Every grade follows the same four steps and is fully reproducible. The deterministic core does the heavy lifting and is the authority; AI only assists on judgment calls, and you can switch it off entirely.
Internally: locate · extract · compare · score.
GridGrade identifies each answer by its content, role, and context — not a hard-coded cell. Students can rename worksheets, move a block ten rows down, or transpose a layout; the answer is still found. When a match is genuinely ambiguous, GridGrade flags it for your review rather than guessing. Every result records exactly where the answer was found.
Sheet "Q3 Analysis" · cells D12:G18
matched by role: summary pivot
confidence: high
We parse what's actually in the workbook: the formula behind a cell, the full pivot-table definition, the steps behind a Power Query, and chart specifications. That's what lets us grade method, not just the final number.
Read this student's work:
• Formula: West-region Q3 cost via SUMIFS
• Pivot: Region in rows, summing Cost
• Power Query: an inner join on Customer ID
GridGrade recognizes that many methods reach the same correct answer and treats them as interchangeable — SUMIFS, SUMPRODUCT, or a pivot; VLOOKUP, XLOOKUP, or INDEX/MATCH. It applies sensible tolerance (rounding, label spelling) and catches subtle traps (filtering after aggregation when it should be before, order-dependent steps, the right number for the wrong reason).
SUMIFS, SUMPRODUCT, or a pivot — same answer → accepted
filter applied after Group By → trap
Credit is awarded with dependency-aware partial credit: a correct step built on a wrong upstream value isn't penalized twice. Low-confidence results are surfaced for your review with the evidence attached. Errors are always explicit — never a silent zero.
Q4 join: 8 / 10 (correct method, upstream value wrong — not double-penalized)
Field placement, rows vs columns (including transposition when orientation isn't the point), aggregation type, filters, sort, grand totals, “Show Values As,” grouping, and whether the source range is stale. We grade the pivot's meaning, not its existence.
We decode the steps behind each query and read them as a transformation: joins (tables, keys, join kind, expanded columns), Group By (keys and aggregations), split/added columns, type changes, and step order. We then corroborate against the loaded output — for example, a row-count delta that proves an Inner join was used where a Left Outer was required. Honest scope: GridGrade reads and reasons about the underlying query (its M code) and checks the output; it does not re-execute your students' queries. The evidence we show is the parsed query plus the output delta.
Graded by what they compute. Helper-column decompositions, function-family equivalents, and coincidental-correctness traps are all handled.
Chart type and appropriateness, titles (including cell-linked), axis labels, and series-range correctness, with flexible label matching — it treats '#' and 'Number', 'Qty' and 'Quantity', 'FY' and 'Fiscal Year' as the same label.
The cross-cutting promise: credit every valid approach a student uses, and find the answer wherever it's placed.
A whole-workbook dependency graph means cascade-aware credit and process-vs-result scoring you can tune per assignment.
Grade the same workbook twice and get the identical result — same points, same deductions, same evidence. Volatile functions like TODAY() are pinned to the assignment's as-of date; anything driven by randomness is graded structurally and flagged, never guessed.
Each deduction carries what the student did, why points came off, and how to fix it — so a regrade request is a quick conversation, not a re-grade from scratch.
GridGrade is review-recommended: you remain the grader of record, you review flagged cases and a sample, and you can disable third-party AI entirely — the deterministic core stands on its own.
Found a mistake in your rubric after grading? Edit it, and regrade just the affected students in seconds — only the changed checks re-run. Every change is audited, your gradebook export shows exactly who changed, and affected students are re-notified automatically. Because grading is deterministic, the regrade is just the same grading, run again — not a mystery.
GridGrade is in active development for a Fall 2026 launch. We roll out each capability behind a measured quality gate and publish accuracy data as it clears — we won't claim a number we haven't proven. Until then, GridGrade is review-recommended: it does the work, you confirm the calls that matter.