# FeedForward: formative feedback for open-ended process modelling tasks

> A SolidJS web application for individual feedback on open-ended BPMN tasks, without grading by AI. Deterministic checks run first, the AI suggests and phrases, and the instructor decides every finding that reaches a student.

- Role: Designer and developer
- Period: 2026, University of Bremen (ISBPM seminar)
- Status: In progress
- Stack: Bun, SolidJS 2, BPMN, LLM workflows, Human in the loop

## Context and problem

Open-ended case-study exams may build argumentation and transfer skills better than tasks with one sample solution. The cost is that no solution key exists. Every student models a different process and justifies different choices, so individual feedback during practice does not scale with cohort size. Students enter the exam without feedback, and the office-hour model may mainly reach students who are already on track.

The seminar is "Information Systems and Business Process Management" at the University of Bremen, taught in English. The gap appears to be structural rather than a matter of instructor capacity.

## Approach

FeedForward is a working environment for the practice phase. It is built around one division of labour: the software handles repetitive checks, the model suggests and rewrites, and the instructor makes every subject-matter judgement.

1. **Deterministic pre-check.** Model syntax, token flow and calculation results are checked against reference values without AI. Students see the outcome immediately.
2. **Suggested findings.** The AI matches a submission against finding patterns the instructor defined per task. Unknown pattern IDs are dropped, and any observation outside the patterns must name a rubric criterion and pass a grade-language check. It never assigns a grade.
3. **Instructor decision.** The instructor confirms, rejects or adds findings in short bullet points.
4. **Explanatory text.** The AI turns the approved bullets into feedback addressed to the specific artifact. It may explain only the bullets it was given.
5. **Review and release.** A student assistant checks the text for fidelity, factual errors and linguistic fairness. The instructor releases it. Nothing reaches a student unapproved.

Pre-check results are mechanical and can be shown without release. The release gate applies to feedback text, and the interface says so.

## Interface

- BPMN workbench, in which students model and submit their process.
- Immediate deterministic pre-check results for the student.
- Instructor findings console to confirm, reject or add findings.
- Composition and release view with meta-review, so nothing reaches a student unapproved.
- Admin UI for task repositories: register, inspect, import, pin and retire, always inspect-then-apply with a visible diff.
- Screen-reader-friendly text description, generated from the linearized process model.

### Content and system are separate

The tool ships no tasks. Task libraries live in separate Git repositories that follow a published content contract. Instructors register, inspect, import, pin and retire a repository in the admin UI. Import is always inspect-then-apply, so an instructor sees the diff before any task changes, and a single broken task fails alone without blocking the others. Released feedback is never rewritten when a task later updates.

### Architecture decisions

- **Bun and SolidJS 2 RC** for the web app, with Drizzle over `bun:sqlite`. The stack was verified against the pinned versions before implementation started.
- **Headless BPMN package.** Parsing, linting and linearization run through bpmn-moddle and bpmnlint with no UI dependency. The linearized model also feeds a screen-reader-friendly process description.
- **AI provider behind a port.** A mock provider and an academic-cloud provider sit behind one interface. Inputs are pseudonymised, and each interaction is recorded with its prompt version, so results can be segmented by prompt later.
- **Post-checks on AI output.** Replies that are not the expected JSON shape fail the run. Unknown pattern IDs are dropped. Grade-like language is rejected in unlisted observations and drafts.

## Status

The build plan is implemented through its final milestone: content contract, import pipeline, Mode A BPMN workbench with submission packaging, deterministic pre-check, AI provider layer, instructor findings console, composition with meta-review and release, and the validation-study export with operations documentation. The suite covers unit, property and integration cases. The repository is private.

Open: the first live course run is planned for winter semester 2026/27. Until real submissions exist, the validation of suggested findings against the instructor's decisions is not run, and no results are claimed.

## Transfer

The same division applies to any LLM-assisted review where a human remains accountable: deterministic checks first, the model limited to selecting from a fixed vocabulary and rephrasing decisions already made, and a release step owned by a person. Separating the content repository from the tool also means another instructor can bring their own tasks without changing the core.