How the tutor works.
PreMedulla runs the diagnostic and planning work in the background while you practice. Here's what that actually looks like.
You commit to an answer before you see the choices.
You read the passage. You read the question stem. Then — before the answer choices appear — you type the answer you'd give in your own words, plus the reasoning that got you there.
This is predict-mode practice. It's the surface the tutor needs to grade how you think, not just which letter you bubble. You can't get diagnostic feedback on a reasoning move you never made visible, so the page makes you make it visible first.
Once you commit, the choices unlock. You pick one. The tutor takes both — your prediction and your final pick — into the grading layer.
The tutor grades the reasoning, not just the bubble.
PreMedulla compares your reasoning to the passage's actual argument structure. If you fell into a specific trap — over-inference, scope creep, swapping a premise for a conclusion — it names the trap against a real error taxonomy, not against a generic rubric.
Then it tells you why the right answer is right in terms of your reasoning move. Not "C is correct." Something closer to: you over-applied the second paragraph's claim to the question's scope, and choice B keeps the scope where the passage left it.
This is the part that earns the "personal tutor" framing. The feedback is about how you think, not about the letter you picked.
Every miss feeds your error-pattern profile.
Every graded attempt becomes a logged event. Across passages and sections, the recurring patterns become visible — the same over-inference move costing you the same points across humanities passages, the same trap-type catching you on social-science passages.
You don't have to assemble this yourself. The tutor does it in the background. That's the medulla part: the work running underneath, so the work running on top is the work that actually moves your score.
The profile gets sharper every passage. Over a few weeks of practice, the picture isn't "you missed 14 questions"; it's "this one reasoning move costs you the most points, and here's where it shows up."
The plan adapts. Debriefs tell you what to do next.
The planning layer reads your profile and tells you which weak spot to fix first this week. Review this trap type. Redo these two passages. Watch this Sketchy / Khan / AnKing / AK Lectures resource if the gap is a content hole.
After a full-length section, the debrief tool surfaces the patterns from that session and what they mean for the next week of prep. You stop deciding what to study and start doing it.
The plan adapts to your profile, not to a generic calendar.
What the tutor covers today.
Honest about scope. The tutor is complete for the sections it covers — not half-built.
Available now
CARS · Psych/Soc
Full predict-mode practice, reasoning grading, error-pattern tracking, and adaptive review for both sections.
Coming next
Bio/Biochem · Chem/Phys
Same engine, different content corpus. They ship when paying users say they're the priority — not on a marketing calendar.
What the tutor doesn't do.
PreMedulla doesn't reteach biology, organic chemistry, or the rest. When your profile shows a real content hole, it points you to the resources that already do it well. The reasoning and pattern layer is the part we own; the content layer already exists and we point at it.
- Sketchy
- Khan Academy
- AnKing
- AK Lectures
Try it: the free CARS Diagnostic.
The smallest version of the full experience. You'll feel the predict-mode loop and get a one-page error-pattern report in a single sitting. No card, no account.
First cohort, early access. Enter your email and you'll get your invite when the Diagnostic opens.
Get early access.
The first cohort opens soon. Leave your email and you'll be first in when PreMedulla goes live.
No spam. No card. One email when it's ready, and not before.
Questions about how it works.
What if I'm bad at explaining my reasoning in writing?
You don't need polished writing — you need an honest sketch of the move you made. A few sentences is enough. The grader is reading for the reasoning step you took, not for prose quality. If anything, students who over-format their predictions slow themselves down; messier and faster is fine.
Does it work without the answer choices showing? What about STEM where the choices are the question?
For CARS and Psych/Soc, predict-mode is the right default — you commit before you see the choices. For Bio/Biochem and Chem/Phys (when they ship), the predict-mode loop adapts: you commit to the reasoning step or the calculation setup before checking the choices, so the grader still has something real to read.
What happens to my data?
Your practice attempts and error patterns are yours. They live in PreMedulla and don't get sold, shared, or used to train anyone else's model. If you cancel, you can export or delete the profile.
Is the AI making things up?
CARS grading is reasoning-against-passage — the grader is checking your move against the passage that's already on the screen, so there's no fact-retrieval to hallucinate. STEM grading is grounded in our reference corpus.
Can I see the trap taxonomy?
Yes. It lives in the post-attempt debrief — we don't gate it. The taxonomy is the shared language the tutor and you use to talk about a miss, so it'd be silly to hide it.
How is this different from just using ChatGPT?
ChatGPT will answer any question you ask it. It won't grade your reasoning against a real error taxonomy, it won't track your patterns across sections, and it forgets you between sessions. PreMedulla is built around the one thing a general chatbot can't do — a persistent profile of your misses that gets sharper every passage.