Royal Coast 
A Royal Coast Labs product
Caddie drafts grading feedback, and the instructor approves every word.
Caddie is an AI workspace for college instructors. It drafts feedback against your own rubric and shows the passage behind every score. Before any AI model reads a student's work, a privacy vault trained on your course replaces your students and the people and places in their lives.
Live at caddiereader.com by invitation. It has not yet read real student work. See two essays go through it.
What it does
Drafts a score and feedback for each rubric criterion, with the passage from the student's own text that justifies it.
What it protects
Students' names, emails and IDs, and the family members, hometowns, workplaces and health details in their writing. You see the original; the model sees tokens.
What it costs you
Your roster and syllabus once, a short rubric interview, and a few quick questions on essays where Caddie is unsure.
The problem
Right now the choice is time or privacy.
Faculty are already using general-purpose AI to grade, often by pasting student work into a consumer chatbot. It saves real hours on the task that eats the most of them.
Institutions cannot approve that, because student work leaves their control the moment it happens, with names attached. So instructors choose between the help and their students' privacy, and nobody is comfortable with either answer.
Caddie removes that choice: instructors get the help, and student identities stay protected.
Built for higher education
Designed around college courses.
Graduate seminars, case analyses, lab reports, and discipline-specific rubrics are the design center, not an afterthought to a school tool. Caddie learns the shape of your course rather than the subject: it does not need to know chemistry, only what a chemistry lab report looks like.
The privacy vault
Nothing is graded until Caddie knows your course.
A student writing about Daisy Buchanan is not writing about a classmate named Daisy, and a grandmother named Ruth is not the Ruth in your case study. Telling them apart takes knowing the course, so Caddie will not grade an essay until it has your roster and your syllabus or course materials.
Before any grading
Your course
Roster, syllabus, readings and lecture notes, the assignment and its rubric. This teaches the vault which names and places are the subject of the course.
Step 1 · certain
Roster and format rules
Full student names, emails, student IDs, phone numbers and the name line at the top of an essay. These are always replaced, and nothing can override them.
Step 2 · judged in context
A layered redactor
Models trained before deployment judge each line, each sentence, each phrase and each word, against your course: "my cousin Bryan" is private, "Daisy Buchanan" is a character, "my client J." is a person. About a tenth of a second an essay.
Step 3 · only if unsure
One question per word
What the vault cannot decide is held, and the essay waits. "Nick" held five times is one question, and your answer covers that essay only.
Then, and only then
One call to the model
The tokenized essay, sealed by the vault, is the only text that leaves. The draft comes back with tokens, and Caddie puts the names back for you alone.
The protection is structural. The only route to an AI model refuses any text the vault has not sealed, and the vault seals only what it has tokenized. If the redactor fails for any reason, every name it has not decided is held and nothing is sent.
See it work
Two essays, from upload to feedback.
Responses to The Great Gatsby from invented students, run through the live pipeline: the layered redactor, the review questions, and a real grading call. Nothing below was edited by hand.
A strong personal response, full of private detail
What Caddie received
Siobhán O'Connell Prof. Helen Marchetti ENG 210: American Literature II September 18, 2026 Borne Back: Gatsby, Nana Bríd, and the Last Tray of Soda Bread[…]I grew up in Hessel Corners, a hamlet with one blinking yellow light, a Methodist church, and my family's bakery. My grandmother, Bríd Ferriter, opened Ferriter's Bakery on the corner across from the church in 1979, two years after she came over from County Kerry. Everyone in town calls her Nana Bríd, even people who are not related to her. Every Saturday of my childhood I stood on a milk crate next to her at four in the morning, shaping soda bread and getting flour in my hair. For me, that kitchen was my green light. It was the thing I measured everything else against.
What the grading model received
STU-TX5V Prof. Helen Marchetti ENG 210: American Literature II September 18, 2026 Borne Back: Gatsby, REF-5D95, and the Last Tray of Soda Bread[…]I grew up in PLC-QDEL, a hamlet with one blinking yellow light, a Methodist church, and my family's bakery. My grandmother, PER-3X4Z, opened PLC-6YLJ on the corner across from the church in 1979, two years after she came over from PLC-4R8J. Everyone in town calls PER-UETW, even people who are not related to her. Every Saturday of my childhood I stood on a milk crate next to her at four in the morning, shaping soda bread and getting flour in my hair. For me, that kitchen was my green light. It was the thing I measured everything else against.
The instructor's 7 questions
Each held word, once, with the sentences it appears in. Answered here as a careful instructor would.
- Methodist: send as written
- her Nana Bríd: replace it
- Every Saturday: send as written
- driving her to chemotherapy twice: replace it
- Nick: send as written
What came back
The model's reply for one criterion, exactly as returned by gpt-5.4-mini. Its evidence quote is in the text it was sent; Caddie finds it there, maps it onto the student's original, and shows the instructor the student's own words.
Show the raw reply
{
"criterion_id": "c3",
"assessable": true,
"score": 4,
"feedback": "The essay uses personal experience in a specific, relevant way to extend the literary argument rather than replacing it. The bakery story, the disrupted routine, and the effort to recreate the past are all clearly tied to the student’s interpretation of Gatsby’s longing.",
"evidence_quote": "What I did this summer was very Gatsby. I tried to rebuild the past in exact detail, as if getting every piece right would bring it back.",
"confidence": "high"
}What the instructor sees
Connection to the text · 4 of 4
The essay refers to specific moments from The Great Gatsby and connects them to a sustained argument about Gatsby’s attempt to recover the past. The strongest section ties the novel’s language to the student’s reading of Gatsby’s desire and the green light: the discussion moves from the quoted line to a concrete interpretation of how Fitzgerald frames Gatsby’s longing.
Evidence, in the student's words: The part of the novel that hit hardest was actually a small moment in Chapter 5. Once Daisy is finally in his house, Nick says the green light has lost its "colossal significance," and that it is "again a green light on a dock."
Insight · 4 of 4
The response goes beyond identifying parallels and makes a clear interpretive claim about what the novel means: Gatsby’s tragedy is not simple foolishness, but a refusal to accept change. That insight is developed consistently, especially in the claim that the novel is about looking at what has already been lost rather than just wanting the past back.
Evidence, in the student's words: What makes Gatsby tragic is not that he looks back. It is that he refuses to let the present be different from the past.
Show the other 2 criteria
Use of personal experience · 4 of 4
The essay uses personal experience in a specific, relevant way to extend the literary argument rather than replacing it. The bakery story, the disrupted routine, and the effort to recreate the past are all clearly tied to the student’s interpretation of Gatsby’s longing.
Evidence, in the student's words: What I did this summer was very Gatsby. I tried to rebuild the past in exact detail, as if getting every piece right would bring it back.
Organization and clarity · 4 of 4
The piece is well organized: it moves from the opening claim, to personal context, to close textual comparison, and then to a concluding reflection that returns to the central idea. Sentences are clear and the paragraphs build toward the final point about recognizing change rather than trying to preserve the past unchanged.
Evidence, in the student's words: What makes Gatsby tragic is not that he looks back. It is that he refuses to let the present be different from the past.
Show the second essay: a weak close reading
A weak close reading, graded as one
Graded against a close-reading rubric. The essay leans on plot summary, and the feedback says so. There is no name in the text at all: Caddie matched the file to the student on the roster and replaced the people she mentions. The instructor answered 7 questions: sent "Nick", "Daisys", "Gatsbys" as written.
What Caddie received
ENG 210 Essay 1 Close Reading Sept 11 The American Dream in The Great Gatsby[…]The passage I chose is when Gatsby is staring at the green light. At the end of chapter 1 Nick sees Gatsby standing on his lawn and "he stretched out his arms toward the dark water in a curious way". Nick looks and sees "a single green light, minute and far away, that might have been the end of a dock." The green light is on Daisys dock. The green light represents Gatsbys dream. Green is the color of money and also the color of go, like a traffic light. So its like Gatsby is being told to go after his dream. It is also far away which shows the dream is hard to reach.
What the grading model received
ENG 210 Essay 1 Close Reading Sept 11 The American Dream in The Great Gatsby[…]The passage I chose is when Gatsby is staring at the green light. At the end of chapter 1 Nick sees Gatsby standing on his lawn and "he stretched out his arms toward the dark water in a curious way". Nick looks and sees "a single green light, minute and far away, that might have been the end of a dock." The green light is on Daisys dock. The green light represents Gatsbys dream. Green is the color of money and also the color of go, like a traffic light. So its like Gatsby is being told to go after his dream. It is also far away which shows the dream is hard to reach.
What the instructor sees
Arguable close-reading claim · 1 of 4
The opening paragraph gives a broad theme claim about the American Dream, but it is mostly summary and not an arguable close-reading thesis about Fitzgerald’s language in a chosen passage. The best evidence of this weakness is, "The Great Gatsby by F. Scott Fitzgerald is a novel about the American Dream."
Quote-by-quote evidence and analysis · 2 of 4
There are several quotations, but the analysis stays general and does not consistently unpack Fitzgerald’s specific wording, and some quoted material is not closely examined sentence by sentence. The strongest passage for this judgment is, "The green light is on Daisys dock. The green light represents Gatsbys dream. Green is the color of money and also the color of go, like a traffic light."
Show the other 2 criteria
Passage focus and relation to the novel · 1 of 4
The essay begins with the green light passage, but most of the body shifts into plot summary and broad theme discussion about the novel as a whole rather than staying anchored in the chosen passage. The connection to the novel’s ending is relevant, but the passage focus weakens because the essay keeps moving away from the close reading itself. The clearest evidence is, "The passage I chose is when Gatsby is staring at the green light."
Organization and progression of analysis · 2 of 4
The essay follows a basic sequence from claim to example to conclusion, but the progression is loose because several middle paragraphs repeat the same general point about the American Dream instead of building a line of analysis from the passage outward. The strongest evidence is, "This shows the American Dream is not real because you can never get to it."
Measured, not promised
Where the vault is today, including where it falls short.
We measure the vault on invented courses it never trained on, with answer keys written before the first run: one set in subjects like the training courses (25 essays, 235 private details), and one in subjects it has never seen, including essays written in French (24 essays, 474 private details). Both sets are small, so each figure comes with its range.
| Held-out essays | Roster matching alone | Caddie today, familiar subjects | Caddie today, new subjects |
|---|---|---|---|
| Private details that reached the model | 68% | 10% (range 7 to 14%) | 17% (range 12 to 22%) |
| The same, weighted by how identifying each detail is | 9% (range 5 to 12%) | 17% (range 12 to 22%) | |
| Course material wrongly removed (characters, authors, places) | 20% | 0.9% | 0.5% |
| Time for the privacy pass, per essay | under 1 s | about 0.1 s | |
By kind of detail
| Kind | Familiar subjects | New subjects |
|---|---|---|
| Names and direct identifiers | 5% of 153 | 6% of 183 |
| Health | 16% of 25 | 43% of 92 |
| Places, schools and workplaces | 15% of 34 | 7% of 73 |
| Other details, such as client initials | 30% of 23 | 18% of 126 |
Names and direct identifiers weigh most in the weighted figure, and health information as much. The tiers were fixed before the breakdown was first run.
What would close the gap to 2%
- Health, the weakest kind. In subjects where students write about health, such as public health and social work, 43% of health details still reach the model: doses, diagnoses described rather than named, and disease words that are also the course's topic.
- Real writing. Every essay the vault has learned from was invented. De-identified essays from past courses, with consent and an ethics review, are the largest lever.
- Asking more often. The review questions already hold 14 to 17% of private details for the instructor instead of sending them. Asking about more borderline words trades a few more questions for fewer leaks.
None of these is certain. Our target before any real student work is under 2%, measured on larger, harder test sets than these.
Your judgment, not ours
Caddie drafts. You decide.
Your rubric, in your words
A short interview, informed by your syllabus, turns what you actually look for into criteria. It only asks what your syllabus could not answer, and it asks the questions nothing else can: what separates excellent work from adequate here, and what should not count against a student.
Evidence for every judgment
Each score comes with the passage that justifies it, highlighted in the student's own text. A quotation that does not appear in the submission is flagged rather than trusted, so the evidence is checkable rather than decorative.
It says when it cannot tell
A criterion it cannot assess from the submission is marked unassessed, not guessed at. Where the vault's edits may have cost a judgment something, the draft says so and points you at the original.
Nothing reaches a student or a gradebook without your explicit approval. Every data movement and every approval is written to an append-only log that an institution can review.
Where it stands
Honest about the stage.
Caddie is a proof of concept. It is not publicly available and it is not in any classroom. Here is exactly what runs today and what does not.
Working
The privacy vault: roster and format rules, including nicknames, personal links and handles, and the layered redactor for family, clients, places, workplaces and health details
Working
Word, PDF and text uploads for essays, syllabi and materials, and no grading until the course has context
Working
Review questions grouped by word, and the append-only audit log
Working
Your own rubric, or one proposed from the assignment, syllabus and readings, with every level described and your point weights; you edit it before grading
Working
Grading drafts with evidence, checked against your answer key and the readings, with length and missing parts checked exactly
Working
Tokenization, re-identification, and the side-by-side view of what the model saw
Working
A full grading cycle: approve each grade, regrade after changing the rubric, export grades, feedback and the gradebook, and export the audit log
Not yet
Accuracy good enough for real student work (target under 2% of private details reaching the model)
Not yet
Study materials, lesson plans and slide decks
Not yet
Direct Canvas and learning management system integration (grades export as spreadsheets today), institutional sign-on, and email
Design partners
Help us find what is wrong with it.
We are looking for faculty who teach writing-heavy courses, grade real stacks of work, and would use an early version on a real assignment. You would be working with something unfinished, which is exactly why your input matters: what you find now changes what gets built.
Or write to hello@royalcoastlabs.com with "Caddie design partner" in the subject.
If you are reviewing this for an institution
Caddie is designed to support an institution's obligations under FERPA by keeping student identities inside a boundary the vendor controls, rather than by contractual promise alone. The vault removes students on the roster with certainty, and removes other personal details with the measured accuracy shown above, which is not yet good enough for real student work. That is design intent, stated before any legal review, and we will say so plainly until a review is done.
The architecture document, the measurements, and the containment tests are available on request, along with a walkthrough of what crosses the boundary and what does not. Ask for the technical brief.