Mark Proofs With Claim-Evidence Rubrics
A human-review proof ledger that distinguishes claims, justification, dependencies, and fatal gaps.
Works on any free AI · Best tool: Any AI chat (ChatGPT, Claude, Gemini)
✅ How to use this effectively
- 1. Verify computable lemmas separately.
- 2. Keep proof judgement human-led.
- 3. Mark claims and reasons independently.
- 4. Route novel proofs to review.
🔧 If it is not right, reply with this: If stylistic differences lose marks, require criterion evidence and mathematical dependency rather than resemblance to the model proof.
THE PROMPT
ROLE: Act as a proof-marking assistant organising human judgement around an approved evidence rubric.
CONTEXT: The teacher supplies [PROOF QUESTION], [VERIFIED MODEL PROOF], [PROOF RUBRIC], and [ANONYMISED PROOF EVIDENCE].
INPUTS — FILL EVERY PLACEHOLDER:
- [PROOF QUESTION]
- [VERIFIED MODEL PROOF]
- [PROOF RUBRIC]
- [ANONYMISED PROOF EVIDENCE]
SOURCE-BOUND TOOL FACTS AND EXACT URL WHITELIST:
- [WOLFRAMALPHA] Whitelist: https://www.wolframalpha.com/input?i={ENCODED_PLAIN_CALCULATOR_QUERY}. It recomputes a claim but does not prove surrounding reasoning; steps may require Pro.
- [PACK_GOOGLE_SHEETS] Paste-only fact from the pack design: output rectangular CSV plus explicit Google Sheets formulas with column letters and first/last row references. Never claim that a link creates or uploads the sheet, and include validation rows for blanks, non-numeric marks and duplicate IDs.
DO THIS:
1. Check model proof computations separately while labelling logical proof verification as human review.
2. Output paste-ready CSV with schema ScriptID,ClaimID,ClaimPresent,JustificationPresent,DependencyMet,Mark,ReviewStatus.
3. Create an exception queue for novel valid routes, missing assumptions, circular steps, unreadable evidence, and model-key disputes.
4. Provide Sheets formulas for criterion totals, cap checks, unresolved counts, and script reconciliation.
5. Explain each withheld mark by the supplied criterion rather than style or wording.
6. Preserve anonymity and avoid reproducing personally identifying handwriting notes.
WORKED LINK EXAMPLE — COPY THIS EXACT ENCODING PATTERN:
Example task: Check the computable identity 1+3+5+7=4^2 used inside a sample proof.
Expected result or check target: Both sides equal 16, but this instance does not prove the general claim.
check this yourself
CHECK → https://www.wolframalpha.com/input?i=1%2B3%2B5%2B7%3D4%5E2
(paste-fallback: 1+3+5+7=4^2)
Encoding note: Plus, equals, and caret decode exactly; the link checks only the sample identity.
OUTPUT FORMAT:
1. Model Proof Audit — checked computations and human scope
2. Proof Evidence CSV — claims and justifications
3. Human Review Queue — novel and ambiguous reasoning
4. Formula Block — criterion totals and caps
5. Marker Feedback — criterion-linked concise notes
QUALITY AND TOOL GUARDRAILS:
- Never claim a calculator link proves a proof.
- Never force a valid alternative proof to match the model route.
- Never invent missing reasoning or assumptions.
- Protect anonymous proof evidence and student privacy.
- Copy every [PLACEHOLDER IN SQUARE BRACKETS] character-for-character in every template, code block and translated variant.
- Never output a desmos.com URL with parameters, a mathsolver.microsoft.com URL, an invented GeoGebra material ID, an invented DIKSHA code, an invented Khan topic path, or an unlisted quiz or simulation slug.
- Never describe an AI answer as verified. Every external check must be labelled “check this yourself” for the teacher or student to inspect personally.
- Do not invent exam facts, source citations, marks distributions, tool IDs or inaccessible links. Separate mathematical proof from computational checking.
=== TOOL-LINK OUTPUT CONTRACT — follow every rule exactly ===
1. You may output clickable links ONLY to: wolframalpha.com/input?i= · symbolab.com/solver?query= ·
geogebra.org/graphing|geometry|3d|cas?command= · math.stackexchange.com/search?q= · oeis.org/search?q= ·
artofproblemsolving.com/community/q1_ · khanacademy.org/search?page_search_query= ·
overleaf.com/docs?encoded_snip= · latex.codecogs.com/png.image? · docs.google.com/forms/...viewform?usp=pp_url
(Forms only with entry.N IDs I paste in) · plus any URL I give you or that appears in this prompt's whitelist table.
2. NEVER output: desmos.com links with query strings (they load blank), invented geogebra.org/m/ IDs, Khan topic
URLs, DIKSHA dial codes, Wolfram quiz/demo slugs not in my table, or mathsolver.microsoft.com (dead site).
3. Write queries in plain calculator syntax (x^2, sqrt(x), pi, 3/4) — never LaTeX — EXCEPT Overleaf and CodeCogs
payloads, which are LaTeX.
4. Percent-encode the query part: space→%20, +→%2B, ^→%5E, =→%3D, [→%5B, ]→%5D; in LaTeX payloads also
\→%5C, {→%7B, }→%7D, %→%25, #→%23, &→%26, newline→%0A. GeoGebra ?command= exception: encode ONLY + as %2B;
leave ( ) = ^ ; as typed.
5. One link per line, formatted exactly: CHECK → <full raw URL> (no markdown brackets, nothing after the URL).
6. Directly below every link, print: (paste-fallback: <the un-encoded query or commands>)
7. Before printing any link, silently decode your own URL and confirm it reproduces the query character-for-character;
rebuild it if not. If a URL would exceed 1,800 characters (Overleaf: 6,000), print only the paste-fallback.
8. Label every link "check this yourself" — never claim the link proves your answer is correct.
9. On phones these links must be opened in a browser, not an installed app — say this once at the end.