AI Case Feedback 2026: Instant Scores Without a Model

No model runs behind this case interview feedback

The file that grades a drill answer is src/lib/grading/rubrics.ts, and it is worth quoting the header of its companion file, types.ts, because it states the design choice in the first line: assessment is deterministic, an identical answer always returns an identical score and identical wording. Checked 29 September 2026, there is no model call anywhere under src/lib/grading.

What runs instead is plain text analysis. Numbers are pulled out with one regular expression. List structure is detected by counting bullets, semicolons and line breaks. Coverage of an expected branch is a synonym lookup: the structure prompt about a grocery chain carries three expected themes, and touching any word from a theme's list scores that theme.

Interpretation is checked by looking for connective words such as so, therefore or which means. Risk is checked by looking for risk, however, caveat, assumes or downside. Sentence count decides whether a structure is short enough to say out loud.

This is much less clever than a language model, and it buys three things a model does not give. The score is reproducible, so two attempts are comparable. It is auditable, so the sentence above about connective words is a fact you could verify in the file. And it is free to run, which is why every drill opens without an account and without a paywall on the feedback.

One correction to make in the open. Our plan cards list an AI debrief among the features, in src/lib/plans.ts. This repository holds no model call for grading, so that line describes something we cannot demonstrate from the code, and the earlier version of this page called the same thing an AI grader. What is in front of you is a rule-based scorer, and calling it that is the only description the code supports.

The case interview feedback scorecard, field by field

Every submitted answer returns the same object, and the debrief renders it in this order.

Two of those rows are thresholds rather than opinions, which is the part worth knowing before you read your first debrief. A dimension at 60 percent or above is not commented at all, so a silent dimension means adequate and not good. And a dimension shown under "Kept" earned every available point, which on a 45-point coverage check is easier than it sounds.

The guard at the top of every rubric is worth knowing too. An answer with no number, fewer than two list marks, fewer than two sentences and under eight words is refused rather than graded. Submit the word dunno and the response is a single dimension called Attempt, scoring 0, with the note that there is nothing to assess yet. That path exists because a genuine three-word answer, such as a number on its own, should be scored by the rubric rather than rejected for length.

Four scores that show where case interview feedback goes blind

Because the scoring is deterministic, its weaknesses can be measured rather than argued about. We ran these on 29 September 2026, and anyone with the repository gets the same four numbers.

1. Wrong number, 72 out of 100. A math answer that derives its way to $4.32M when the answer is $1.92M loses 28 points on the Answer dimension and keeps everything else, because Working, Units and Interpretation score the shape of the reply.
2. Three keywords, 85 out of 100. On the structure prompt, the string revenue; cost; competitor clears both coverage and separation and only drops on depth.
3. An unchecked estimate, 80 out of 100. A market sizing answer that frames what it counts, chains three assumptions and writes the multiplication out, then stops without a plausibility check, scores 80. Sanity check is the one dimension in that rubric which pays zero when it is absent, so the closing sentence of the answer is worth 20 points on its own.
4. An exhibit read with no conclusion, 30 out of 100. Quoting both figures from the chart pays 25 out of 25 on observation. Stopping there gives up the 30 points for the so-what and 20 of the 25 for the implication, plus the 20 for naming a next test.

Rows two and four are the interesting pair. The scorer is generous where language is cheap and strict where a candidate has to commit to something. A model would judge rows one and two better than we do. It would also return a different score on Tuesday, and neither we nor you could say why.

What the firms publish about AI in your preparation

This is the part the search results skip, and it is the reason to read a firm's own pages before choosing a practice tool. McKinsey publishes a table of AI uses on its interviewing page, and the split is clean.

The firm's own words on the first group: "When used responsibly, we see AI as a valuable tool that supports preparation." Its assessment integrity expectations page carries the other side, and it is specific about the sitting itself. Candidates are asked to refrain from "using any applications or websites, generative AI, a calculator, or prepared notes", and to "use only pen and paper for notetaking and calculations". The same page lists what breaking that can cost: "disqualification from our process, rescinding of an offer, restrictions on future applications, or legal action."

So drilling with a scored tool sits squarely in the encouraged column, and carrying one into the assessment sits in the other. Test publishers describe their side in the same register. HireVue tells candidates on its candidate FAQ that its interviews and assessments "are intended to be first steps in the overall interviewing process", with "person-to-person interviews before hiring decisions are made by people." That page does not say whether a candidate is shown a score, and we are not going to guess.

The part of case interview feedback a person still does

A rubric reads what you wrote. It does not hear that you were close, notice that your voice sped up on the number, or decide that today the useful correction is confidence rather than coverage. It cannot ask a follow-up question designed to see whether you will defend a weak branch or abandon it.

The deeper limit is stated by the firms themselves. Bain writes on its case interview page that it is "not looking for a single right answer; we're looking for how you think, structure problems, and build on ideas as the conversation evolves." An answer builds on ideas across a conversation, and a drill has no conversation in it. A structured, confident, well-quantified answer resting on a false premise gets a high score from our code and fails in front of an interviewer, and that gap is not something a threshold can close.

What a rule-based scorer wins on is the axis where a human cannot compete: availability and cost. Corrections arrive between two reps rather than between two sessions. A coach who has watched fifty candidates on the same prompt will tell you more in ten minutes than this page can, and will charge for the hour. Both are worth having, in that order.

Where to see it run, and what it costs

The weights and the reasoning behind every dimension are set out on how an answer is assessed. That page executes the six rubrics on a sample answer and prints what they return, so it reports the live code rather than a copy of it. Then open any drill from the drills index or the six linked from our home page and submit something deliberately incomplete, because a debrief on a weak answer tells you more about the scorer than a debrief on a strong one.

Cost, plainly. Every drill and its full scorecard is open with no account and no email. A subscription is $49 a month, $120 once for a 90-day window, or $249 a year, and none of it improves the grader, since the paid product runs the same rubrics. What the money buys is volume and guided case runs.

And the honest gap to name before you leave: the six assessment simulators are not built. Our assessments page counts them at "0 of 6 simulators are live", so if you want a scored rehearsal of a specific test, this is not yet the place, and the drills train the underlying skills instead.