Benchmark · AI IELTS grader

Target7

An AI IELTS-writing grader, benchmarked essay-by-essay against a teacher's hand-marked bands — measured honestly, and held conservative where the data said to.

Ground truth: teacher Bình Lưu's hand-marked essays — 9 essays, 200+ comments, official per-criterion bands. One teacher, small sample — stated up front, because the point of a benchmark is that you can trust it.

Live · target7.vn ↗

95%
Error recall (Task-2) — catches 95 of every 100 errors the teacher marks
0.31
Band-score accuracy (MAE) — within a third of a band of the teacher
0
Over-grades — bias −0.31, safely conservative
100%
Regrade consistency — same band ±0.5 across 3 runs (temp 0)
855/855
Automated tests passing
49.2s
p95 latency — under the 58s abort wall; detectors add 0 ms

Error detection vs the teacher

The job that matters: catching what a real examiner catches. Recall is measured by automated span-overlap against every error the teacher marked.

EssayTaskTeacher errorsCaughtRecall
Xuân MaiT2322784%
Trần Hải LongT2232296%
Giang TrịnhT2454396%
HGT23535100%
Hương TràT21919100%
Task-2 margin total15414695%

Two tracked-changes essays (50% / 57%) are excluded from the margin: ~85% of the teacher's marks there are the student deleting words for style, not grammar errors, so span-overlap under-counts them. Task-1 misses are chart-data errors a text grade can't see — now covered by live Task-1 data checks.

Band-score accuracy

Predicted overall band vs. the teacher's official band, across 8 scored essays.

EssayOur bandTeacherΔ
Giang Trịnh5.55.50
HG5.05.00
Thanh HW26.56.50
Thanh HW37.07.00
Xuân Mai5.56.0−0.5
Trần Hải Long5.56.0−0.5
Hương Trà6.06.5−0.5
Hương Trà T15.56.5−1.0
0.31
Mean absolute error — below the ~0.45 human examiner-to-examiner floor
7 / 8
Within half a band · exact match on 4/8 · bias −0.31 (never over-grades)

Where the gap is

Every band miss traces to one criterion — Lexical Resource — and nothing else.

CriterionBias vs teacherVerdict
Task Response~0On the money
Coherence & Cohesion~0On the money
Lexical Resource−0.44The one systematic weakness — stingier on vocabulary
Grammatical Range~0On the money

The honest call — a data-backed "don't ship"

Three prompt formulations were A/B-tested against the teacher's real Lexical-Resource bands.

VariantBand error (MAE)BiasResult
Current (shipped)0.438−0.44Conservative — held
Descriptor anchor0.375+0.13Inflates weak essays (+1)
Hard floor0.625−0.63Crushes strong essays
Finding

The band-5/6 line is bimodal — it can't be held by prompt wording

The lowest-error variant only wins by inflating weak essays a full band — the opposite of what a grader students trust should do. So Lexical Resource was held conservative rather than degraded: a measured decision not to ship a change, backed by the A/B data. Shipping the honest non-decision is the point.

Deterministic layer & reliability

0 FP
In-browser detectors (parallelism, wordiness, comma-splice, structure) — zero false positives on all 9 essays, 0 ms / 0 API calls
~89% / ~90%
Fix quality / precision (skeptical LLM-judge) — corrections are genuine, few false alarms
On the job that matters — catching what a real examiner catches — it runs at ~95%, with band scores inside the human-examiner margin and a bias that never over-grades. The one weakness, I proved with data can't be safely fixed — so I held it.

Honest caveats

So the numbers stay trustworthy: ground truth is one teacher, 9 essays — strong signal, small sample. Recall matching is automated span-overlap (±a few per count). The eval reproduces the v100 grader logic via DeepSeek; the live edge function adds caching + chart vision + band finalization on top.