A provisional mark should come with visible boundaries.

Remark sends the uploaded PDFs to Anthropic, and Claude reads the pages directly — the scheme to learn what each mark is for, the script to judge what you wrote. Nothing is transcribed by a separate service first, so diagrams, graphs and working reach the model as they appear on the page.

Uploading a paper queues a background job that reads the marking scheme, attempts every question it can locate, and stores a result per question and per criterion. That job does not produce an official grade. Each question is marked twice, independently, and the two readings are reconciled; where they disagree the question is flagged for you to check rather than silently averaged.

• The completed answer-script PDF.
• The official marking-scheme PDF.
• The selected syllabus, paper, session, variant and test-condition metadata.
• Instructions defining the matching, legibility and marking boundaries.

Your name, email address and payment details are never sent to the marking provider; the request carries the documents and the paper metadata only. Anthropic states that it does not use inputs or outputs from its commercial API to train its models by default. Uploaded files are also checked for malware by the configured scanning service before marking. Do not upload real student work unless you are authorised to do so and comfortable with those transfers. See the privacy notice for storage and deletion details.

1. Validate that both files are PDFs and compare their apparent paper identity with the confirmed metadata.
2. Parse the marking scheme into named criteria, and locate every question the system can find in the script.
3. Run independent marking passes that return a met/not-met judgment, visible evidence, legibility state, note and confidence information for each named criterion.
4. Validate the structured responses and calculate provisional marks awarded and available in server code.
5. Retry low-confidence full-paper questions within a bounded one-minute retry budget. A retry promotes a result only when it establishes corroborated high confidence; unresolved ambiguity remains flagged.
6. Show mismatch, unreadable-document, missing-question and low-confidence warnings rather than silently presenting every result as certain.

The exact configured model may change during development. The result records the model identifier returned by the server so a demo output can be traced to that configuration.

• Handwriting, faint scans, page rotation, crossed-out work, diagrams and answers spanning pages may be read incorrectly.
• The model can pair the wrong working with a question, misunderstand a scheme convention, omit evidence or produce a plausible but incorrect judgment.
• A confidence label is the system’s assessment of reading certainty; it is not a calibrated probability and does not prove correctness.
• Document-matching checks and instructions to ignore commands inside PDFs reduce some risks but do not guarantee protection from malicious or confusing document content.
• Model availability, rate limits, network failures and output-schema changes can prevent analysis from finishing.

Treat every live result as provisional. A student or teacher should verify the selected question number, visible evidence, scheme wording and arithmetic before relying on it. Low-confidence or illegible evidence needs particular attention.

Remark must not be used as the sole decision-maker for official grades, admissions, predicted grades, school discipline, access arrangements or another decision that could materially affect a student.

The product includes structured criterion scoring, independent-pass reconciliation, confidence routing, durable job leases and an evaluation harness.

Before launch, the operator should test against representative, teacher-marked scripts across handwriting styles, scan quality, supported papers and relevant student groups; publish meaningful error measures; define escalation thresholds; and monitor model-version changes. No claim of examiner-level accuracy, bias-free performance or validated educational impact is made today.

Marking is not the only place a model is involved. Three other things on your report are generated:

• The explanation of why marks were lost, written against the criteria the scheme actually contains.
• The worked solution shown for questions you lost marks on. The model is given the printed question and the official marking points and asked to work the question through, naming the mark each step earns. It is a model answer, not the examiner’s, and it can contain arithmetic or reasoning errors. Check it against the scheme, which is shown directly above it for that reason.
• The revision links in your improvement plan. These are found by a web search restricted to a fixed list of known revision sites, and every link is fetched and confirmed to open before it is stored. We do not control what those sites publish, and a page can change after we check it.

A worked solution is written once per question per paper and shown to every student who sat that paper. The model is instructed to work only from the printed question and to ignore the candidate’s handwriting, so one student’s answers do not travel to another. Nothing you wrote is shared with another student.

• Stop before upload if sending the documents to Anthropic for marking is not appropriate.
• Download or delete a saved upload from Past Papers and delete the account from Settings.
• Use the support guide for failed requests and marking concerns. A monitored public escalation channel is still a launch requirement.