AI Grading Assistant for Teachers

Grading by Hand Eats the Hours You Do Not Have.

The Problem

In-class handwritten work is some of the most honest work a student produces, and some of the slowest to grade. Students and parents now expect grades that are fast, consistent, and defensible. The studio reads each answer, a scanned handwritten page or a typed submission collected online, and drafts a score against your own rubric, criterion by criterion. You stay in charge on every single paper.

Twenty-eight papers, one at a time, in one sitting

One stack in, sorted by student, and one tick puts the shakiest first

A single stack swallows an evening

Grading runs in the background, so you can close the tab

The last paper rarely gets the attention the first did

The same criteria all the way down the stack

A tired grader at 11pm, with no note of why

A draft score and a short rationale on every criterion

Defensible only if you can still explain it

Your confirmation or override recorded on every line

The Sample

Watch One Paper Get Graded.

Sample Only. Ms. Carla Jiménez, the AP Biology teacher from the Syllabus sample, gave the enzyme-lab short answer from that curriculum to her class and scanned their handwritten pages. Here is how The Rubric! Studio read and scored one student, Jordan Vega, a fictional student invented for this sample (the same student whose recommendation is drafted in the Letter sample). This is the full sample that used to sit on its own page, now built right into this one.

AP Biology Short Answer 3 Criteria · 10 Pts 22 Students Matched 2 Overrides Recorded 2 of 66 Drafted Scores Changed

The Job

The Question and the Rubric

AP Biology Short Answer 3 Criteria · 10 Pts

Question Given to Students: An enzyme's reaction rate climbs with temperature, peaks, then falls off sharply. Explain what is happening at the molecular level across that whole range, and predict what a graph of rate versus temperature looks like past the peak.

The Rubric Carla Built

Mechanism (4 pts): explains collision frequency rising with temperature, then denaturation past the optimum.
Prediction (3 pts): describes the curve past the peak and ties it to the mechanism.
Use of Terms (3 pts): uses active site, denature, and optimum correctly, not as decoration.

The Scan and the Scores

One Paper, Criterion by Criterion

Select a criterion on the right to trace the sentences it scored in Jordan's scan, or one of the quotes under it to light the exact words the studio cited and checked against the transcript. In the studio, every quote that passed the check is lit the same way inside the answer.

Check This Deduction. On another paper in this stack the studio flagged its own draft: The rationale names "substrate" as missing, and the transcription contains it. That criterion came back Low, the row went to Needs Review, and Accept All left it alone; the draft had deducted a point for it, and Carla restored the point with the transcript open.

The Scan, Transcribed

Jordan Vega – Enzyme Lab, Q3 As temp goes up the molecules move faster so they bump into the active site more often and the rate goes up. But after the best temp (the optimum) the heat shakes the enzyme so hard the active site changes shape; it denatures, and the substrate cant fit anymore. So the graph goes up to a peak then drops fast, and it doesnt come back down slow like it went up, it falls off a cliff because once the shape is wrecked its wrecked.

The studio transcribes the handwriting first, so you can read a hard-to-decipher page as clean text while you check the score.

Drafted, Yours to Change

Mechanism 4 / 4

High Confidence Both halves are on the page: "bump into the active site more often" as temperature rises, then "the heat shakes the enzyme so hard the active site changes shape" past the optimum. The shape change is given as the reason the rate falls, which is the mechanism the criterion asks for.

Evidence: bump into the active site more often the heat shakes the enzyme so hard the active site changes shape Evidence Checked

Prediction 3 / 3

High Confidence The curve past the peak is there, "goes up to a peak then drops fast", and "it falls off a cliff because once the shape is wrecked its wrecked" ties the asymmetry to the mechanism. Both parts of the criterion are in the student's own words.

Evidence: goes up to a peak then drops fast it falls off a cliff because once the shape is wrecked its wrecked Evidence Checked

Use of Terms 2 / 3

Medium Confidence All three terms are here and used correctly: "the active site" twice, "the optimum" and "it denatures" once each. Drafted a 2 rather than a 3 because the shape change is explained in plain words with "denatures" attached to it once, and a strict reading of "not as decoration" could ask the term to carry more of the explanation; a 3 is defensible.

Evidence: the active site the optimum it denatures Evidence Checked

Override Note, Recorded → 3 / 3. "Attaching denatures to the exact moment the shape changes is the opposite of decoration. The understanding is in the plain words, and the term is pinned where it belongs. Full marks." The studio drafts; the teacher decides, and what the studio records on this criterion is her note, in place of its own rationale, so the comment pass reads her reason and not its guess.

No Override Recorded. The Studio's Drafted 2 / 3 Stands.

A medium pill is the studio asking for a read: Accept All High-Confidence never takes a row with one, so it waits for you.

As Carla Recorded It

Final Score10 / 10

As the Studio Drafted It

Final Score9 / 10

The studio drafts; the teacher decides.

How Carla Checked 22 Papers

Flagged First, Uncertain Next, Confident Last

Flagged First. Four papers came back flagged: a low-confidence read, a page the studio could not read cleanly, or a rationale whose evidence failed the check against the transcript. Sort by Confidence put them at the top, and she read those with the transcript open; on the one flagged for a term it had called missing, she restored the point.
Uncertain Next. Rows a medium pill kept out of Accept All came next, Jordan's among them. Each is a deduction the studio was not sure of, so the rationale is the thing to read against the transcript; on Jordan's she disagreed, recorded a 3, and left a note saying why.
Confident Last. The rest came back high on every criterion, with a clear page and a sure roster match. Accept All High-Confidence recorded those drafts in one click, and every one stayed open to change.

None of this is a measured error rate. It is the order the review table gives you, and when the last row was finalized the page printed her own count: You changed 2 of 66 drafted scores across 22 papers, both of them on criteria the studio had marked Medium or Low. Your first grading job is on us, so you can run a stack of your own and count.

The Feedback Comment (Optional)

One Comment, Tailored per Student

Carla set a short comment for each score band once; the studio then wrote each student a personalized version. Jordan's was first written against the drafted 9 of 10. When her override moved the total, the row said so ("The Score Changed After This Comment Was Written"), and Refresh Comments for Changed Scores rewrote it against the recorded 10 of 10, reading the note she left on Use of Terms as the reason for that score. Jordan's, from the top band, as refreshed:

"Excellent, Jordan. You nailed the two-part mechanism and, more impressively, explained why the curve is asymmetric instead of just drawing it. Your vocabulary did real work too: active site, optimum, and denatures each land exactly where the mechanism needs them. This is the write-up I'll be showing next period."

The Class Roster

Two Ways to Collect the Stack

Scanned Pages

Carla pasted her roster once, then uploaded a photo of each student's page. The studio read the name off each scan and matched it to the roster, flagging anything it wasn't sure about for her to confirm.

Jordan Vega ✓ Sofia Reyes ✓ Marcus Webb ✓ Aisha Khan ✓ + 18 More

Fictional students, invented for this sample.

JPG / PNG Photos Multi-Page PDF, Split in Your Browser

Collected Online

Prefer no scanning? Run the same job in Collect-online mode. Every student on your roster gets a private writing link and a due date, and opens it with an EduFinder account of their own, which never needs a membership. Pasting is off and the editor saves as they write, so what you grade is their own typing. Light signals like active writing time arrive as conversation starters, never verdicts.

Private Per-Student Links Due Date You Set A Student Account of Their Own

Maya Chen · Typed Online

Submitted 9:41 PM

Report a Mistake

The Evidence

How Close the Studio's Drafts Come to Two Trained Raters.

How It Is Measured

A Public Dataset, Scored by Two People

ASAP-SAS is the short answer set from the Hewlett Foundation's Automated Student Assessment Prize, published on Kaggle in 2012. Its training file holds 17,043 answers written by students in grades 8 and 10 to ten questions in science, biology, English, and English language arts, and two trained raters scored every answer from the question's published scoring guide. The first rater's score, Rater 1 in the tables below, is the dataset's official one; Rater 2 is the second.

Agreement is reported as quadratic weighted kappa, the standard measure for this dataset: 1 is perfect agreement, 0 is what chance alone would give, and a score one point off costs less agreement than a score two points off. Beside the studio's figure is the same measure between the two raters themselves, on the same answers and across every answer to the question, which is how closely a second careful person agrees. The last row of each table is the whole run: its kappas average the questions' kappas, and its percentages are taken over every answer.

Measured October 13, 2026

800 Answers to Eight Public Questions

ASAP-SAS 800 Answers Balanced Tier Mapping B After the Evidence Check

We drew 100 answers at random from each of the eight questions in ASAP-SAS, 800 in all, and the studio scored each one through the same code that scores answers collected online: the same prompt, the same structured output, and the same single repair attempt, on the balanced tier. Each question became a rubric with one 1-point criterion per key element its published scoring guide lists, in the guide's own words, scored present or absent under a note saying there is no partial credit; the guide's own rule for counting elements into a score was then applied in code, so the score compared with the raters' is on their scale, and the question students saw became the job's question as before.

How Each Guide Became Criteria

Question	Elements	The Guide's Rule for the Score
1. Acid Rain	8	The score is the number of pieces of needed information credited, up to 3 (the guide's 3-point rubric gives 3 for three pieces, 2 for two and 1 for one).
2. Polymer Investigation	6	The score is the guide's: a valid conclusion with two improvements is 3, a conclusion with one improvement or two improvements without a conclusion is 2, and a conclusion alone or one improvement alone is 1.
5. Protein Synthesis	8	The score is the guide's: four key elements is 3, three is 2, one or two is 1, and none is 0.
6. Cell Membrane	11	The score is the number of key elements credited, up to 3 (the guide gives 3 for three key elements, 2 for two and 1 for one).
7. Trait of Rose	2	The score is the guide's: no credit without a creditable trait; 1 for the trait, and an additional point for a creditable supporting detail.
8. Mr. Leonard	2	The guide scores the requirements of the task, which the question states: all of them is 2 (Proficient), some is 1 (Partially Proficient), and none is 0.
9. Organization of Article	2	The guide scores the requirements of the task, which the question states: all of them is 2 (Proficient), some is 1 (Partially Proficient), and none is 0.
10. Doghouse	3	The score is the guide's: a color with both its effect inside the doghouse and a result from the experiment is 2, a color with one of the two is 1, and anything else is 0.

Agreement, as Quadratic Weighted Kappa

Question	Studio vs. Rater 1	Studio vs. Rater 2	Rater 1 vs. Rater 2	Raters, All Answers
1. Acid Rain	0.81	0.81	0.96	0.94
2. Polymer Investigation	0.56	0.63	0.89	0.91
5. Protein Synthesis	0.59	0.60	0.94	0.95
6. Cell Membrane	0.70	0.70	1.00	0.96
7. Trait of Rose	0.46	0.45	0.99	0.97
8. Mr. Leonard	0.45	0.52	0.87	0.86
9. Organization of Article	0.54	0.55	0.74	0.83
10. Doghouse	0.72	0.71	0.91	0.88
All Eight Questions	0.61	0.62	0.91	0.91

Exact Matches and Misses

Question	Range	Answers	Studio Matches Rater 1	Rater 2 Matches Rater 1	Studio Off by 2 or More
1. Acid Rain	0 to 3	100	60%	91%	2%
2. Polymer Investigation	0 to 3	100	54%	84%	10%
5. Protein Synthesis	0 to 3	100	84%	94%	3%
6. Cell Membrane	0 to 3	100	80%	100%	3%
7. Trait of Rose	0 to 2	100	62%	98%	15%
8. Mr. Leonard	0 to 2	100	56%	86%	9%
9. Organization of Article	0 to 2	100	57%	75%	3%
10. Doghouse	0 to 2	100	79%	91%	2%
All Eight Questions	0 to 2 or 0 to 3	800	67%	90%	6%

Of the 800 drafts, 532 matched the first rater exactly, 46 were one point low and 175 one point high, and 4 were two or more points low and 43 two or more points high. The studio reported high overall confidence on 528 of the 800 drafts, medium on 271 and low on 1; it routed 1 to Needs Review, and none took its one repair attempt.

Before its one repair attempt, the evidence check flagged none of the 800 drafts for quoting words the answer does not contain or naming a term as missing that it holds; after it, none still carried a flag.

Accept All High-Confidence would have recorded 499 of these 800 drafts unread, the ones that came back complete and high on every confidence the studio reports; of those, 373 (75%) matched the first rater exactly and 19 (4%) were two or more points off.

To check a figure yourself, download every answer's result (each answer id with what the studio drafted for it) and join it to the raters' scores in your own copy of ASAP-SAS.

On every one of the eight questions, the studio sits below the two raters: its draft agreed with the first rater less closely than the second rater did.
Each answer was scored element by element against its question's published guide, and the guide's own rule turned the count of elements credited into the score, so this measures whether the studio credits the elements the raters did; a draft can still land on the right score with the wrong elements credited, which the score alone cannot show.
The answers were scored as typed text, so these figures say nothing about reading handwriting from a scan.
The answers come from a public dataset released in 2012 that may sit in a modern model's training data, so read these figures as an optimistic bound.
Questions 3 (Koala/Panda) and 4 (Invasive Species) were left out: each of their published guides describes one dimension of reading at three levels rather than elements to count, which no count can stand in for.
Some of the questions' answers were handwritten and later transcribed, so they may carry transcription errors.
Question 10's answers carry stray spaces inside some words, an artifact of the dataset.
Question 10's guide prints its data table as a picture, so the table was typed into the question, as a teacher entering it would.

Measured with scripts/benchmark-rubric-scoring.ts at commit 7749bd9c, from train_rel_2.tsv with seed 20261013. A run is never edited once it is published; a later run is added above it, with its own date.

Measured October 13, 2026

1,000 Answers to Ten Public Questions

ASAP-SAS 1,000 Answers Balanced Tier Mapping A After the Evidence Check

We drew 100 answers at random from each of the ten questions in ASAP-SAS, 1,000 in all, and the studio scored each one through the same code that scores answers collected online: the same prompt, the same structured output, and the same single repair attempt, on the balanced tier. Each question became a rubric with one criterion worth the question's top score, whose description is the question's published scoring guide word for word, and the question students saw, with its reading passage or data, became the job's question.

Agreement, as Quadratic Weighted Kappa

Question	Studio vs. Rater 1	Studio vs. Rater 2	Rater 1 vs. Rater 2	Raters, All Answers
1. Acid Rain	0.83	0.82	0.96	0.94
2. Polymer Investigation	0.74	0.76	0.89	0.91
3. Koala/Panda	0.52	0.50	0.77	0.76
4. Invasive Species	0.38	0.42	0.67	0.75
5. Protein Synthesis	0.62	0.62	0.94	0.95
6. Cell Membrane	0.65	0.65	1.00	0.96
7. Trait of Rose	0.40	0.39	0.99	0.97
8. Mr. Leonard	0.57	0.52	0.87	0.86
9. Organization of Article	0.62	0.59	0.74	0.83
10. Doghouse	0.68	0.67	0.91	0.88
All Ten Questions	0.60	0.59	0.87	0.88

Exact Matches and Misses

Question	Range	Answers	Studio Matches Rater 1	Rater 2 Matches Rater 1	Studio Off by 2 or More
1. Acid Rain	0 to 3	100	66%	91%	2%
2. Polymer Investigation	0 to 3	100	66%	84%	3%
3. Koala/Panda	0 to 2	100	54%	80%	1%
4. Invasive Species	0 to 2	100	61%	77%	0%
5. Protein Synthesis	0 to 3	100	71%	94%	1%
6. Cell Membrane	0 to 3	100	76%	100%	6%
7. Trait of Rose	0 to 2	100	54%	98%	15%
8. Mr. Leonard	0 to 2	100	60%	86%	1%
9. Organization of Article	0 to 2	100	69%	75%	0%
10. Doghouse	0 to 2	100	77%	91%	2%
All Ten Questions	0 to 2 or 0 to 3	1,000	65%	88%	3%

Of the 1,000 drafts, 654 matched the first rater exactly, 97 were one point low and 218 one point high, and 4 were two or more points low and 27 two or more points high. The studio reported high overall confidence on 714 of the 1,000 drafts, medium on 285 and low on 1; it routed 1 to Needs Review, and none took its one repair attempt.

Before its one repair attempt, the evidence check flagged none of the 1,000 drafts for quoting words the answer does not contain or naming a term as missing that it holds; after it, none still carried a flag.

Accept All High-Confidence would have recorded 714 of these 1,000 drafts unread, the ones that came back complete and high on every confidence the studio reports; of those, 505 (71%) matched the first rater exactly and 14 (2%) were two or more points off.

To check a figure yourself, download every answer's result (each answer id with what the studio drafted for it) and join it to the raters' scores in your own copy of ASAP-SAS.

On every one of the ten questions, the studio sits below the two raters: its draft agreed with the first rater less closely than the second rater did.
Each answer was scored as a whole against its question's published guide, so this measures how often a drafted score lands where the raters' did, not whether every line of a rationale is right.
The answers were scored as typed text, so these figures say nothing about reading handwriting from a scan.
The answers come from a public dataset released in 2012 that may sit in a modern model's training data, so read these figures as an optimistic bound.
Some of the questions' answers were handwritten and later transcribed, so they may carry transcription errors.
Question 10's answers carry stray spaces inside some words, an artifact of the dataset.
Question 10's guide prints its data table as a picture, so the table was typed into the question, as a teacher entering it would.

Measured with scripts/benchmark-rubric-scoring.ts at commit 29fad138, from train_rel_2.tsv with seed 20261013. A run is never edited once it is published; a later run is added above it, with its own date.

Measured October 13, 2026

1,000 Answers to Ten Public Questions

ASAP-SAS 1,000 Answers Balanced Tier Mapping A Before the Evidence Check

Agreement, as Quadratic Weighted Kappa

Question	Studio vs. Rater 1	Studio vs. Rater 2	Rater 1 vs. Rater 2	Raters, All Answers
1. Acid Rain	0.85	0.85	0.96	0.94
2. Polymer Investigation	0.75	0.76	0.89	0.91
3. Koala/Panda	0.51	0.50	0.77	0.76
4. Invasive Species	0.38	0.47	0.67	0.75
5. Protein Synthesis	0.63	0.62	0.94	0.95
6. Cell Membrane	0.73	0.73	1.00	0.96
7. Trait of Rose	0.44	0.43	0.99	0.97
8. Mr. Leonard	0.57	0.52	0.87	0.86
9. Organization of Article	0.62	0.59	0.74	0.83
10. Doghouse	0.72	0.72	0.91	0.88
All Ten Questions	0.62	0.62	0.87	0.88

Exact Matches and Misses

Question	Range	Answers	Studio Matches Rater 1	Rater 2 Matches Rater 1	Studio Off by 2 or More
1. Acid Rain	0 to 3	100	69%	91%	1%
2. Polymer Investigation	0 to 3	100	66%	84%	2%
3. Koala/Panda	0 to 2	100	54%	80%	1%
4. Invasive Species	0 to 2	100	61%	77%	0%
5. Protein Synthesis	0 to 3	100	71%	94%	1%
6. Cell Membrane	0 to 3	100	81%	100%	4%
7. Trait of Rose	0 to 2	100	58%	98%	14%
8. Mr. Leonard	0 to 2	100	60%	86%	1%
9. Organization of Article	0 to 2	100	69%	75%	0%
10. Doghouse	0 to 2	100	78%	91%	1%
All Ten Questions	0 to 2 or 0 to 3	1,000	67%	88%	3%

Of the 1,000 drafts, 667 matched the first rater exactly, 95 were one point low and 213 one point high, and 3 were two or more points low and 22 two or more points high. The studio reported high overall confidence on 759 of the 1,000 drafts, medium on 238 and low on 3; it routed 3 to Needs Review, and none took its one repair attempt.

Accept All High-Confidence would have recorded 759 of these 1,000 drafts unread, the ones that came back complete and high on every confidence the studio reports; of those, 539 (71%) matched the first rater exactly and 17 (2%) were two or more points off.

To check a figure yourself, download every answer's result (each answer id with what the studio drafted for it) and join it to the raters' scores in your own copy of ASAP-SAS.

On every one of the ten questions, the studio sits below the two raters: its draft agreed with the first rater less closely than the second rater did.
Each answer was scored as a whole against its question's published guide, so this measures how often a drafted score lands where the raters' did, not whether every line of a rationale is right.
The answers were scored as typed text, so these figures say nothing about reading handwriting from a scan.
The answers come from a public dataset released in 2012 that may sit in a modern model's training data, so read these figures as an optimistic bound.
Some of the questions' answers were handwritten and later transcribed, so they may carry transcription errors.
Question 10's answers carry stray spaces inside some words, an artifact of the dataset.
Question 10's guide prints its data table as a picture, so the table was typed into the question, as a teacher entering it would.

Measured with scripts/benchmark-rubric-scoring.ts at commit 311bcfe3, from train_rel_2.tsv with seed 20261013. A run is never edited once it is published; a later run is added above it, with its own date.

Six Steps. From One Paper or a Whole Class to Grades You Can Defend.

You own five of the six steps. The studio does exactly one: the first-pass draft.

Your Step The Studio's Step

Set Up

You

Create a Grading Job.

One question, plus your rubric. Each criterion has a label, its maximum points, and (optionally) a short description of what earns full marks.

You

Add Your Roster.

Paste your class roster, one name per line, optionally with a comma and a student ID. Save it as a class and your other teacher tools can reuse the same roster. A job built from an assessment in a class-linked curriculum arrives with that class already picked, so its roster copies over without a paste.

Collect

You

Add the Answers.

Upload scans as JPG or PNG photos, or a per-student PDF that becomes that student's answer, split into page images in your browser (up to four pages each). Or drop the whole set in as one class stack and confirm how the pages group by student. For work no scanner can read, Score by Hand creates one blank row per student on the roster instead, and stores no student work at all. Or have the job collect typed answers online, where every student on your roster gets a private writing link and a due date you set.

Grade

The Studio

Start Grading.

It runs in the background, so you can close the tab and come back later. Tick Sort by Confidence on the review and the shakiest answers come first. Once you have finalized a few, it can rescore whatever you have not reviewed yet against your own first three, and never changes a score you already recorded.

Decide

You

Review Each Paper.

The scan or typed submission, the student it was matched to, a transcript of what was written, and a draft score with a rationale on every criterion.

You

Confirm, Release, and Download.

Confirm or override each line, then export the results as a CSV, an Excel workbook, or a printable PDF grade sheet. Take the confident ones in one click with Accept All High-Confidence, sixty rows at a time. A medium or low pill on any criterion keeps a row out of that click, so what you accept unread is only what came back clean on every signal. Release grades to the students who submitted online with an account, which sends the per-criterion points, the total, and your comment, and never the transcription, the rationale, or the confidence. Download a Google Classroom Sheet or a Generic Gradebook CSV beside those three. The recorded grade is the one you decided, not what the tool guessed.

Want the reading first? The guide to writing recommendation letters explains how a letter that informs rather than praises is written from collected material, which is what a graded job carries into The Letter! Desk.

Grade the Whole Class at Once.

One stack goes in. What comes back is a review you can sort so the shakiest answers come first, grades you release when you are ready, and a read on what the class as a whole still needs, which a job built from one of your curricula can feed straight back into that week.

One Class Stack Sorted by Student You Confirm the Groups Scored, Yours to Review You Accept or Override Released and Downloaded Class Insight Brief Re-Plan Week N

One Stack, Sorted Before Anything Is Graded

Ask students to write their name, and their ID if you use one, in the top-left corner of their first page, and keep each student's pages together in the pile
Drop the whole class set in as one PDF or a pile of photos, up to 240 pages a stack and four pages a student, and the pages are grouped into one student's answer at a time
Groups come back worst first, a page that landed with the wrong student moves in one click, and a name that matched nobody on your roster is left for you to pick rather than guessed at
Nothing is graded until you press Confirm Groups, and every grade that follows is still yours to review and finalize

Score by Hand, for Work No Scanner Can Read

A project, a performance, or an oral defense is graded in the room, so there is no page to scan and nothing to transcribe
Set Up Hand Scoring creates one blank row per student on your roster, and every one of those rows reads Scored by Hand
No model runs on a hand-scored job and no student work is stored: the studio holds your rubric, your roster, and the points you record
You enter and finalize each grade in the same review table you would use for scans, and export it the same way

Nothing here emails anyone, releases a grade, or records a decision without your own click, and nothing is shared that a student did not agree to.

What You Get on Every Paper.

What You Give It

Your Question Your Rubric Your Roster Scans or Online Answers

What Comes Back

Transcript Draft Score per Criterion A Rationale for Each Your Confirmed Grade CSV / Excel / PDF

Two Ways to Collect Work

Upload scans as JPG or PNG photos, or a per-student PDF that becomes that student's answer, split into page images in your browser (up to four pages each, and several photos can combine into one answer)
Or run the job in Collect-online mode: every student gets a private writing link and a due date you set, and opens it with an EduFinder account of their own, which never needs a membership
The online editor saves as they write and pasting is turned off, so what you grade is the student's own typing
Watch an online collection fill in live: who hasn't opened it, who's writing, and who submitted or ran late. Extend, reopen, or lock any one student's link
Scanned pages are matched to your roster by the name and student ID written on them, and an uncertain match is flagged, not guessed
Your rubric can be typed in directly or imported from a Word, Excel, or typed PDF file

What Comes Back on Every Paper

For scans, a transcript read from the page itself: handwritten prose, the math it reads written out, and what a diagram or sketch shows (typed answers are graded as submitted)
A draft score and a short rationale on every criterion, tied to the wording of your rubric
A confidence pill on every criterion, so a deduction the studio was not sure of says so before you read it
The same criteria applied all the way down the stack, so the tenth answer gets the same standard as the first
Light writing-process signals on typed submissions, like active writing time, as conversation starters, never conclusions
A review you can sort by confidence, worst first, plus a rescore of everything still unreviewed against your own first three finalized answers, which never touches a score you already recorded
Accept All High-Confidence records the drafts where every signal came back clean, sixty rows a click, and each of those rows stays open for you to change afterwards
A one-way release to the students who submitted online with an account: the per-criterion points, the total, and your comment, and never the transcription, the rationale, or the confidence
Your confirmation or override on every line, then export to CSV, Excel, or a printable PDF grade sheet. It is a consistency aid, checked by you, not a black box that hands out marks

From the Sample Job · One Rubric, Ten Papers

The Tenth Paper Is Held to the First Paper's Standard.

Mechanism4 Pts

Prediction3 Pts

Use of Terms3 Pts

Full Marks on That Criterion Partial Credit Little or None You Overrode the Draft

Every cell is a draft until you confirm it, and every cell is yours to change. The ringed one is Carla's override in the sample above: the studio drafted 2 of 3 on Use of Terms under a medium pill, she read the rationale against the transcript, recorded 3, and 3 is the grade the studio keeps.

100pages per job, student by student 240pages in one class stack 4pages per student 5downloads: CSV, Excel, PDF, and two gradebooks 60rows per Accept All click

Scans and typed submissions are stored privately in your workspace. Grading runs through an AI provider's API under terms that do not use your data to train their models, and we never use student work to train our own.

The Rubric! Studio is FERPA-aware and operates under your institution's direction; you remain the decision-maker on every grade. Built for teachers at US and international high schools, and for the department chairs and academic leaders who want consistent, defensible grading across every section.

One Class, Three Tools.

One fictional class runs through all three tools. Keep a class in your account and they share it: plan the course once, grade with its roster, and recommend its students with the work they actually did.

1Plan The Syllabus! Planner You Are Here

Unit 1 Assessmentbecomes a ready-to-review grading job

2Grade The Rubric! Studio You Are Here

Jordan's Graded Labis carried into the letter in words, never as a raw score

3Recommend The Letter! Desk You Are Here

How the Handoffs Work

From This Class, End to End

Keep a class here once (a name and a roster) and all three teacher tools share it.
The Syllabus! Planner drafts the course, and each assessment hands off to this studio as a ready-to-review grading job: scan-gradable written work with the questions and a rubric to score from your scans, live-graded work with a task brief and a rubric to score by hand.
Pick your class on a new job and its roster copies over, so you skip the paste step.
When you finish reviewing a student's grade, the row offers to draft their recommendation letter in The Letter! Desk, described in words, never as a raw score.

How This Sample Was Made

One Click between Each Tool

The question and rubric here came straight from Carla's curriculum in The Syllabus! Planner with one click. She graded the class against them here, with her final say on every score. And because Jordan is a student in this class, that graded work is ready as evidence when Carla writes a recommendation letter in The Letter! Desk, described in words, never as a raw score.

Fair Questions, Straight Answers.

What we would want to know before trusting a grading tool.

How Much Does the Rubric! Studio Cost?

Your first grading job is on us on any account, up to ten students, with its comment pass. After that, grading is part of the Pro Membership, $39 per month (USD), with no per-page charge on top of it and one membership covering every tool on your account. Uploading student by student, a job holds up to 100 pages; a class stack is a separate intake with its own ceiling of 240 pages. Either way, one student's answer is at most four pages.

What Can I Upload?

Upload your scanned answers student by student, as photos or as a per-student PDF, or drop the whole class set in at once as one class stack: a single PDF or a pile of photos that the studio sorts into students for you to confirm. Each photo is one page, and each page of a PDF is one page; a failed upload is never counted. Prefer no uploads at all? A job can instead collect typed answers online through private per-student links.

Can I Grade a Whole Class in One Pass?

Yes. Upload the whole class set as one class stack and the studio groups the pages into one student's answer at a time, reading the name written on each first page. You check the groups on the sort screen, which shows the shakiest ones first: move a page that landed with the wrong student, split a group that runs long, and pick a student from the roster wherever a name matched nobody. Nothing is graded until you press Confirm Groups, and every grade that follows is still yours to review and finalize. Ask students to write their name, and their ID if you use one, in the top-left corner of their first page, and keep each student's pages together in the pile.

Can It Grade Work a Scanner Cannot Read?

Yes. Score by Hand covers the assessment there is no page for: a project, a performance, an oral defense. Set Up Hand Scoring creates one blank row per student on your roster, each marked Scored by Hand, and you enter and finalize every grade in the same review table you would use for scans. No model runs on a hand-scored job and no student work is stored; the studio holds your rubric, your roster, and the points you record.

Can Students Type Their Answers Online Instead?

Yes. Create a job in Collect-online mode and each student on your roster gets a private writing link with a due date you set; they open it with an EduFinder account of their own, which never needs a membership and is also what lets you release their grade back to them afterwards. Typed submissions land in the same review as scans, and you also see light writing-process signals, such as active writing time, presented as conversation starters, not conclusions. As they work, a live view shows who hasn't opened the link, who's writing, and who has submitted or run late, and you can extend, reopen, or lock any student's link individually.

Is My Students' Work Safe? What about FERPA?

Scans are stored privately in your workspace. Grading runs through an AI provider's API under terms that do not use your data to train their models, and we never use student work to train our own. The Rubric! Studio is FERPA-aware and operates under your institution's direction.

Does the AI Decide My Students' Grades?

No. It drafts a score and a short rationale for each rubric criterion, and you review every paper, adjust anything you disagree with, and decide the final grade. Accept All High-Confidence is the one shortcut, and it is still your decision: it records the drafted score on the rows where every signal came back clean, up to sixty at a click, and each of those rows stays open for you to change afterwards. The recorded grade is the one you decided, not what the tool guessed.

How Often Is a Drafted Score Wrong?

On October 13, 2026, the studio drafted a score for each of 1,000 short answers from ASAP-SAS, a public dataset that two trained raters had already scored: its draft matched the first rater's score exactly 65% of the time, where the two raters matched each other 88% of the time. Drafts were two or more points off the first rater on 3% of answers (31 of the 1,000). On quadratic weighted kappa, the agreement measure this dataset is scored by (1 is perfect agreement and 0 is chance), the studio's drafts reached 0.60 against the first rater, where the two raters reached 0.87 against each other on the same answers. On the same 1,000 answers, scored twice on October 13, 2026, exact matches with the first rater went from 67% (667) to 65% (654), drafts two or more points off from 3% (25) to 3% (31), and the kappa against the first rater from 0.62 to 0.60. The one change to how the studio scores between the two runs was the evidence check. Two runs of the same answers never land exactly alike, so part of any difference between two runs is run-to-run variation rather than a change in the studio. Every drafted score now cites the words of the answer that earned it and the terms it says are missing, and the studio checks both against the answer before the draft is stored: what fails is flagged on the row as Check This Deduction or Check This Evidence, the criterion is marked Low, and the row waits for you in Needs Review. Before its one repair attempt, the evidence check flagged none of the 1,000 drafts for quoting words the answer does not contain or naming a term as missing that it holds; after it, none still carried a flag. Once every row of a job is finalized, Review Grades prints your own count in one line, You changed N of M drafted scores across K papers. The answers were scored as typed text, so these figures say nothing about reading handwriting from a scan. The answers come from a public dataset released in 2012 that may sit in a modern model's training data, so read these figures as an optimistic bound. The Evidence section on this page has the method, a row for each of the ten questions, and the date and size of every run. Your first grading job is on us, so you can count what you change on a stack of your own.

What File Formats Can I Upload?

Student answers go in as JPG or PNG photos, or as a per-student PDF that becomes that student's answer, split into page images in your browser (up to four pages each). A class stack takes the same photo and PDF formats, split in your browser the same way. Your rubric can be typed in directly or imported from a Word, Excel, or typed PDF file.

Does It Work with the Other Teacher Tools?

Yes. Plan a course in The Syllabus! Planner and every assessment has a one-click handoff into this studio that drafts a ready-to-review grading job, scan-gradable written work as the questions plus a rubric scored from your scans and live-graded work as a task brief plus a rubric you score by hand; the curriculum then tracks each job's status, and every job links back. When that curriculum is linked to a class, the handoff picks the class too, so its roster rides along without a paste, and any class kept in your account can be picked on a new job for the same reason. Work goes back the other way once half a class is finalized: a Class Insight Brief reads what the class as a whole still needs, without naming a student, and from it Add to Next Week's Plan files the reteach lesson as that week's first meeting while Re-Plan Week N drafts that week's daily plans around the brief. And once a student is matched to your class, their row offers to draft their recommendation letter in The Letter! Desk, with that graded work carried along as evidence.

Do I Need My School's Approval to Use It?

No. Your own account is the whole setup: no institutional license, no IT ticket, no waitlist, and no roster sync, because you type or paste your class roster yourself. The studio is FERPA-aware and operates under your institution's direction, so handle student work the way your school's policies already require. Every recorded grade stays yours.

One Stack. Grades You Can Defend.

Grade one stack tonight.

The Rubric! Studio

A Whole Class Graded against Your Own Rubric.

Handwritten scans, or typed answers collected online through private per-student links
Handwriting, math, and diagrams transcribed straight from the page
A draft score and a rationale on every rubric criterion, in your own wording
Every paper matched to the right student, with uncertain matches flagged, not guessed
Your confirmation or override on every line, then export to CSV, Excel, or a printable PDF grade sheet
One class stack sorted into students, with Accept All High-Confidence to take the clean rows sixty at a time
A Class Insight Brief once half the class is finalized, and one click to re-plan that week of your curriculum around it
Reuse the same rubric on another class in one click, and retry any paper that didn't finish
One shared class across the Syllabus! Planner, this studio, and The Letter! Desk

Running a tool again and updating what it made are part of the Pro Membership.

Free

Your first grading job is on us, up to ten submissions with their comments.

$39/mo

Pro Membership

Per month in USD. One membership covers every tool on your account.

Open the Rubric! Studio

No school sign-off needed: your own account grades tonight, with no IT ticket and no roster sync.

Grade a Whole Class Stack

Grading by Hand Eats the Hours You Do Not Have.

Watch One Paper Get Graded.

The Question and the Rubric

One Paper, Criterion by Criterion

Flagged First, Uncertain Next, Confident Last

One Comment, Tailored per Student

Two Ways to Collect the Stack

How Close the Studio's Drafts Come to Two Trained Raters.

A Public Dataset, Scored by Two People

800 Answers to Eight Public Questions

1,000 Answers to Ten Public Questions

1,000 Answers to Ten Public Questions

Six Steps. From One Paper or a Whole Class to Grades You Can Defend.

Create a Grading Job.

Add Your Roster.

Add the Answers.

Start Grading.

Review Each Paper.

Confirm, Release, and Download.

Grade the Whole Class at Once.

One Stack, Sorted Before Anything Is Graded

Score by Hand, for Work No Scanner Can Read

What You Get on Every Paper.

Two Ways to Collect Work

What Comes Back on Every Paper

The Tenth Paper Is Held to the First Paper's Standard.

One Class, Three Tools.

From This Class, End to End

One Click between Each Tool

Fair Questions, Straight Answers.

How Much Does the Rubric! Studio Cost?

What Can I Upload?

Can I Grade a Whole Class in One Pass?

Can It Grade Work a Scanner Cannot Read?

Can Students Type Their Answers Online Instead?

Is My Students' Work Safe? What about FERPA?

Does the AI Decide My Students' Grades?

How Often Is a Drafted Score Wrong?

What File Formats Can I Upload?

Does It Work with the Other Teacher Tools?

Do I Need My School's Approval to Use It?

One Stack. Grades You Can Defend.

A Whole Class Graded against Your Own Rubric.