‘Bluebook’ Is Not a Citation Manual and ‘AP’ Is Not Accounts Payable: Why Test-Prep Centers Bought Tuned Models in 2026

Here is a two-minute test you can run before you finish your coffee. Open whatever general AI assistant your staff already use and type: "What is Bluebook?"

You will get an answer about a legal citation manual. It is a perfectly good answer. It is also the wrong one, because in your building Bluebook is the College Board's testing app — the thing every student sits in front of for the digital SAT, the PSAT/NMSQT, the PSAT 10 and the digital AP exams, and the thing your tutors are talking about when they write "practiced two Bluebook full-lengths this week" in a session note.

Now type: "Summarize AP performance across the center this term." A general model has a real chance of asking you about your accounts payable.

The Words This Trade Uses That a General Model Fumbles

Every trade has a private vocabulary. Test prep has an unusually dense one, because it borrows from four different testing organizations, three different special-education traditions and whatever your state calls its own assessments. A short and incomplete list of the ones that go wrong:

• Module 1 and Module 2. On the digital SAT, how a student does on the first module decides whether the second one is the harder path. A general model reads "Module 2" as a chapter in a training course and writes a progress report that means nothing.
• RW and M. Reading and Writing, and Math — the two section scores, 200 to 800 each. Not "verbal." Not "critical reading," which has not existed since 2016.
• Selection Index. The National Merit number, and the single most common place a general model produces confident nonsense.
• Composite versus section score. On the ACT, composite is an average. On the SAT, the total is a sum. Models blur them.
• Stanine versus percentile. The ISEE reports stanines, 1 to 9. The SSAT reports percentiles against same-grade, same-gender test takers. Tell a Manhattan or Palo Alto family in November that their daughter's 7 is a percentile and you have lost the room.
• RIT. The NWEA MAP growth scale. Not the university in Rochester.
• OG. Orton-Gillingham. Your reading specialists say it forty times a day.
• Superscore, Score Choice, FRQ, ESY, SLD, 504, SHSAT, TACHS, HSPT, TSI, Accuplacer, Regents, STAAR. Each one is an ordinary word in your hallway and a coin flip everywhere else.

A vertical model is an AI system tuned on one trade's own documents and vocabulary, so it treats "Module 2" as the second adaptive section of the digital SAT rather than a chapter heading, and knows that a stanine and a percentile are not the same animal. In 2026 this stopped being a curiosity and became a category, because general models kept missing exactly the things a trade takes for granted.

The Selection Index Example, Because It Costs Real Money

A junior comes back from October with a PSAT/NMSQT: Reading and Writing 700, Math 690. A parent wants to know whether National Merit is realistic in your state.

A general model, reasoning from ordinary arithmetic, averages the two and offers something soothing. The actual Selection Index doubles the Reading and Writing side: two times 70, plus 69, equals 209. Whether 209 clears your state's Semifinalist line is a conversation your Academic Director can have. Whether the number is 209 or the model's guess is not a matter of opinion, and getting it wrong in an email to a parent who has already looked it up on a forum is the kind of small error that ends a $3,500 package renewal.

flowchart TD
  A["Tutor note: 'RW 610 M 590, Bluebook Mod 2 lower path'"] --> B["Tuned model reads the note"]
  B --> C["Checks digital SAT section names and ranges"]
  B --> D["Checks Bluebook adaptive module meaning"]
  B --> E["Checks student's 504 extended-time plan"]
  B --> F["Checks the November test date on file"]
  C --> G["Drafts the parent progress report"]
  D --> G
  E --> G
  F --> G
  G --> H["Academic Director edits and signs"]

What Tuning Actually Fixes — and What It Is Fed

Tuning a model for a test-prep company is not a science project. It is feeding it your own paper: three years of score reports, your session-note conventions, your curriculum names, your diagnostic scoring sheets, your accommodation language, and a current reference on how each exam is actually built this year.

That last part matters more than owners expect, because the tests keep moving. The SAT went fully digital and adaptive. The essay is gone. The guessing penalty has been gone since 2016 and a general model will still occasionally advise a student to skip rather than guess. The ACT's enhanced form made science optional on national test dates, which changes what the composite is made of — and a model trained mostly on older internet writing will tell an ACT family the science section is required, which is both wrong and expensive when your tutor spent six sessions on it.

The other thing tuning fixes is tone. Your progress reports have a house voice: what improved, what the next two weeks target, what the parent should and should not do at home. A tuned model writes in that voice because it has read three hundred of yours. A general model writes like a brochure.

Friday Afternoon, Progress Report Day

Take a center with 180 active students on monthly written progress updates. The Academic Director's Friday used to run like this: open each student's session notes, cross-check the last diagnostic, write six to eight sentences, send. Fifteen minutes each when the notes are good, longer when the tutor wrote "went well."

With a tuned model in front of her, the draft is on screen before she opens the file. It used the right section names, referenced the correct upcoming test date, noted that this student has extended time under a 504, and flagged the three students whose tutors wrote nothing useful. Her job becomes judgement: which parent needs a phone call instead of an email, which student is plateauing because the tutor pairing is wrong.

The Rework Arithmetic

The honest way to value this is not "hours saved on writing." It is rework — the drafts that come back wrong and have to be caught. Assume the following, and substitute your own numbers after a one-week trial:

The $151 is not the point. The point is the bottom row. Two or three wrong statements a month reaching paying parents — a wrong test date, a wrong score conversion, a confident sentence about a guessing penalty that has not existed in a decade — is a credibility leak in a business where the entire product is expertise. One lost 20-hour package at $88 an hour is $1,760, which swamps everything else in the table.

Where a Tuned Model Still Has No Business

It does not know your student. It knows that a 40-point Math gain over six weeks is within normal range; it does not know that this particular sophomore's mother is going through a divorce and that the real reason for the plateau is that he has not slept. Nothing in the file will tell it that. Your tutor knows and your Center Director knows.

It should not be the source of truth for test dates and registration deadlines. Tuned or not, these systems will produce a plausible date. The College Board and ACT calendars are the source of truth; keep them in your own document and let the model read from it rather than from memory.

And it should not write anything that goes into an IEP meeting or a school communication about a student with a disability. A tutoring company sitting at that table is a guest, the language is legally loaded, and the person who signs the summary should be the person who was in the room.

What to Do Before Next Friday

Collect twenty of your best progress reports, your scoring conventions and a one-page cheat sheet of each exam's structure for this testing year. That folder is 80% of the work. Then run the same ten student files through your general assistant and through one tuned on that folder, and have your Academic Director mark the errors in red. You will know inside an hour whether this is worth paying for.

Frequently asked questions

Do I have to hand over student data to do this?

No, and you should not start there. Strip names and use score patterns, section labels and your own written conventions. Under FERPA your district contracts almost certainly restrict what leaves your systems, and families of students under 13 bring COPPA into it. Tune on your language first; personalization on real records is a later decision with a signed agreement behind it.

Is this the same thing as writing a better set of instructions for ChatGPT?

Partly, and for a small center a good instruction document plus your cheat sheet may get you most of the way. The difference shows up at volume and at the edges — the abbreviations your tutors invented, the scoring quirks of your local Catholic school entrance exam, the way your own diagnostics are graded. That is where written instructions run out and tuning on your documents starts to pay.

Which tests does a general model get most wrong?

In our experience the worst offenders are the regional admissions exams — SHSAT, TACHS, HSPT — and anything involving score conversion between the SAT and ACT. The national exams are covered well in general writing; the exam that only 30,000 eighth-graders in one city take is not.

How often does it need updating?

Once a testing year at minimum, plus any time an exam changes format. Put it on the same calendar reminder as your curriculum refresh, right after the June test date when your floor is quiet.

One place this shows up outside the written reports is the phone. Parents call and ask questions in exactly this vocabulary — "she got a 7 stanine, is that good," "does he need the science section," "when is the next Bluebook date" — and whoever answers has about four seconds to sound like they know. CallSphere builds AI voice and chat agents for the center line and website chat that answer from your own approved material, book the diagnostic, and pass anything they should not answer to a human. The vocabulary problem is the same problem; it just arrives out loud.