Grading accuracy
Every number checked, and tested like it matters
- Graded numbers tested
- 213,624
- Right verdict on first pass
- 99.41%
- Sent to a human instead
- 1.75%
- Wrong and silent
- 0
Last full run October 17, 2026. The suites are rerun whenever the grader changes, and a single silent error fails them.
Every number, not a few checkpoints
When you build the model yourself, every line and every year is graded against a reference solution, not a handful of highlighted cells.
Built the way people really build
Each correct model is rebuilt dozens of ways: shifted rows, your own labels, sections in a different order, costs entered as negatives, figures in thousands. Then mistakes and cheats are planted in it.
Unsure means a human looks
When the grader isn't sure, it says so and the number goes to review instead of being marked wrong. “Wrong and silent” is the failure nobody would catch, and the target is zero.
Correct models rebuilt, then broken on purpose
Every Speed Session model at six seeds and every Deal Flow deal on a blank canvas, each rebuilt the ways below. The first group are correct models; the second contain one planted mistake or cheat the grader has to catch.
| How it was built | Graded numbers | Right verdict | Sent to review | Wrong and silent |
|---|---|---|---|---|
| Correct build in the guided layoutCorrect model | 6,651 | 100.00% | 0.00% | 0 |
| Correct build started 3 rows lower and one column rightCorrect model | 6,651 | 100.00% | 0.00% | 0 |
| Random blank rows and side notesCorrect model | 6,651 | 100.00% | 0.00% | 0 |
| Each graded line labelled with one of its accepted aliasesCorrect model | 6,651 | 99.29% | 0.71% | 0 |
| Candidate's own wording: case, units, abbreviations, colonsCorrect model | 6,651 | 99.67% | 0.33% | 0 |
| Sections built in a different orderCorrect model | 6,651 | 99.50% | 0.50% | 0 |
| Look-alike rows (Unlevered vs Levered, GP vs LP…) next to the real onesCorrect model | 6,651 | 99.46% | 1.52% | 0 |
| Year headers written as FY2027… instead of Yr 1…Correct model | 6,651 | 96.72% | 3.28% | 0 |
| Offset + spacing + own wording + decoys combinedCorrect model | 6,651 | 99.38% | 1.43% | 0 |
| Candidate's own link lines (=B7) between the assumptions and the formulas that use themCorrect model | 6,651 | 100.00% | 0.00% | 0 |
| One given input typed into a formula as a number (=B7*23) instead of referencedContains a mistake or cheat | 6,651 | 97.43% | 3.62% | 0 |
| One graded line built with a wrong formula (×1.05)Contains a mistake or cheat | 6,651 | 100.00% | 0.50% | 0 |
| One graded line typed in as a numberContains a mistake or cheat | 6,651 | 99.55% | 0.47% | 0 |
| Cheat: =<one assumption>*k tuned to land on the right answerContains a mistake or cheat | 6,651 | 99.38% | 2.21% | 0 |
| One graded line never built (row deleted)Contains a mistake or cheat | 6,651 | 99.94% | 0.62% | 0 |
| Two graded lines' labels swapped (labels lie about the numbers)Contains a mistake or cheat | 6,651 | 99.62% | 3.41% | 0 |
| Real line relabelled as a look-alike (e.g. 'Unlevered IRR' holding the levered number)Contains a mistake or cheat | 6,651 | 93.26% | 10.75% | 0 |
| Every formula pasted as valuesContains a mistake or cheat | 6,651 | 100.00% | 0.00% | 0 |
| Nothing built below the assumptionsContains a mistake or cheat | 6,651 | 100.00% | 0.00% | 0 |
Student-built models
The same models built the way students actually build them: typos, their own formatting, costs entered as negatives, figures in thousands, missing sections, half-finished sheets and scratch math in the margins.
| How it was built | Graded numbers | Right verdict | Sent to review | Wrong and silent |
|---|---|---|---|---|
| Misspelled labels (dropped, doubled, swapped letters)Correct model | 5,817 | 99.48% | 0.86% | 0 |
| Bullets, numbering, dot leaders, stray punctuation, random caseCorrect model | 5,817 | 99.83% | 0.17% | 0 |
| Step numbers (1, 2, 3…) typed in a column left of the labelsCorrect model | 5,817 | 100.00% | 0.03% | 0 |
| Costs entered as negative numbers, used consistentlyCorrect model | 5,817 | 99.98% | 0.19% | 0 |
| Dollar lines built in $000s, used consistentlyCorrect model | 5,817 | 99.79% | 0.21% | 0 |
| Rate lines as 5.5 instead of 0.055, used consistentlyCorrect model | 5,817 | 99.95% | 0.05% | 0 |
| Offset + gaps + typos + sloppy labels + negative costs, all correctCorrect model | 5,817 | 99.67% | 0.96% | 0 |
| A whole section never built (downstream lines break)Contains a mistake or cheat | 5,817 | 99.40% | 2.66% | 0 |
| Only Year 1 built on a few schedule rowsContains a mistake or cheat | 5,817 | 99.72% | 2.48% | 0 |
| Year 1 formula pasted across with anchored referencesContains a mistake or cheat | 5,817 | 99.98% | 2.11% | 0 |
| Later years read the prior year's driverContains a mistake or cheat | 5,817 | 99.98% | 7.05% | 0 |
| #REF! and #DIV/0! in a few formulasContains a mistake or cheat | 5,817 | 100.00% | 0.46% | 0 |
| A rate typed as 5.5 instead of 5.5% without fixing the formulasContains a mistake or cheat | 5,817 | 99.97% | 1.31% | 0 |
| Wrong lines with the right number typed in a scratch area beside themContains a mistake or cheat | 5,817 | 99.98% | 5.43% | 0 |
| Typos + sloppy labels + half-built rows + a missing section + scratch mathContains a mistake or cheat | 5,817 | 99.62% | 6.96% | 0 |
What it still gets wrong
Fewer than one verdict in a hundred is wrong before review. Nearly all of those are correct numbers the grader couldn't place with confidence: a line labelled very differently from the case, or a look-alike row (levered next to unlevered) holding the number. Those go to review rather than being failed. The hardest cases are year headers written as fiscal years and inputs typed straight into formulas, which is where the review rate is highest.
Think a grade is wrong? Every graded cell shows what it expected and why, and you can tell us. Professors and recruiters see the same reasoning on every submission.