Agent Collusion Explorer

A static site for reading the trajectories behind Emergent Collusion in Long-Horizon LLM Agent Interaction. It carries its own copy of the data, so a clone needs nothing but a local web server.

Run it

python3 -m http.server 8000 --directory docs

Then open http://localhost:8000/.

A server is required: the page is ES modules and fetch, neither of which a browser will run from a file:// path.

What you are looking at

Three levels, in order:

1. Explore — one card per condition: the model or pairing, what the condition varies, and how its fifty runs came out.
2. A condition — its fifty runs, each a strip of ten episode outcomes (Alice's verdict above Bob's) with the collusion pathway and whether it converged.
3. A run — one trajectory, episode by episode: what each agent did on its own task, the messages the two exchanged, the verdict each returned, and the feedback and private reflection that followed.

What is in the repository

Path	What	Size
docs/	the viewer: one page, plain ES modules, no build step	~1 MB
docs/data/index.json	the browse index — every condition and run, with a compact per-episode encoding of verdicts, task correctness, judge labels and task type	~1 MB
docs/data/viewer/<condition_id>/repNNN.json.gz	one trajectory each, reduced to what the viewer draws	~120 KB each

The bundles are gzipped JSON, fetched on demand and unpacked in the page with DecompressionStream, then kept for the session.

Where the viewer looks for a trajectory

In order, stopping at the first that answers:

1. docs/data/viewer/ — this repository's own copy.
2. The Hugging Face dataset, under viewer/, for a deployment that ships without the bundles.

?data=<base> overrides both and uses only what it names, which is how you point the page at a freshly built copy:

http://localhost:8000/?data=/some/other/build/#/explore

Regenerating the data

Both tiers come from viewer/build.py in the research repository, run against a release directory:

python viewer/build.py --release release --out release/viewer

That writes docs/data/index.json and one bundle per run. Copy the bundles here as docs/data/viewer/.

Browser support

Unpacking a trajectory needs DecompressionStream (Chrome 80+, Safari 16.4+, Firefox 113+); the page says so plainly if it is missing. Everything else degrades quietly — motion is dropped under prefers-reduced-motion.