Highest grading accuracy at the lowest cost per activity
Three numbers, all from production: accuracy as the share of teacher-reviewed AI grading that was not corrected, LLM cost per graded activity, and the share of papers the teacher had to fix. Measured from Oct 1.
Targets locked Nov 1
Grading accuracy
Target not set90.66%
all configs excl. beta · 104,516 teacher-reviewed rubric evals
LLM cost per activity
Target not set$0.81
October 2026 · $3518 ÷ 4,339 activities
Papers needing manual review
Target not set24.0%
7,580 of 31,556 reviewed papers corrected
Locked on 2026-11-01. Status on the overview is judged against these.
Q4 to date vs the October baseline. Only teacher-reviewed work counts (result viewed or score corrected), from teachers with at least 3 activities of 10+ papers; rubric types include every strictness config, custom and system. Cells under 30 evaluations are hidden as noise.
| Type | Q4 | October | Volume |
|---|---|---|---|
Problem Solving All configs · rubric level | 92.48% | 92.43% | 75,439 rubric evals |
Essay All configs · rubric level | 85.93% | 85.84% | 29,077 rubric evals |
Visualsbeta All configs · rubric level | 94.16% | 94.66% | 3,064 rubric evals |
Code on Paperbeta All configs · rubric level | 89.35% | 89.35% | 4,603 rubric evals |
Multiple Choice Objective · answer level | 99.62% | 99.60% | 810,953 answers |
Identification Objective · answer level | 94.70% | 94.59% | 165,493 answers |
Enumeration Objective · answer level | 77.66% | 77.14% | 5,811 answers |
Multiple Answer Objective · answer level | 90.10% | 91.41% | 2,414 answers |
Same pipeline as the Q3 accuracy goal — full daily history. Visuals & Code on Paper are beta: shown, never judged.
Langfuse production traced cost (USD, Langfuse price table) ÷ activities with at least one graded paper in the month. Ratio of monthly totals; the current month is month-to-date and never judged.
| Month | LLM cost | Activities | Per activity | Per paper |
|---|---|---|---|---|
| November 2026MTD | $149 | 664 | $0.22 | $0.006 |
| October 2026baseline | $3518 | 4,339 | $0.81 | $0.013 |
A paper needs manual review when the teacher changed at least one score on it. Share of reviewed papers (viewed or corrected) per grading day; days under 30 reviewed papers are hidden. Corrections lag grading, so recent days still move.
Features listed under this goal in the Q4 direction deck. Ship dates come from the release sync when a release name matches.
New Rubrics Generation
PlannedNo matching release yet
Free-form Rubric
PlannedNo matching release yet
Adaptive Personality Generator
PlannedNo matching release yet