2026.RA.Pure-Rounds-Sweep — does more negotiating time help a Bayesian table close?

Five automated negotiators must agree on one package out of 256. Each holds a private score sheet and a private walk-away threshold; a deal needs all five to accept. This corpus sweeps the one thing the frozen campaign never varied — the number of negotiation rounds before the forced final vote — across 4 to 256 rounds, for the two lineups that contain no language model and therefore cost only CPU: all_rational (five copies of a Bayesian agent that maintains a posterior over hypotheses about each opponent's sheet) and all_oracle (five agents that are simply told every sheet).

The headline: more time buys nothing. Deal rate across all_rational deadlines 4 → 128 ([4, 8, 16, 32, 64, 128]) reads 0.233 / 0.267 / 0.208 / 0.208 / 0.250 / 0.242 and normalized score 0.195 / 0.230 / 0.182 / 0.178 / 0.207 / 0.203, with every instance-clustered interval overlapping every other. Rounds-to-agreement tracks the deadline almost exactly (4.96 / 8.97 / 17.00 / 32.96 / 64.97 / 128.97 against forced finals one round later), so the agents use the whole horizon and finish where they started. The mechanism is in belief_turns.csv.gz: belief accuracy peaks at turn 5–6 and then drifts slowly downward for the rest of whatever horizon it is given, so extra rounds deliver no extra identification of opponents, and without identification there is nothing to concede to. The omniscient arm closes 1.000 of games at every deadline, confirming the horizon itself is not the constraint. Full analysis: research note 0061 in the source repo.

What is in here

file	rows	what it is
episodes.csv	1560	one row per played negotiation: outcome, welfare, Gini, IR violations, rounds used, and per-episode belief scalars
belief_turns.csv.gz	159015	the novel instrumentation — one row per (episode, believer seat, turn): posterior_mass_true_type, expected_utility_rmse, accept_set_f1
cells.csv	13	per-cell aggregates with instance-clustered 95% intervals
episodes/<cell_id>.jsonl.gz	13 files	complete episode records — every turn, parsed action, and belief block
bank/	24	the frozen instance bank, without which surplus columns cannot be recomputed

Key columns: experiment_name (append-safe grouping), cell_id, arm, deadline_rounds, host, instance_id, seed, deal_rate, normalized_primary (realized normalized utilitarian welfare over the instance's feasible ceiling; a no-deal scores 0), rounds_to_agreement, belief_auc_accept_set_f1.

Reproducing the headline belief-accuracy numbers

The belief-F1 figures quoted in research note 0061 (0.210 / 0.228 / 0.241 / 0.244 / 0.241 / 0.235 across all_rational deadlines [4, 8, 16, 32, 64, 128]) are cells.csv.belief_auc_accept_set_f1 — a per-episode time-average, then aggregated across episodes. Taking a flat mean of the accept_set_f1 column in belief_turns.csv.gz instead runs about 0.003 lower at every deadline, because that weights turns equally rather than episodes and so over-weights long episodes. Both are defensible summaries of the same data; use cells.csv to match the published table.

Read that column as a time-average, not as "how much this cell learned". The per-turn curve in belief_turns.csv.gz peaks at turn 5–6 at the same value (~0.252) in every cell from 32 rounds up, then drifts down for the rest of the horizon — at 128 rounds it falls monotonically across the whole game, 0.243 → 0.226 in eight blocks, ending 0.024 below its peak. A longer cell therefore averages in more of the decayed tail and posts a lower number (r128's 0.235 against r32's 0.244) while having learned exactly as much, exactly as fast. The decay is not a composition artifact: 119 of the 120 r128 episodes contribute all 129 turns.

Experiment-name mapping

experiment_name	description
pure-rounds-sweep-v1	the 13 complete cells below — all_oracle at [4, 8, 16, 32, 64, 128, 256], all_rational at [4, 8, 16, 32, 64, 128] — at 120 episodes each (24 frozen parameter sets × seeds 0–4), unanimity, private information. The grid is deliberately incomplete: all_rational at [256] rounds is absent (see Provenance).

cell	arm	deadline	episodes	host	deal rate	score
all_oracle_r4	all_oracle	4	120	tiger7	1.000	0.888
all_oracle_r8	all_oracle	8	120	tiger7	1.000	0.884
all_oracle_r16	all_oracle	16	120	tiger7	1.000	0.877
all_oracle_r32	all_oracle	32	120	tiger7	1.000	0.899
all_oracle_r64	all_oracle	64	120	tiger7	1.000	0.890
all_oracle_r128	all_oracle	128	120	tiger7	1.000	0.884
all_oracle_r256	all_oracle	256	120	tiger7	1.000	0.877
all_rational_r4	all_rational	4	120	jagupard29	0.233	0.195
all_rational_r8	all_rational	8	120	viscam1	0.267	0.230
all_rational_r16	all_rational	16	120	jagupard34	0.208	0.182
all_rational_r32	all_rational	32	120	jagupard34	0.208	0.178
all_rational_r64	all_rational	64	120	iliad1	0.250	0.207
all_rational_r128	all_rational	128	120	john13	0.242	0.203

Two caveats that change how these numbers read

1. all_oracle is a flat control for AGGREGATES ONLY. Closure and mean score are horizon-invariant, but per episode — same host, same instance and seed — it lands on materially different deals at different deadlines (91–97 of 120 differ, median |Δscore| 0.14). Do not run a paired within-episode oracle contrast across deadlines. Its small residual aggregate variation is periodic in (deadline + 1) mod 5, the identity of the rotating seat that holds the forced final.
2. Episode-level identity is host-dependent. Around 3–6% of episodes in the all_rational arm sit on near-ties whose resolution depends on the host CPU's float-summation order, usually flipping a whole deal/no-deal outcome. The host column is provided for that reason; aggregate comparisons are unaffected (the rounds=4 cells are statistically indistinguishable from the frozen campaign), but episode-level pairing across cells on different hosts is not clean. The all_oracle arm is unaffected.

Regenerating this

# create the grid sweep, fan it out across Slurm CPU nodes, then close any preemption holes
uv run python experiments/rational_agents/launch_rounds_sweep.py --create-only
experiments/rational_agents/submit_rounds_sweep.sh --sweep-id <entity>/<project>/<id> --jobs 5 --cpus 16 \
    -- --shards-per-cell 16 --shard-timeout-s 10800 --bootstrap-samples 2000
uv run python experiments/rational_agents/launch_rounds_sweep.py --sweep-id <id> --fill-missing
# then package and upload
uv run python experiments/rational_agents/package_hf_rounds_sweep.py --out /tmp/rounds_sweep_hf
uv run python experiments/rational_agents/upload_hf_publication.py --folder /tmp/rounds_sweep_hf --repo siddharthmb/2026.RA.Pure-Rounds-Sweep

Provenance

• W&B sweeps (a real method: grid sweep over arm × rounds, one run per cell, plus the fill passes — a killed sweep run never releases its grid point, so a refilled cell appears under its own sweep id):
  • https://wandb.ai/siddharth-stanford/rational_agents/sweeps/ufrdew0j — the original arm x rounds grid sweep — 12 cells
  • https://wandb.ai/siddharth-stanford/rational_agents/sweeps/l1f1larz — fill pass for all_rational at 128 rounds — the 13th cell
  • https://wandb.ai/siddharth-stanford/rational_agents/sweeps/txi1tje1 — fill pass for all_rational at 256 rounds — cancelled, produced no cell
• Cluster artifacts: /nlp/scr/siddharth/ii_mats/rational_agents/pure_rounds_sweep_v1/ — per-cell run directories in the ordinary schema, per-cell records under control/cells/, manifest at control/campaign_manifest.json (status partial-final-13)
• Slurm logs: /juice2/u/siddharth/ii_mats/logs/rounds_sweep/<jobid>.out, first line of each is the exact invocation
• Bank: experiments/rational_agents/instances_five_seat_private_v2, manifest SHA-256 cd32ebdef8072da3…
• Known gap, permanent: all_rational_r256 is absent and will not be run. Its cost is superlinear in the deadline for an engine reason, not a solver one: the runner re-serializes the entire growing episode record once per wave, so a single r256 record reaches ~0.9 GB (against ~0.18 GB at r128) and each shard worker holds 16–17 GB resident. Sixteen concurrent shards then exceed the job's 64 GB cgroup and are OOM-killed. The fill reached 64 of 120 episodes in 98 hours (0.65 episodes/h against r128's 7.86, a 12× slowdown) with 12 of 16 workers already dead, and was cancelled on 2026-08-24. Nothing rests on it: the horizon null is flat over five octaves (4 → 128) and belief accuracy has peaked and begun decaying by turn 6, so r256 would be confirmatory only. Re-running it needs the engine fix first, not more wall clock.

Related

• siddharthmb/2026.RA.Quorum-Rounds-Sweep — the agreement-rule axis: relaxing unanimity does move closure (0.217 → 0.908), which is the contrast that makes this null informative
• siddharthmb/2026.RA.Agent-Variants-Closure — the preference-sharing axis, the other lever that works
• siddharthmb/2026.RA.Five-Seat-Frontier-Negotiation — the frozen campaign these cells are anchored against