siddharthmb's picture
game-structure generality v1 (note 0086)
a81d5b1 verified
|
Raw History Blame Contribute Delete
6.35 kB
metadata
pretty_name: 2026.RA.Game-Structure-Generality
tags:
  - negotiation
  - multi-agent
  - cheap-talk
  - llm-agents
configs:
  - config_name: episodes
    data_files:
      - split: train
        path: episode_rows.parquet

2026.RA.Game-Structure-Generality

Does a free-form public message channel matter more for language-model negotiators when coordination is harder? This dataset holds every episode of the game-structure generality experiment (research note 0086 of the rational_agents project): tables of Claude Opus 4.8 negotiators with the talk channel open (talk) and closed (mute, moves only), and tables of a model-free Bayesian negotiator (bayes), played on four game structures:

experiment-name structure bank rule seeds episodes per role
game-structure-generality-v1/S3 3 parties banks/instances_n3_private_v1 (24 instances) unanimity (3 of 3) talk/mute 0–4, bayes 0–4 120
game-structure-generality-v1/S5 5 parties instances_five_seat_private_v2 (24 instances; in the project repo) unanimity (5 of 5) 0–4 120
game-structure-generality-v1/S7 7 parties banks/instances_n7_private_v1 (24 instances) unanimity (7 of 7) 0–4 120
game-structure-generality-v1/Q4 5 parties bank v2 re-materialized at 4 of 5 (veto seat retained) 4 of 5 talk/mute 0–2, bayes 0–4 (table keeps 0–2) 72
game-structure-generality-v1/V1ctx, V1q4ctx 5 parties, Bayesian only (context) instances_five_seat_private_v1 unanimity / 4 of 5 0–4 120

The S5 talk and mute rows are derived from the model-breadth lane's Opus 4.8 cells (mutebreadth_opus48_v1__all_llm, mutebreadth_opus48_v1__all_llm_moves_only); their raw episodes are not re-uploaded here. Every other cell's raw episodes are in runs/.

Headline (95% instance-clustered bootstrap, 10,000 draws; E1 = mute − talk deal rate, paired; E2 = talk − Bayesian, paired)

structure talk mute Bayesian E1 E2
3 parties 0.925 0.808 0.175 −0.117 [−0.242, +0.017] +0.750 [+0.608, +0.858]
5 parties (bank v2) 0.908 0.600 0.233 −0.308 [−0.425, −0.192] +0.675 [+0.558, +0.783]
7 parties 0.817 0.467 0.267 −0.350 [−0.483, −0.225] +0.550 [+0.383, +0.708]
4 of 5 (bank v2, seeds 0–2) 0.958 0.875 0.431 −0.083 [−0.167, −0.014] +0.528 [+0.375, +0.681]

Muting cost difference 7 vs 3 parties (cross-bank): −0.233 [−0.408, −0.058]; 4-of-5 vs unanimity on identical keys (paired difference-in-differences, n=72): +0.194 [+0.083, +0.306]. Full tables, the discovery/coordination census, ICCs and seeds 0–2 estimates are in analysis/results.md and analysis/summary.json.

Files

  • episode_rows.parquet — one row per episode (config episodes): experiment-name, role (talk / mute / bayes), source_run, model, instance and seed keys, deal, normalized score, parties left below threshold by a closed deal, walks, the zone-of-agreement census columns (zopa_tabled, deal_given_tabled, failure_type, ...), voided and fabricated turn counts, published-message count, cost.
  • runs/*.tar.zst — raw run directories (episodes with every turn's exact prompt view and model output, instances, manifest, usage ledger, run.log with the invocation). *_merged = source cell plus its fill (missing keys re-run after a budget-guard stop); *_fill01 and the unmerged source dirs are kept for provenance; *__s34 = the seeds 3–4 extension cells. Unpack with tar -I zstd -xf <file>.
  • transcripts/*.jsonl — per-turn transcript projections of the hosted cells (seeds 0–2).
  • banks/ — the 3- and 7-party banks and outcome-blind feasible-set statistics of every bank.
  • analysis/ — results.md, summary.json, the analysis config and log, preregistration.md (with amendments A1–A4), exact_prompts.md (verbatim first-turn prompts of each hosted cell), spend.txt.

Protocol

Hosted seats: anthropic:claude-opus-4-8, thinking on (adaptive), 16,384-token turn floor, private score sheets and whole-number thresholds, datacenter framing, canonical prompt scaffold, four rounds with rotating proposer then a forced final vote. Mute = protocol arm moves_only (no message is published). Claude Opus 5 and Claude Fable 5 decline this prompt at the provider level (the block was localised to the scratchpad instruction), so the experiment uses Opus 4.8; no reworded prompt was used for any reported result. 0 refused turns and 0 fabricated turns in every hosted cell.

Regenerate

From the project repo (experiments/rational_agents), with PYTHONPATH=libs/interlens/src:.:

# banks
python generate_five_seat_bank.py --out instances_n3_private_v1 --n-parties 3 --first-seed 53000
python generate_five_seat_bank.py --out instances_n7_private_v1 --n-parties 7 --first-seed 57000
# one cell (repeat per structure/table; see run_game_structure_cell.sbatch)
sbatch --partition=john --cpus-per-task=2 --mem=16G --export=ALL,RUN=gsg_v1__n7__all_llm,TABLE=all_llm,BANK=instances_n7_private_v1,SEEDS="0 1 2",MIN_ACCEPT=7,BUDGET_USD=210,EST_USD=2.0,CONCURRENCY=8 run_game_structure_cell.sbatch
# fill a cell stopped by the working-capital guard
sbatch --export=ALL,SRC=gsg_v1__n7__all_llm_moves_only,TABLE=all_llm_moves_only,BANK=instances_n7_private_v1,SEEDS="0 1 2",MIN_ACCEPT=7,BUDGET_USD=36,EST_USD=2.0,CONCURRENCY=2 fill_game_structure_cell.sbatch
# analysis
python analyze_game_structure_generality.py --config game_structure_generality_v1/config_final.json --out game_structure_generality_v1/analysis_final

Every cell's exact command is the first line of its run.log. Quorum cells use tables all_llm_quorum_4, all_llm_moves_only_quorum_4, all_rational_quorum_4 with MIN_ACCEPT=4.

Cluster paths

Run dirs: /nlp/scr/siddharth/ii_mats/rational_agents/gsg_v1__*; transcripts: /nlp/scr/siddharth/ii_mats/transcripts/rational_agents/game_structure_generality_v1/; Slurm logs: experiments/rational_agents/game_structure_generality_v1/slurm/. No wandb runs. Total API spend $718.79 (all smokes included).

Related: siddharthmb/2026.RA.Five-Seat-MovesOnly-Control (the original Opus 5 bank-v1 mute control), siddharthmb/2026.RA.Quorum-Rounds-Sweep (Bayesian quorum sweep, bank v2).