Proofline is a research preview for turning handwritten mathematics into editable LaTeX without silently correcting the writer's reasoning.
- a browser-only demo for upload, review, editing, and export;
- an evaluation harness for transcription fidelity and consistency;
- training-data utilities for public datasets; and
- a small, fully synthetic public fixture set.
The deployed site currently runs in an explicitly labeled demo mode with placeholder transcription. It does not send uploads to a model backend.
The original development corpus contained private academic records and has been removed from the public repository, along with raw transcriptions and model outputs derived from it. Git history was rewritten so those source files are not retained on the public branch.
Only synthetic examples are published. They contain no real student work,
names, email addresses, student identifiers, signatures, or grades. See
data/README.md for details.
python3 -m http.server 8080 --directory siteThen open http://localhost:8080.
Regenerate the synthetic PDFs:
python3 scripts/create_public_fixtures.pyRun a candidate model against data/golden_set/manifest.json with
scripts/run_golden_set.py. Generated transcriptions and scorecards are
written under baseline_results/ and intentionally ignored by Git.
The public repository demonstrates the product interaction and evaluation method. A live fine-tuned checkpoint and production inference service are not currently included.