Building the future of marking

Fairer Graph and Calculation Marks

Scheme reconcile no longer overrides incomplete visual marks, and correct numerical finals plus clear conceptual equivalents are scored more like a human marker.

• Visual-gated reconcile: If vision already found missing graph or diagram criteria, all-met checklist reconcile will not force full marks.
• Correct finals count: On multi-mark numerical questions, a correct final answer earns full marks unless the scheme explicitly splits working and answer credit.
• Conceptual equivalents: Standard alternative wording (for example permeability vs free space, or “diode” for flyback suppression) is accepted when the concept is clear.

Faster, Fairer Batch Marking

Large handwritten papers no longer spend most of their time on wasted image compares. Vision and marking calls now run concurrently, equation crops stay display-only, and over-awards can be corrected after adjudication — the same way under-awards already were.

• Skip wasted image compares: Mark-scheme equation crops are kept for display but no longer trigger a full image-to-image compare on every calc question.
• Concurrent vision & marking: Page scans, reference compares, ink reads, marking chunks, and students now run in parallel within safe concurrency caps.
• Symmetric scheme reconcile: Over-awards vs the checklist are flagged for a second look; if adjudication lowers the mark, that lower mark is stored with an audit note.
• Single-mark substantive credit: On 1-mark questions with a compound scheme point, meeting the main requirement earns the mark; missing elaborations are noted in feedback instead of scoring zero.

Fairer Automarking Scores

Two accuracy fixes for question-template marking: the marker can no longer tick every scheme point and then award zero, and near-miss handwritten maths (e.g. reading “sin²t” instead of “sin 3t”) gets a higher-zoom confirmation pass before the mark is locked in.

• Checklist/score reconcile: When the AI marks every scheme point as met but under-awards, the question is re-adjudicated and — if still contradictory — awarded the full mark with an audit note.
• Ink near-miss confirm: Handwritten answers that sit just off the expected scheme answer are re-read at higher zoom (capped per script) so a single digit/superscript misread does not cost the mark.
• Method-credit prompt: When correct working ends in a wrong final answer, results show a “Method credit?” badge so you can add the mark — the automatic score still follows the printed scheme only.

Handwritten & Digital-Ink Answers on PDFs

Automarking now detects freehand / digital-ink annotations on learner PDFs (the “draw” tool in Adobe, Preview, browsers, etc.) and vision-reads only those pages — so maths and other written-on-paper scripts are no longer marked as blank when the text layer only has the printed questions.

• Evidence-triggered vision: Pages with Ink, Stamp and similar annotations (or large pasted images) are rendered with annotations included and read by vision. Fully typed submissions cost nothing extra.
• Additive merge: Handwritten transcriptions are appended to the submission text rather than replacing it, so typed answers and printed structure are kept.
• Never silently unattempted: If ink is present but unreadable, the question is flagged for teacher review with a crop of the learner’s writing — not reported as blank.

Question Templates Read Visual Content

Building a question template from an uploaded paper now vision-reads every page — so equations, diagrams, circuits, figures, tables and other non-text content are captured instead of being dropped by plain PDF text extraction.

• Always-on vision ingest (flagged): Template Builder merges the PDF text layer with a high-detail vision read of each page. Works across subjects — maths notation, chemical structures, circuits, labelled figures, graphs, maps, music, code and tables.
• Mark schemes from the page: When AI generates a mark scheme, it can see the source paper images, so it no longer invents “not present in the supplied OCR” excuses for missing expressions or diagrams.
• Review warnings & regenerate: Hollow or low-confidence schemes are flagged on the review screen, with a one-click regenerate button. Alignment disagreements between text and vision raise a banner so numbering can be checked before saving.
• Safe rollout: Vision ingest ships behind TEMPLATE_INGEST_VISION_ENABLED (off by default) with a shadow-compare script so existing templates can be checked before turning it on for all users.

Re-upload a paper through Question Templates to rebuild any template that previously lost visual content.

Diagram & Chart Marking Accuracy

Visual questions (graphs, sketches, diagrams) now score by official mark points, show clearer evidence in results, and support per-question re-marking.

• Mark-point scoring: Criteria map to the scheme’s actual mark points so a wrong axis label no longer rounds away a deduction (e.g. 1.5 silently becoming 2).
• Clearer evidence: Higher-resolution crops, mark scheme text on results, mark-point checklist with criterion notes, and a cross-student visual moderation grid.
• Accurate diagram locate: Page hints from the evidence scan, crop confirmation, and double-marking against the mark scheme so the wrong page is far less likely to be scored.
• Re-run the batch: To update diagram marks after editing criteria, re-run the whole automarking job. Editable mark points on templates let you set weights before a job runs.

Open an automarking results job and use Needs visual review or Visual moderation to check diagram answers.

Reporting Hub Modernization

Classes → Reporting hub is now a shared teacher/manager workspace with dual risk engines, analytics panels, and clearer intervention actions.

• Intervention risk & Mark risk: Switch engines without mixing signals. Teachers default to a High+Medium focus queue; managers open on full cohort oversight.
• Analytics panels: Grade distribution, weekly outcome trend, question/criterion hotspots, and ALS/RPL vs peers — reused from existing class and dashboard compute.
• Attention list open by default: Paginated (25 per page), richer weekly digest names, integrity queue strip, and drawer deep-links to progress reports, notes, meetings, and marking sparklines.
• Export & docs: CSV includes confidence, similarity, outcome notes, class lifecycle, and ULN when enabled. Help and knowledge base now say Classes → Reporting hub.

Open Classes → Reporting hub to try the new layout. Preferences for risk engine and filters are remembered on your account.

Attempt Limits for Rubric & Automarking

Set how many times a student can be marked against the same rubric or question template. Work beyond the cap is still marked in full, but held off the profile until you choose to record it.

• Optional max attempts: On each rubric and question template, turn on a limit (1–20). Attempts count per student against that rubric or template across every class and assignment. Only completed marks count; failed or skipped runs do not.
• Over-limit stays readable: Scripts beyond the cap are still AI-marked with full feedback. They show as Too many attempts on the results page and are not written to the student profile until a teacher overrides.
• Override when it should count: Record one mark, or every over-limit mark in the job, from the batch or automarking results page. Final attempts are flagged in red so you can see when a learner is on their last allowed go.
• See it before you spend credits: ZIP upload previews show the next attempt number when a cap is set. Student profiles show Attempt X of Y on capped history rows, and blocked marks stay off the profile until recorded.

Set a limit under Rubrics or Automarking → My Question Templates (Edit). Ad-hoc automark jobs with no saved template stay unlimited.

Class Overviews, Image Recognition & Multi-File Marking

Results pages now give you a saved cohort summary after a job finishes, and automarking reads more of the work students actually submit — including handwritten Word ink, page images, and folders with more than one file.

• Overview of marked results: On both batch (rubric) marking and automarking results pages, a class overview appears under the student table — average score, highest and lowest, common strengths and gaps, plus a short class summary. It is generated once when you open a completed job and saved on that job for next time.
• Image & handwriting recognition: Automarking now Vision-OCR’s image-heavy and ink Word files (handwriting drawn in Word, embedded page scans) when there is little or no typed text. Those submissions are no longer skipped as “no readable text” when the pictures still contain the answers.
• Multiple files per student: When a learner folder has more than one upload (for example a PDF plus a Word file), unread Office/PDF/image files are OCR’d and merged with any text that did extract — so one readable sibling no longer leaves the rest of the work ignored. You can also download a student’s source files from the results view alongside the marked PDF.
• Clearer password messaging: If a file opens successfully with your organisation’s default password, results show a quiet info note — not a yellow “with issue” warning. Warnings stay reserved for files that could not be read.

Open any completed automarking or batch results job to see the new overview. Re-run a job if you need the summary rebuilt after fresh marks.