2026.RA.Agent-Variants-Closure

Episode and turn records from a design loop that took a five-seat private-information negotiation agent from a 0.233 deal rate to 1.000, and the controls that say why.

What this is

Five automated negotiators must agree unanimously on one package out of 256 within four rounds. Each holds a private score sheet (what each package is worth to it) and a private threshold (the value below which it prefers no deal). The published agent infers its counterparts from their offers and closes 0.233 of games; an omniscient reference shown every sheet closes 1.000 of the same games, so every failure is a missed feasible deal rather than an impossible one.

A forensic audit found the cause: the agent's belief learns only from an opponent conceding, which happens under one time per seat per episode, so its posterior stays near-uniform, it never believes any package can pass, and it proposes selfishly and never converges. This dataset holds the designs built to fix that, with the two that graduated confirmed on the full 24-instance bank.

Headline

arm	deal rate	normalized score	USW	min-seat surplus	IR violations
all_oracle (ceiling)	1.000 [1.000, 1.000]	0.888 [0.868, 0.908]	193.4	8.40	0.000
declaring (preference sharing)	1.000 [1.000, 1.000]	0.857 [0.823, 0.890]	186.3	7.70	0.000
conceding_novotes (best without sharing)	0.408 [0.283, 0.542]	0.339 [0.229, 0.453]	74.8	3.55	0.000
all_rational (published baseline)	0.233 [0.142, 0.333]	0.195 [0.120, 0.280]	45.4	3.04	0.000

120 episodes per arm (24 instances x 5 seeds), 95% intervals from an instance-clustered bootstrap with each instance's five seeds resampled together. Zero IR violations anywhere: nothing here buys closure by signing a package that leaves a seat below its own walk-away value.

Letting the agents state their preferences closes every game. Every design that keeps them guessing tops out near 0.4. The inference problem, not the bargaining problem, was the blocker.

Caveat that belongs next to that number

Truthful revelation is a design choice for cooperative-by-construction agents, not an equilibrium claim. The declaring seats emit numbers read straight off their own sheets and the readers trust them. Nothing here shows a self-interested agent would reveal truthfully, or that revelation survives an incentive to misreport — a seat overstating its threshold would be handed a larger slice by these readers. Robustness to strategic misreporting is future work.

Tables

Two tables, each its own config with a single train split, because their schemas differ:

from datasets import load_dataset
REPO = "siddharthmb/2026.RA.Agent-Variants-Closure"
eps   = load_dataset(REPO, "episodes")["train"]
turns = load_dataset(REPO, "turns")["train"]

episodes — one row per episode: experiment_name, variant, instance_id, seed, rounds, deal, normalized_primary, usw, min_seat_surplus, ir_violation, closing_round, n_turns, per_party_surplus (JSON), finalized_by, episode_id.

turns — one row per turn: the same identifiers plus round, seat, phase, action, deal_named (JSON), offer_id, message, content. The preference declarations are in message and content, so the sharing variant is auditable from this table alone.

experiment_name mapping

• confirm_declaring_fullbank_r4 (120 episodes) — PREFERENCE SHARING, full-bank confirmation. Five Bayesian seats each publish their TRUE score sheet once on the public message channel; every seat parses the declarations out of its view and, once all five are in, plans on a full-information table. 24 instances x 5 seeds, 4 rounds, unanimity.
• confirm_conceding_novotes_fullbank_r4 (120 episodes) — BEST NON-REVELATION design, full-bank confirmation. Deadline-indexed concession: demand at least (1 - t/T) of your own attainable surplus and, subject to that floor, table whatever is best for the believed worst-off seat. No preference sharing, no extra signal. Same 120-episode grid.
• subset_declaring_r4 (50 episodes) — The design-iteration bench run for the winning variant: the same preference-sharing agent on the fixed 10-instance subset (two instances per difficulty level L0-L4) x 5 seeds, 4 rounds.
• reference_all_rational_r4 (120 episodes) — The PUBLISHED private-information Bayesian agent, unmodified — the baseline every variant is measured against. Same bank, same seeds, same four-round protocol (the rounds-sweep r4 cell, which is anchor-certified identical to the frozen campaign).
• reference_all_oracle_r4 (120 episodes) — The OMNISCIENT reference: the same decision machinery shown every seat's sheet. Closes 1.000, which is what makes 1.000 the target rather than an aspiration.

Regenerate

cd experiments/rational_agents
# design loop on the fixed 10-instance subset (fast; subset frozen in agent_variants/subset.json)
uv run python -m agent_variants.bench --variant published oracle declaring conceding_novotes
uv run python -m agent_variants.bench --variant declaring --transcripts
# full-bank confirmation, 120 episodes per graduate
uv run python -m agent_variants.run_confirmation --variant declaring
uv run python -m agent_variants.run_confirmation --variant conceding_novotes
# this dataset
uv run python -m agent_variants.package_hf_variants --push

Cluster paths and runs

• Confirmation run dirs (episodes, instances, manifest.json with an agent_variant provenance block, HTML transcripts): /nlp/scr/siddharth/ii_mats/rational_agents/variantconfirm_declaring/ and /nlp/scr/siddharth/ii_mats/rational_agents/variantconfirm_conceding_novotes/
• Reference cells: /nlp/scr/siddharth/ii_mats/rational_agents/pure_rounds_sweep_v1/all_rational_r4/ and .../all_oracle_r4/
• Subset bench transcripts: /nlp/scr/siddharth/ii_mats/transcripts/rational_agent_variants/
• Weights & Biases: project rational_agents, group agent_variants_confirm_v1, runs variantconfirm_declaring and variantconfirm_conceding_novotes
• Code: experiments/rational_agents/agent_variants/ (variants.py, bench.py, run_confirmation.py, RESULTS.md leaderboard); the audit that motivated it is research note 0057 and this loop is written up in note 0059

Provenance note

Every variant is a subclass of the frozen campaign code — no negotiation solver, policy or scenario file was edited to produce this data, because an unrelated experiment was executing from the same working tree throughout. The published baseline in the table is the frozen agent, unmodified.