← checkpoint 25 analysis · the two attractors
divide_dollar — every published episode
canonical family holdout: a deterministic divide-the-dollar preset, never trained on
30 transcripts across 2 arm(s), re-rendered from the
stored episode JSONs with the current viewer. actions lists the parsed act of each turn in order:
none means the seat emitted no negotiating act at all, which is a different failure from walking
away. max share is the largest normalized surplus any single seat realized — 1.000 means one seat
took the entire available surplus. Fabricated turns: 0.
transcript is the full text of the episode — every turn, the game setup, the outcome. chart is the interactive view that additionally places each deal against the IR-feasible frontier (the efficient deals that are also acceptable to everyone, so that a deal which is Pareto-optimal only by pushing a seat under its threshold is drawn as unreachable rather than as a target) and shows each turn beside its post-hoc oracle counterfactual. Charts are rendered for 30 of 30 published episodes: in full for the canonical presets, which are this hub's exhibits, and as a bounded per-run sample for the generated banks, where a complete chart export would run to gigabytes. Every episode has its text transcript regardless.
| arm | instance | seed | deal | named deal | normalized surplus per seat | max share | turn actions | transcript | chart |
|---|
Source artifacts
- baseline:
/nlp/scr/siddharth/ii_mats/rational_agents/grpov2eval_baseline_divide_dollar - trained:
/nlp/scr/siddharth/ii_mats/rational_agents/grpov2eval_lam1_checkpoint-25_divide_dollar
These are the run directories the pages were rendered from; each also holds its own
manifest.json (with the exact run.py invocation) and run.log.