AI insights from revision history

# AI insights from revision history

Yawp records every keystroke. Two related ideas surface from that data: (1) summarizing a student's writing process for the teacher, and (2) flagging suspicious patterns (e.g., a student going from blank doc to a 92% submission in under an hour).

## Problem

The version history is rich (Kevin called it "SUPER helpful") but consuming it is manual — the teacher scrubs through edits to see the process. Two specific gaps:

**Thread 1: Process insights for the teacher.**
A teacher looking at a graded paper can't quickly see what happened during writing — where the student struggled, where they got into flow, what ideas they explored and discarded. Today they reconstruct this by sifting through revision history and tutor conversations. Not scalable.

**Thread 2: Integrity signals.**
Reported 2026-02-25: a student submitted a "suspiciously good paper" — went from minimal content to a 92% submission in under an hour. The teacher's options today are (a) scrub the revision history manually, or (b) go off gut feeling. Google Docs has a feature that plays the revision history as video; worth considering whether something analogous makes sense here.

## Goals (rough)

For teachers:

- A surfaced summary of a student's writing process per submission: time on task, where they spent the most effort, what changed in revisions, themes they explored.
- An integrity flag (subtle, not accusatory) when a submission's process looks anomalous — e.g., short total active time, large pasted blocks (overlap with [bugs/paste-alert-detection-broken.md](../bugs/paste-alert-detection-broken.md)), no exploration phase.

For students (later):

- A reflection nudge: "your strongest essays had ~3 hours of revision time spread over ~2 days." Self-awareness signals.

## Non-goals

- Disciplinary verdicts. Yawp surfaces the signal; the teacher decides what it means.
- Replacing paste-alert detection. Augments it; doesn't replace it.

## Domain notes

Depends on the existing revision-history infrastructure being queryable in aggregate. Connects to [features/student-growth-portfolio.md](student-growth-portfolio.md) — the same data feeds both surfaces.

## Open questions

- [ ] What's the integrity threshold? Probably needs calibration against real student behavior.
- [ ] Privacy/trust: how does this feel to students who know their process is being analyzed? Frame as feedback, not surveillance.
- [ ] Does this need its own ML model, or is the LLM enough given a structured timeline of edits as input?

## Engineering handoff

Not ready. Idea-stage with two distinct threads that could ship separately. Recommend tackling the process-insights thread first; integrity signals are more sensitive and need pilot validation before broader release.