HN Showcase
Back to browse
Glass Box: writing editor that exports a verifiable PDF of your process

Glass Box: writing editor that exports a verifiable PDF of your process

by normanbell·Feb 24, 2026·2 points·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerZero to OneBold BetSolve My Problem

Proves AI usage with a verifiable PDF receipt instead of probabilistic detection guesswork.

Strengths
  • •Addresses real failure: Stanford data shows 60%+ false positive rates on AI detectors for non-native writers.
  • •Process transparency (keystrokes, AI calls, pastes) creates evidence chain; PDF export is portable proof.
  • •Local-first + open-source (Python/NiceGUI); no keystroke telemetry or cloud lockdown.
Weaknesses
  • •Requires academic institution buy-in; students can't unilaterally defend themselves against skeptical professors.
  • •Unknown adoption rate or real-world effectiveness; framing assumes professor will accept PDF as ground truth.
Category
Productivity
Target Audience

Students defending against AI detection accusations, educators seeking integrity verification, academic integrity advocates

Similar To

Turnitin · Grammarly · OpenAI text classification

Post Description

We built a local-first writing editor that records the composition process (keystrokes, paste events, AI interactions) and exports a PDF with annotation highlights + an AI usage appendix.

The motivation: AI detectors are genuinely broken. Stanford research showed >60% false positive rates for non-native English writers on TOEFL essays. The "solution" from institutions has been to buy more detectors or make the policies so broad that either the student can do nothing if they are accused or the teacher can do nothing if the student does not make an admission.

Instead of trying to detect AI use after the fact, we give students a way to record their process as they write. The exported PDF shows a professor exactly what was human-written, what was AI-assisted, and what was pasted from external sources.

Tech stack: - NiceGUI (Python) for the editor - Quasar Q-Editor (contenteditable) with JS paste/ghost-completion hooks - Dual PDF engine: WeasyPrint (CSS rendering) with ReportLab fallback (pure Python, works on Windows) - Ollama integration for local AI (recommended model: qwen2.5:0.5b, ~400MB) - TWFF (Tracked Writing File Format): ZIP-based container with XHTML content + process-log.json with SHA-256 chained hash

The format spec (TWFF) is open. Anyone can implement it. We're trying to get at least one other tool to support it before calling it a "standard".

Demo: https://firl.nl/twff Spec: github.com/Functional-Intelligence-Research-Lab/twff

Similar Projects