Building a Web of Trust to Fight LLM Spam

Tangled now natively supports vouching, letting users vouch for or denounce other users they've interacted with.

Vouched users get a green shield next to their profile picture, while denounced users get a red shield. This serves as a signal to filter out submissions made with LLM tools that look plausible on the surface but are subtly wrong.

The core design is as follows.

• Vouches and denouncements include a text reason field.
• Judgments are subject to attenuation, so only the judgments made by the user and their own circle are visible.
• Vouching or denouncing leaves a public record on the user's PDS.
• The Tangled appview aggregates these to display a vouch "hat" over profiles on issues, comments, pull requests, and pull request comments.

Currently, denounced users are not blocked; only a warning label is attached to parts of the UI. However, decay, where vouches weaken over time, and an evidence trace feature, which leaves the relevant PR as evidence if it is vouched for right after being merged, may be added.