TA guide

One page: check a whole batch of sealed submissions at once. Time: a few minutes for a full section, most of it unzipping.

The five steps

1. On the Canvas assignment page, choose Download Submissions. Canvas gives you one zip of every student's file.
2. Unzip it.
3. Open knobe.org/studio and choose the Cohort check tab.
4. Load the journey file (or the assignment spec) under What should these chain to? This sets the expectation: submissions should link to that file's fingerprint.
5. Drag the unzipped folder onto the drop zone. Every readable file is verified on the spot, in your browser. Nothing is uploaded.

Reading the table

One row per file. Problem rows sort to the top.

• Integrity: verified means the file is byte-identical to what the student sealed. failed means it was altered after sealing. unreadable means it is not a sealed file at all (often a plain essay uploaded without sealing).
• Chain: linked means the file references the assigned spec or journey by exact hash. no link means it chains to something else, or to nothing.
• AI disclosure: the student's own declaration of what an AI did, quoted from the file. none declared is not an anomaly by itself; under a disclosure policy it means the student declared no AI use.
• Review: the submission status the student sealed, usually submitted.

A failed or no link row is a conversation, not a verdict. The table tells you what the file says about itself; what that means is course policy.

Exporting and grading

Download the table as CSV gives you the same rows for gradebook cross-reference. The CSV contains only what students declared inside their files; no Canvas data enters the page. Grading itself stays in SpeedGrader, where a sealed file displays as plain text.

Seal a cohort receipt

After a check, Seal a cohort receipt (under the table) seals a KNOBE attesting that these files, with these payload hashes, produced these verdicts when you ran the check. Keep it with the course records; in a dispute, it fixes what existed at grading time, and anchoring its single payload hash anywhere with an independent timestamp (an email to yourself, a gradebook note) timestamps the whole batch.

Two things that are normal, not anomalies. Large cohorts produce large receipts: a 500-file check seals in about a second to a file of roughly 200 KB, which is fine for the protocol and for storage. And a resubmission that Canvas renamed (the same bytes as essay1.knobe.md arriving again as essay1-1.knobe.md) appears as two rows with different file names and identical payload hashes; identical bytes under the identical name are merged into one row.

One student's history

Drop a single student's files from across the term and press Chain view. Each file lists what it builds on and whether the fingerprints match, which is the record you want in a regrade or integrity conversation.

If you run an AI agent

The same checks run agentically: install knobe-mcp (see knobe.org/mcp) in Claude Desktop or any MCP client, point it at the unzipped folder, and ask it to verify the batch and tabulate declarations. It runs locally, so the privacy posture is unchanged.