Perturbation Workbench
Research tool
I built the Perturbation Workbench to support the assignment-design procedure described in the AIED paper. It keeps the question versions, rubric, and graded attempts together so an instructor can review what was tested.
You run the chatbot queries yourself. Copy the question into the target chatbot, paste its response back into the workbench, and grade it against the rubric. The tool does not call a model or decide whether the answer is mathematically correct.
The workbench can export an assignment and a run manifest. The manifest records hashes, counts, settings, and timestamps. Those records support an audit of the workflow; they do not independently prove that the pasted response came from the named model or that the grading was correct.
Your session is stored in this browser, including question text and notes. A session export contains that working material. The run manifest is a separate export designed to omit question text, responses, and ledger notes. Review the file you intend to share.
Citing the method
Cite the AIED paper for the method. To describe use of the software, also name Perturbation Workbench and link this page.
What the test result means
A resistance label applies to the recorded question, target model, rubric, and attempts. It is not a guarantee about future model versions or extended student-chatbot conversations.