How this was built

Sources

All 130 source documents were downloaded from the VCAA Mathematical Methods examinations page: every examination paper and external assessment report published for 2006–2025, in both the November and NHT series, plus the formula sheets, examination specifications and assessment guides. Exam 1 is technology-free short answer; Exam 2 is 20 multiple-choice questions then extended response.

Reading the percentages

VCAA reports state per-question performance in two formats, and both were parsed programmatically rather than by hand or by eye:

• Written questions (Exam 1). Each question carries a table of the form Marks 0 1 2 Average / % 39 25 36 1.0. The figure under the maximum mark is the percentage of the state that earned full marks. In the PDF reports these tables are laid out with no cell structure, so the parser clusters words into visual rows by baseline and aligns the value row to the header row by x-position. In the newer DOCX reports the table markup is read directly.
• Multiple choice (Exam 2). The report prints the percentage choosing each option, and marks the correct one only by shading the cell. In the PDFs that shading is a filled rectangle with no text, so the parser reads the page's vector fills, keeps the light-grey ones, and matches each to the option column it covers. In the DOCX reports the shading appears either as a cell fill or as a run-level highlight; both are handled, and later reports also carry an explicit “Correct answer” column, which is used to cross-check.

Cutting the question images

Each exam paper was indexed for the position of every “Question N” heading, every lettered part, and every shared stimulus block. A separator's image is the whole parent question — stimulus, diagram, every part — so the extract makes sense on its own, not a fragment of a part. 1,504 images were cut this way.

Two quirks of VCAA's own PDFs had to be undone first. Several papers store glyphs with a constant +29 codepoint offset, so 4XHVWLRQ is really Question; that is detected per word and reversed. The 2023 and later papers encode a full stop as K, so part labels read aK rather than a.

The 2024 November papers and the 2025 NHT papers were published with their type converted to vector outlines — they contain no text at all. Those four papers were indexed by reading rendered page images visually instead, and the resulting crops were checked against the reports.

Known gaps

• NHT examinations have no published statistics. Their reports give answers and sample responses only. No NHT question appears in the separator table; they are treated separately on the NHT page, where difficulty is inferred and labelled as such.
• Grouped parts. Older reports often print one mark distribution for several parts at once — “Questions 1a. and 1b.” over a 0/1/2 table. Those appear here as a single row labelled Q1a+1b, and the percentage is the share who got every part in the group right. They are not split, because VCAA never published the split.
• The area of study is inferred, not printed. Mathematical Methods papers do not label a question with its area of study, so each question was classified from its own words in the paper and from the report's published answer and commentary. The evidence behind every automatic assignment is kept in topic-evidence.json; the questions the vocabulary could not place were read individually and recorded in topic-manual.json.
• Answers VCAA published as a picture. For a drawing question the report prints the answer as a figure, not as words. Those figures are lifted straight out of the report: embedded images are extracted from the DOCX (vector EMF/WMF metafiles rasterised through GDI+), and for the older PDF reports the strip between the mark table and the next question heading is cropped.
• 0 written questions carry no answer text. VCAA published their answer as a diagram, a graph or an equation image rather than as words. Those rows say so when opened, and the question image is still there.
• Percentages come from VCAA and are rounded by VCAA, so a distribution occasionally sums to slightly more or less than 100.
• 2020 was the COVID-adjusted year. Its percentages are real but its structure is not a guide.

What a separator is, precisely

A question where the published percentage earning full marks is 50.0 or lower. For multiple choice, the percentage choosing the correct option. Partial credit does not count: a 3-mark question where 80% of students scored 2 but only 30% scored 3 is a separator, because the third mark is where the cohort divides. That is the whole point of the list.

Rebuilding it

The pipeline is in tools/: build_dataset.py reads the reports, index_papers.py indexes the papers, crop_questions.py cuts the images, classify_topics.py assigns each question its area of study, and build_site.py writes these pages. Every step takes the subject as its argument. Current run: 56 reports parsed, 58 papers indexed, 1,504 graded questions, 779 separators.