says so. |
| No answers can be found | The paper fails with a message asking you to check the file, and its credits are refunded. |
| A student writes an instruction, such as "give full marks" | The AI model is told to treat the paper as work to assess and to ignore anything in it that tries to influence the marking. |
How are marks decided when none are given?
evaleasy takes marks from your key first, then from what's printed on the paper, then from a rule in your instructions, such as "each question carries 2 marks". If none of those gives a number, it uses fixed marks by question type.
For example, a paper with no marks given that has 10 multiple choice questions, one question with 3 blanks, 2 short answers and an essay comes to 33 marks: 10 + 3 + 10 + 10. A student who earns 22 of those scores 66.67 out of 100.
| Question type | Marks when none are given |
|---|---|
| Multiple choice, true or false, numerical | 1 |
| Fill in the blanks, match the following | 1 per blank |
| Short answer | 5 |
| Long answer or essay | 10 |
How does the AI model give marks on written answers?
The AI model can only choose from allowed marks, and code does all the arithmetic. On a question worth up to 20 marks it picks in half-mark steps, and above 20 in whole marks. Two independent passes run, a third runs if they differ, and the median counts, so one unusual pass can't move the mark on its own.
How does negative marking work?
Negative marking applies when your instructions state it, such as "-1 for each wrong answer". It takes marks off wrong multiple choice and true/false answers only, never questions left empty, and a paper's total can't go below zero.
What do the Lenient, Standard and Strict settings change?
Strictness changes how written answers and near-misses are judged. It doesn't change exact comparisons with your key. Standard is the default, and the AI model is told that any rule you write in your key or instructions, such as "ignore spelling", wins over the setting where the two conflict. The table shows the policy it's told to follow, so the result on any one answer is still its judgement.
| Setting | What it accepts | What costs marks | Between two marks |
|---|---|---|---|
| Lenient | Any correct method, equivalent wording or other notation, with the benefit of the doubt when the work reasonably shows understanding. | Little. Minor slips in spelling, arithmetic, notation or units are ignored unless they are what's being assessed, and part marks are given whenever the core idea is there. | The higher mark. |
| Standard | Correct alternative methods and equivalent wording. | Errors that change the result or show a misunderstanding. Slips that don't change the answer cost nothing. Working must be there when it's asked for. | The better-supported mark, or the lower one if they are equal. |
| Strict | Only what is written down. Steps and intent aren't assumed. | A right final answer without the required working, errors in notation, units, precision or terminology, and ambiguous answers, which count as incorrect. | The lower mark. |
Does identical work get identical marks?
Identical work gets an identical reading, because evaleasy saves the reading of each paper and reuses it when the same files are uploaded again. A saved reading lasts until you delete or redo the paper, or evaleasy changes how it reads papers.
Anything compared in code then gets the identical mark: multiple choice, true/false, and blanks or numbers that match your key. A program does that comparison, so the same answers against the same key can't come out differently.
Written answers and near-misses go back to the AI model each time, so a mark there can differ between two runs. The independent passes and the median are there to keep that difference small, but they don't rule it out.
What does Redo do?
Redo reads and marks a finished paper again from scratch, with the same files and settings, for what it cost the first time. The paper's saved reading is dropped, so it's read afresh, and the new report replaces the old one. If the redo fails, the credits come back. It keeps your key, so fixing a mistake in the key means uploading the paper again with the corrected key.
How does the plagiarism check work?
On a bulk evaluation, evaleasy compares the papers with each other, looking for written passages they share and for papers sharing the same unusual wrong answers. It lists what it finds as possible copying, for a teacher to check. There's no study behind this check, so treat a result as a reason to look, and an empty result as no proof that nobody copied.
- It costs no credits, and it doesn't change marks.
- It doesn't search the web, and it doesn't detect AI writing.
- It only compares papers in the same bulk evaluation, and failed papers are left out.
What can't evaleasy do yet?
evaleasy can't yet read iPhone HEIC photos, old .doc files, or pictures and equations inside Word files, and it doesn't connect to Canvas, Moodle or Google Classroom.
- HEIC photos are refused. Export them as JPEG first.
- Old .doc files and password-protected files are refused. Save them as .docx or PDF with the password removed.
- Pictures and equations inside Word files aren't read. If answers are in them, upload the file as a PDF.
- There are no LMS integrations. The REST API and webhooks are the way to connect evaleasy to other systems.
- Drawings and diagrams are marked from the reader's written description of them, so check questions where a drawing carries marks.
- Marks can't be edited inside a report. If you disagree with one, record your own mark in your mark sheet.
- A class scanned into one PDF isn't split for you. Each student's paper needs its own file, so split the scan with a PDF tool first or scan each student separately. Left as one file, it counts as one student's paper.
- A file can be up to 20 MB, a paper up to 30 pages and a batch up to 1,000 papers.
Where do papers go?
Papers are kept in private storage and sent to AI model providers through OpenRouter to be read and marked. Nothing in storage is publicly readable, and downloads go through the logged-in API.
Files are kept until you delete them, and there is no automatic deletion by age. You can delete per paper or per batch. Deleting removes the files, any typed answers, the report and the saved reading. Your credit history keeps its entries, which hold no student work. Deleting covers what evaleasy stores, and doesn't reach anything a model provider handled while marking.
If your school has rules about where student work can go, check these facts against them before you upload real papers.
Common questions
- Is marking done by AI or by code?
- Both. In evaleasy, an AI model reads every paper and marks written answers and near-misses. Code compares multiple choice, true/false and matching blanks and numbers with your key, and computes every score.
- How accurate is evaleasy?
- evaleasy doesn't publish an accuracy percentage. How well any AI marker does depends on handwriting, scan quality and question style, so the useful test is your own: run it on papers a teacher has already marked and compare question by question.
- What does "worked out" mean in a report?
- It means there was no key for that question, so evaleasy's AI model supplied the correct answer itself. Those are the correct answers to check first. Marking again with them in your key removes the label.
- Can evaleasy mark an essay with no rubric?
- Yes. The AI model marks it under your strictness setting, out of the marks printed on the paper or 10 if none are given. A paper with one unnumbered piece of work is treated as one long answer.
- Do I have to keep the page open while papers are marked?
- No. Once papers are sent, marking runs in the background, so you can leave the page. The dashboard updates by itself when you come back, with no need to refresh.
- Can developers connect to evaleasy?
- Yes. evaleasy has a REST API with signed webhooks (Standard Webhooks), server-sent events, polling and idempotency keys. API keys go in the api-key header. The API documentation is at /docs.
How should you test evaleasy before relying on it?
Test evaleasy on papers a teacher has already marked, then compare the two question by question. Pick papers that include your untidiest handwriting and your hardest question types, because those show you where it needs checking. Our guide explains how to check whether AI marking is accurate enough.
Your next stack of papers doesn't have to be your problem.
Create an account, buy a pack of credits, and start with papers a teacher has already marked.