Pnyx

Listening infrastructure for public discourse. A Chrome extension and demo that makes listening visible on social media feeds.

Built for the Agora Hackathon x TUM.ai E-Lab (April 2026), Challenge Cluster 02: "Shift digital interaction from performative reaction to visible meaningful listening."

What it does

Pnyx runs two AI models entirely on-device (no data leaves the browser) and adds four behavioral layers to social media:

SEE - Badges on posts showing what claims are being made and whether reasoning is present. Powered by the Habermas model (DeBERTa-v3-small, F1 0.974 against its training-label distribution — see the note below) and an 85-signal heuristic tier for AI-generated text detection.

PAUSE - When you click Reply, an overlay appears showing the extracted claims as checkboxes. Pick what you're actually responding to before writing.

SHOW - Your reply carries a visible tag: "Responded to: [claim]". Other readers can see what was heard.

EXPLORE - Branching deliberation on any claim. Three beats (steel-man, perspective-take, consequence-trace) via DeepSeek V3 through OpenRouter.

Theoretical framework

Three thinkers encode directly into the technical choices:

• Habermas - The scoring model maps two validity claims from Theory of Communicative Action: Wahrheit (truth/claim risk) and Richtigkeit (rightness/argument quality). Discourse legibility, not judgment.
• Arendt - The Erscheinung model detects genuine human presence (Erscheinungsraum). The third validity claim: Wahrhaftigkeit (sincerity). Is there a "who" behind the text?
• Mouffe - Explore mode treats disagreement as a democratic resource (agonistic pluralism). Escalation is a valid deliberation move, not a failure state.

Quick start: Demo page

cd scaffold/frontend
python -m http.server 8080
# Open http://localhost:8080/index.html

Models load on first visit (~460 MB, cached in browser after first load). Seven pre-scored posts demonstrate all four layers.

Model files

ONNX model files are not included in this repo (too large). Download from Hugging Face:

Tokenizer files and configs are included in the repo. If model files are missing, the demo falls back to heuristic-only scoring.

Quick start: Chrome extension

cd scaffold/frontend/extension
bash setup.sh

Then in Chrome:

1. Open chrome://extensions
2. Enable Developer mode
3. Click "Load unpacked" and select the scaffold/frontend/extension/ directory

The extension injects Pnyx into LinkedIn and X/Twitter feeds. setup.sh copies model files, detection modules, and vendor libraries into the extension directory.

Architecture

All scoring runs on-device via ONNX Runtime Web (WASM). Zero API calls for inference.

How inference works

1. Post text enters a Web Worker (inference-worker.js)
2. Erscheinung heuristic tier runs first (<1ms) - if score >= 4.0, ML inference is skipped
3. Both ONNX models run in parallel (Habermas: 256 tokens, Erscheinung: 128 tokens)
4. Scores feed into claim-extractor.js for sentence-level claim extraction
5. Badges, Pause Layer, and Reply Tags render from the extracted claims

Training pipeline

The scripts/ directory contains the full data generation and training pipeline:

1. 01_source_texts.py - Source text collection from disinformation datasets
2. 02_label_texts.py - Automated labeling with discourse quality rubric
3. 03_balance_and_report.py - Dataset balancing (2,500 samples per cell)
4. 04_generate_boundary_data.py - Targeted examples for failure modes
5. train_v3_modal.py - Model training on Modal (T4 GPU, focal loss, ~27 min/epoch)

Final dataset: 10K balanced samples + 453 boundary examples. Habermas v3 achieves F1 0.974 (0.977 claim risk, 0.972 argument quality).

What that number measures. Read it as a reported internal macro-F1 on a random row split of model-labeled data, not as discourse-detection accuracy.

Labels come from google/gemini-2.0-flash-001 (scripts/02_label_texts.py), with boundary cases from scripts/04_generate_boundary_data.py. The score therefore measures how closely the distilled DeBERTa model reproduces the labeling model; no human-annotated evaluation set exists in this repository.

Two caveats on the split. scripts/train_v3_modal.py shuffles rows and takes a percentage validation slice rather than splitting on unique text before boundary examples are oversampled, so isolation of duplicate texts across the train/validation boundary is unverified. scripts/eval_v3.py samples from the full labeled file rather than a reserved split. The dataset, checkpoint, and split manifest behind 0.974 are not published here, so the figure cannot be reproduced from this repository alone.

Project structure

scaffold/frontend/
  index.html             Demo page (all four layers)
  inference-worker.js    Web Worker for ONNX inference
  claim-extractor.js     Sentence-level claim extraction
  pause-layer.js         Pause overlay before reply
  reply-tags.js          Visible listening tags
  explore.html/js/css    Branching deliberation
  detection/             Erscheinung heuristic pipeline
  model/                 Habermas ONNX + tokenizer
  model-slop/            Erscheinung ONNX + tokenizer
  extension/             Chrome MV3 extension

scripts/                 Training pipeline
data/                    Demo datasets + source credibility
docs/                    Model contracts (architecture, I/O specs)
kb/                      Research notes (problem space, prior art)
pitch/                   Presentation slides + demo recording

Design

Espresso palette. The visual language avoids traffic-light color coding (no red/green quality judgments) in favor of saturation and density to signal discourse complexity. See DESIGN-LANGUAGE.md for the full design system.

License

MIT