Three models. Three marking phases.
One inventory valuation script.
All totals are out of 21. These are the awards reported in the AI outputs.
| Model | 1 · Initial marking | 2 · HKEAA guidance | 3 · Official quiz scheme |
|---|---|---|---|
| Hark | 21Provisional; stated range 20.5–21 | 21Earlier range withdrawn | 21Teacher discretion: 20–21 |
| Muse Spark | 21Provisional | 21Quiz weights still inferred | 21Scheme-based claim |
| ChatGPT | 21Provisional | 20–21Final-profit point conditional | 21Recommended award |
Each final breakdown is Q1: 8/8, Q2(a): 4/4 and Q2(b): 9/9. These are AI assessments, rather than a verified teacher award or HKEAA endorsement.
The same inventory valuation answer script was marked by Hark, Muse Spark and ChatGPT in three successive phases. Each phase produced a directly annotated PDF, labelled Mode A in the supplied files.
Marking skill, question paper and student answer sheet. The quiz-specific marking scheme was not supplied.
Re-marking with the 2025 and 2026 BAFS Paper 2A briefing documents. These concern other examination questions and do not establish this quiz’s precise weights.
Re-marking with the school quiz’s supplied solution and marking scheme, alongside the earlier materials.
Based on the provided skill, question paper and answer sheet, please mark the paper using mode (a) - direct annotation, and output the pdf to me. name the pdf such that your AI model, given info (i.e. skill, QP and AS) is clearly shown in a standardised format.
Now, please refer to this official document by the HKEAA in addition to the previous materials: https://www.hkeaa.edu.hk/DocLibrary/HKDSE/Subject_Information/bafs/PowerPoint-BAFS-2026-03-rev.pdf and https://www.hkeaa.edu.hk/DocLibrary/HKDSE/Subject_Information/bafs/PowerPoint-BAFS-2025-03-rev.pdf, and remark the paper accordinly.
Now, you are given the official marking scheme, please remark.
Reported request-to-PDF times, supplied by the experiment organiser. Approximate values remain approximate.
| Model | 1 · Initial | 2 · HKEAA | 3 · Quiz scheme | Sum of phases |
|---|---|---|---|---|
| Hark | ≈ 3 min 15 sec | ≈ 2–3 min | 34 sec | ≈ 5 min 49 sec– 6 min 49 secCalculated from the reported phases |
| Muse Spark | ≈ 5 min | ≈ 1 min | ≈ 1 min | ≈ 7 min |
| ChatGPT | 14 min 59 sec | 6 min 18 sec | 5 min | 26 min 17 sec |
Hark: Phase 1, 15:44:17–15:47:32. Phase 2 took roughly 2–3 minutes and finished at 15:51:45; its start was not supplied. Phase 3, 15:52:31–15:53:05.
Muse Spark: About 5 minutes, then 1 minute and 1 minute. Muse attributed much of phase 1 to reading handwriting at high zoom and building the annotation layout. No absolute timestamps were supplied.
ChatGPT: Phase 1, 15:50:03–16:05:02. Phase 2, 16:05:46–16:12:04. Phase 3, 16:12:53–16:17:53. The reported durations include file processing, PDF checks and connection delays.
All absolute times shown use Hong Kong time. The timing record was supplied by the organiser, rather than independently measured from system logs.
Hark and ChatGPT flag the correct but unlabelled HKD 97,350 result. ChatGPT makes its final-profit point conditional in phase 2, then recommends awarding it in phase 3. Hark retains full marks but acknowledges teacher discretion in its final PDF. Muse’s final feedback does not flag this omission.
Hark’s phase 2 introduces inferred half-mark weights and treats certain totals as not separately scored. Its phase 3 replaces these with quiz-specific weights. ChatGPT also rebuilds its allocation after receiving the scheme. Muse’s final description assigns a mark to total expenses; Hark and ChatGPT instead describe six trading points, a shared salaries-and-rent point, an actual-loss point and a final-profit point.
That difference merits checking against the original scheme. The scheme itself was not supplied for this publication, so this comparison cannot independently adjudicate its allocation.
Muse flags the handwritten ‘Add’ subtotal as HKD 7,806 and proposes HKD 7,808, without deducting a mark. Hark and ChatGPT do not flag this as a student error. The arithmetic is clear: 6,032 + 1,200 + 576 = 7,808. Whether the handwriting actually reads 7,806 remains a reading judgement; Muse’s annotation is recorded here as its claim.
All three final PDFs flag a printed deduction subtotal of HKD 11,460 in the scheme. The listed deductions independently sum to HKD 9,460: 4,560 + 2,000 + 500 + 2,400. Subtracting 9,460 from 52,588 gives the agreed HKD 43,128. The arithmetic can be checked; the claim about what the original scheme prints comes from the marked PDFs.
All nine supplied marked outputs, preserved without alteration.
| Model | 1 · Initial marking | 2 · HKEAA guidance | 3 · Official scheme |
|---|---|---|---|
| Hark | |||
| Muse Spark | |||
| ChatGPT |
Nine PDFs · 23 pages · 23.70 MB in total. The ZIP includes the original filenames and a checksum manifest. Download file manifest.
This is a case study of one answer script, three labelled models and successive re-marking in the same workflow. It shows how the supplied evidence changed these particular outputs. Later phases are revisions of earlier work, rather than independent trials.
It does not establish general marking accuracy, reliability across weaker scripts, or a controlled speed ranking. There is no independently verified teacher-marked baseline, repeated trial, matched computing setup, cost record or complete chat histories in the supplied publication materials. More complete evidence and repeated blind marking would be needed for those conclusions.
The practical finding is that a shared total can conceal different feedback and mark allocations. Read the annotated reasoning alongside the headline score.
The results above were transcribed from the nine linked PDFs. Timings were supplied by the organiser. The original outputs remain available so readers can inspect the evidence and the models’ own wording.
Background documents identified in the outputs: HKEAA 2025 BAFS Paper 2A briefing and HKEAA 2026 BAFS Paper 2A briefing. They provide background context and are not this school quiz’s marking scheme.