Cross-paper analysis

# Cross-paper analysis

Across a batch of submissions for the same assignment (e.g., "all the Macbeth essays"), Yawp surfaces common patterns: what students consistently nailed, what they consistently struggled with. Teachers use this to target the next lesson.

## Problem

Today a teacher has to read 30 essays and form their own gestalt about where the class is strong and weak. The information is sitting in 30 graded submissions; nothing rolls it up. As a result, the next-lesson decision is based on whichever essays the teacher remembers most vividly — usually the worst and the best.

This is high pedagogical value: the teacher's next lesson is the lever, and informing it from data they're already producing is high-leverage.

## Goals (rough)

- For a batch of submissions tied to a single AssignmentType, generate a class-level report:
  - Top 3-5 areas where students excelled.
  - Top 3-5 areas where students struggled.
  - Specific examples / quotes (anonymized) illustrating each.
- Teacher can use this to shape the next lesson or revision push.

## Non-goals

- Replacing per-student feedback. This is additive analysis on top of individual grades.
- Predictive scoring. This is descriptive of what the class actually did, not predictive of what's next.

## Dependencies

- [archive/released-grades-organization.md](released-grades-organization.md) — submissions must be reliably grouped by AssignmentType before batch analysis is possible.
- The GA must produce structured per-category feedback that can be aggregated (likely already true).

## Open questions

- [ ] Surface placement: a button on the AssignmentType folder view? Auto-generated when the batch is fully graded?
- [ ] What's the right output format — narrative summary, bulleted list, both?
- [ ] Privacy: anonymize student names by default in the cross-paper view, or attribute? Likely anonymize.

## Engineering handoff

Not ready. Idea-stage. Depends on released-grades-organization shipping first.