Negotiation capsule

Question: How does a recorded agreement constrain the next movement?

This is a complete teaching workflow over two built-in scenarios. The explicit study.json freezes the methods, settings, seeds and sampling claim. It is not an article benchmark or evidence of a method advantage.

python examples/capsules/negotiation/run.py --output runs/negotiation-capsule

A core-only installed wheel suffices. Run from the repository root, or copy this whole directory elsewhere and run python run.py --output /path/to/new/output. The output directory must not exist. The study uses one worker, at most 120 batch seconds and 256 MiB of workspace artifacts; per-process limits are in study.json. Exports add a small amount of storage beyond the workspace quota.

Inspect receipt.json, experiment-card.json, and all all-trials.json rows. Each completed run has a checked bundle.json, offline replay.html and initial.svg. An agreed negotiation also produces negotiation-story.html and its exact event/frame references. The receipt records observed wall time and output bytes with the platform; peak RAM is explicitly unmeasured.

Expected qualification: all declared teaching trials should yield independent valid_solution receipts. The program preserves other outcomes and exits 2 if this expectation is not met; it never silently retries. A completed job alone does not imply a solution. Runtime/IDs and process timing may vary across machines.

Next: use the headless guide to declare your own population and budgets; retain unsuccessful outcomes in the denominator.