Setting up grading rubrics

Create a rubric

1. Open Activity settings → Rubrics → New rubric in the MATCHA dashboard.
2. Choose a settings group and enter the title. In an LMS launch, your own settings can be reused across your courses; organization-shared rubrics can be viewed and used, but not changed.
3. Add 1–12 criteria. Give each a positive weight with at most two decimal places; the weights must total exactly 100%.
4. For each criterion, choose no grid or 2–6 levels. Each level has a distinct percentage between 0 and 100, a label, public description and private staff notes. Public and private text can also be entered for the rubric, criteria and adjustments (up to 2,000 characters per field).
5. Choose whether ordinary scores may fall between levels and whether totals may exceed 100%.
6. Optionally add up to ten bonuses or penalties, with a minimum, maximum and assessor. They change the total by percentage points: 72% minus a 30-point penalty is 42%. Only instructor-assessed adjustments may be uncapped. System adjustments require human confirmation; no automatic scoring is implied.
7. Select Save rubric. Correct any errors shown beside the fields. A complete definition must fit within 64 KiB.

Attach it to an activity

Save the activity first, then open its Grading settings and select a rubric. You can attach or edit a rubric after the activity has started.

Attaching a rubric immediately requires instructors and graders to use a rubric-capable MATCHA desktop release. Update every staff client to the latest compatible release and reopen grading. Older clients may show a generic error when accessing the activity or downloading submissions. The dashboard shows the required version once attached. Removing the rubric does not undo this staff requirement.

Students can continue participating and reading released feedback. Existing grades and feedback are preserved. Attaching a rubric does not infer criterion scores from a prior total; those submissions must be assessed with the rubric before changing their grade. Preserve and reconcile any unsynced staff edits after updating.

Read and acknowledge the update notice, then select Save rubric settings. These settings save separately from other activity settings. An activity instructor must make this change; ordinary grading access alone is insufficient.

Context and student notice

The curriculum text starts from the course's curriculum label and can be customized up to 120 characters. An optional target mean from 0 to 100 is informational; saving it never changes grades. Leveling is a separate instructor action.

Assignment instructions and instructor-provided template text are included in AI grading context by default; each can be switched off. Existing instructor comments are always included when present. Work metrics and analysis are also used. AI grading is unavailable at Privacy Level 3 and above. Saving context options does not start an AI run.

You can enable and customize a student notice. It appears only on released AI-assisted assessments, not purely manual assessments. The wording is saved with released assessment feedback.

Grading with a rubric

Open a submission from the grading list in the MATCHA desktop app. On an activity with a rubric, the grade field shows the grade calculated from the rubric and cannot be typed into. Select Rubric beside it, or the Rubric tab, to grade the submission.

1. For each criterion, choose a level or enter a percentage from 0 to 100. If the rubric does not allow scores between levels, choose one of the levels.
2. Write feedback for the student under each criterion. The private rationale, evidence and staff notes are visible only to instructors and graders; they are never released to students.
3. Apply any bonuses or penalties. They add or subtract percentage points from the weighted total.
4. Check the total as you work, then select Save rubric. Discard changes returns to the last saved version.

A criterion without a score leaves the assessment incomplete. An incomplete assessment has no grade and cannot be selected for release; this is different from a score of 0. There is no separate total to edit: change the criteria or adjustments instead.

A grade entered before the rubric was attached stays unchanged and can still be released. MATCHA does not infer criterion scores from it. To grade that submission with the rubric, choose Fill rubric; saving replaces the earlier grade with the rubric total.

If scores were leveled, each leveled criterion shows its base score, factor and current score. Directly editing a criterion’s score sets a new base and removes leveling for that criterion only; the other criteria keep their bases.

If the LMS grade scale is unavailable, the rubric shows percentages only. Reopen grading from the LMS before saving.

Reviewing and selecting rubric assessments

Opening or editing a rubric does not mark it reviewed. Select Mark reviewed to record that you checked this exact version without selecting it for release. Selecting a submission for release in the grading list also records review of the assessment shown in the list. Graders and instructors can both do this.

Saving a changed rubric assessment requires it to be reviewed again and clears its release selection. Editing only the general comment keeps the review. A change waits for the server before it can be reviewed; the grading list shows Syncing until then.

The grading list summarizes each submission as Not assessed, Incomplete, Not reviewed, Reviewed or Reviewed · ready. Submission changed means the student’s file changed after the assessment; review it again before releasing.

If another grader saves first, your unsaved edits are not applied to their newer version. Copy anything you want to keep, choose Reload latest version and reapply your changes. A read-only review window shows the rubric without allowing changes. Saving the project or closing the submission or window saves unsaved rubric edits first; if they cannot be saved, the submission stays open.

Releasing rubric feedback

When the release selection includes a rubric assessment, Release student feedback (or Submit grades and feedback) first asks how to write the general comment. Students using a current MATCHA version, and students on an LMS course page, see the released rubric: each criterion's score, level and feedback, applied bonuses and penalties, the total, and the student notice of an AI-assisted assessment. Private rationales, evidence, staff notes, base scores and leveling factors are never released. The general comment is shown as well.

Add a compact rubric summary to the general comment follows your comment with criterion scores, levels, applied adjustments and the total, then the complete student notice of an AI-assisted assessment. This is what the LMS and older MATCHA versions show. The option is on by default for LMS activities; elsewhere it starts from the submissions' previous choice. Your editable comment never changes: each release generates the summary again from the exact assessment being released, so it is never added twice or left over from an earlier version.

The preview shows the general comment each selected student will receive, including the grader name when Add grader names to comments is on. The general comment is limited to 2,048 characters, including the grader name. Only the generated summary is shortened to fit; it then ends with “(full rubric in MATCHA)”. MATCHA never shortens your comment or the AI notice: if they do not fit, that submission is not released and the dialog explains why. Shorten the comment, turn off the summary or turn off grader names, then release again.

Changing these options for a submission first saves a new feedback version, without changing its scores or review. Previously released feedback stays available to the student and LMS until the release succeeds. Later comment-only updates keep the chosen options. Releasing again with unchanged options creates no new version. Leveled, undone and other changed assessments must still be reviewed and released explicitly, and their summary always shows their current scores.

Leveling scores

An instructor with authority for the activity can select Level scores in the desktop grading toolbar. Assigned graders cannot level or undo leveling. This manual action needs no AI subscription or credits and is available for manual rubrics even where AI grading is unavailable.

1. Choose All submissions, AI-graded (reviewed or not), or AI-graded (not reviewed). There is no default scope. AI origin is retained after human edits; pending comments alone do not make a reviewed assessment unreviewed.
2. Enter a target mean from 0 to 100. A saved activity target is offered as the starting value; it never applies itself.
3. Select Preview changes. Check the eligible and skipped counts, rubric revision groups, base/current/proposed scores and totals, factors, achieved mean and residual. Expand individual rows for criteria, levels and rank changes. Ungraded, incomplete, stale-submission and scalar-only grades are skipped; historical rubric revisions remain usable. The scope covers current submissions across the activity, independently of list searches.
4. Select Apply bounded preview only when you want the displayed change. Cancel preview leaves grades untouched. To alter the scope or target, select Change scope or target and preview again.

Each criterion is multiplied from its last directly assigned base and capped at 100%. A new leveling replaces earlier factors: a base of 50 becomes 60 at ×1.2 and stays 60 after another ×1.2. A direct edit to 55 establishes a base of 55 for that criterion, so a later ×1.2 gives 66. Other criteria keep their existing bases. Bonuses and penalties do not change.

The applied factor is limited to ×0.8–×1.2, inclusive. For a base mean of 50, target 75 requests ×1.5 but applies ×1.2, achieving 60. Target 25 requests ×0.5 but applies ×0.8, achieving 40. Factor limit reached shows the achievable result; there is no override. Zero bases cannot increase, and clamping can change levels and rank. A preview with no score changes leaves revisions and release selections intact.

Leveling preserves qualitative criterion feedback and in-text comments, including the initial prose of AI drafts. It does not regenerate feedback; inspect wording that no longer fits. Review carries forward only from already reviewed assessments. Unreviewed work stays unreviewed, and pending comments still need normal review.

Release remains separate. Changed rows become new drafts and their release selections are cleared, even if previously released. Students keep the previous published feedback and LMS grade until a separate explicit release. Students never see direct-score bases or leveling factors.

Undo is available on the computer where the leveling was applied. To undo, inspect a committed operation under Saved operations (this device), choose Preview undo, then Apply undo. This restores exact saved scores, bases and prior factors, including an earlier leveling; it does not divide by the factor. Undo also creates drafts requiring explicit release. If any affected assessment has changed since the operation, the whole undo is refused and intervening edits are kept.

After a conflict, save open reviews, sync outstanding changes, then choose Refresh preview inputs and preview again. If a reply was lost or the app restarted, inspect the saved operation or use Check saved status and access before retrying. Retrying an existing operation uses the same identity; it never applies an extra factor. A pending preview may be cancelled after its status is known.

Edits, conflicts and interrupted saves

Editing a reusable rubric affects future assessments in every activity using it. Existing assessments keep their original rubric revision and grade. Deleting a rubric removes it from the picker and clears current activity selections; it preserves assessment history, grades and the staff update requirement.

If another instructor saves first, your draft stays visible. Copy any text you want to keep, then choose Reload saved version (or Reload activity configuration) and reapply your changes. Reloading asks before discarding unsaved edits. A stale draft never silently overwrites a newer revision.

If attachment is pending or the network response was lost, use Retry saved request in the same browser tab. Its exact request is retained through a reload when browser session storage is available. Do not start a replacement attachment. Grading may be temporarily unavailable until the transition finishes. Another tab can check the status; if the original request is no longer available, contact support with the displayed operation identifier.

Organization AI credits

Account administrators manage the pool in Organization admin → AI credits. The pool is pay-as-you-go, with no included allowance or expiry/reset. It is separate from student seat wallets, seat licenses and personal credits. Self-organizations offer personal funding only. These account administration controls are unavailable in an LMS launch; LMS instructors can use organization funding when starting an authorized grading run.

The page shows the balance, reservations and available credits, with usage by course, instructor and feature. Refresh before making changes if the last read failed. Course quotas cover the course's lifetime; instructor quotas renew each UTC calendar year on January 1. Limits include spent and reserved usage. Blank means unlimited and zero blocks spending. A lower quota preserves past usage and blocks further calls while the limit is exceeded. A quota does not grant instructor access.

Allow AI-assisted grading defaults on. Turning it off stops new calls, including personally funded grading for this organization, while retaining completed assessments and settling confirmed usage. Service enablement is separate: an enabled organization preference cannot turn on a feature that is unavailable for the service.

Use Buy organization AI credits to choose packs at the displayed server price and currency, then continue to secure checkout. Prices and credits per pack come from the same catalogue as personal pay-as-you-go credits. Organization funding does not require a personal subscription. After returning, use Check payment and refresh balance until payment is verified; an unpaid or canceled checkout adds no credits. Retrying a failed checkout keeps its original request ID, quantity and price. Do not create a replacement for an uncertain purchase; contact support with its request ID if recovery is needed.

Once reconciliation confirms a checkout unpaid, Discard unpaid checkout removes its saved shortcut from this browser tab and lets you change the quantity for a new purchase. It does not cancel the original Stripe checkout: if that checkout is paid later, its credits are still added to the organization. The action asks you to confirm this. An uncertain or failed reconciliation must be recovered with the same request ID before a replacement is offered.

Each model call reserves its maximum cost, then charges confirmed usage and frees the remainder. The same usage markup applies as for personal AI. Refused or malformed output and a separately reserved repair still incur known usage. Canceling stops future calls, without refunding confirmed usage. An unknown-usage hold is released after 24 hours and MATCHA bears that unknown cost. A paid result lost in transit requires an explicit, newly estimated paid retry; it is never repeated automatically.

Successful partial refunds reverse purchased credits proportionally. An open or lost dispute reverses the whole purchase without adding the refund twice; a dispute win restores only credits not permanently refunded. A reversal can leave pool debt, which blocks new spending until covered. Debt does not cause automatic collection or delete assessments and usage history. A refund recorded as successful that later fails requires support reconciliation rather than automatic restoration.

MATCHA superusers can grant organization credits from the organization's superuser detail page. A reason is required and retained with the actor and operation ID. A failed grant or settings save can be retried with its exact saved ID and values in the same browser tab, including after reload when session storage is available. A revision conflict requires refreshing before making a new settings request.

A definite grant validation rejection leaves the entered amount and reason editable. Correct the draft and submit a new grant. An uncertain failure or operation conflict keeps its original ID for recovery. Balance and usage refreshes preserve grant drafts and confirmations; use Refresh price to update the credit catalogue explicitly.