siddharthmb/2026.RA.Five-Seat-Frontier-Negotiation

Five-Seat Frontier Negotiation

⚠ ERRATUM (2026-08-10) — the omniscient-oracle arms in this bundle carry a spoiled ballot

OmniscientBestResponsePolicy — the computable seat in every *_oracle lineup here — cast its forced-final vote on whichever live offer it valued most instead of on the one offer under the up/down vote. The protocol rejects that as a legality error, the seat spends its single retry repeating itself, and the turn is recorded as a pass: a silent abstention, because a pass parses cleanly. Fixed in commit ca20157 (2026-08-10), which postdates every episode in this bundle. Arms without that policy — all-LLM, the private Bayesian lineups, and the fairness/DP composed policies — are clean (0 mismatches over 9,589 re-derived turns).

What this means. The episodes are a faithful record of what happened, but the oracle seat in them is a defective agent, so any closure claim about an oracle lineup is an artifact rather than a negotiation result. Where the repair has been measured: on the frozen Opus campaign, all_oracle's deal rate goes 0.875 → 1.000 (all 15 of its no-deals were spoiled ballots) and its score 0.791 → 0.907; on a preregistered fresh-bank replication, one_oracle − all_llm on utilitarian score is −0.068 [−0.136, −0.001] rather than −0.412. Do not train on or evaluate against the oracle seat's forced-final vote turns without accounting for this.

Full account: research notes 0045 (the gate that found it), 0039 (replication verdict) and 0043 (fairness-basket erratum) in the source repo.

This is the public evidence bundle for a matched five-party datacenter negotiation experiment. Every LLM receives only its private preference sheet and whole-number acceptance threshold, can propose packages and send public persuasive messages, and uses explicit API thinking. Five primary conditions compare five LLMs, four LLMs plus one rotating private-information rational agent, four LLMs plus one rotating omniscient oracle, five rational agents, and five oracle agents. Rational-advisor, lower-quorum, and personality arms are explicitly deferred and absent from this release.

The generating model is anthropic:claude-opus-5. The frozen campaign name is five_seat_private_opus_v3, and this card was rendered in strict release mode. Raw episodes preserve provider-returned reasoning summaries/tokens when the API exposes them; this dataset does not claim access to hidden chain-of-thought. Each committed/scorable turn has both a private-information rational counterfactual and an omniscient oracle counterfactual. Unparseable emissions that commit no move (including their stored retry attempts) remain visible in the raw episode and transcript but correctly have no annotation turn of their own.

⚠ Erratum (2026-08-14) — both oracle arms' closure numbers in this release are a spoiled ballot

Read this before using one_oracle or all_oracle deal rates, no-deal counts, or scores. The omniscient seat (OmniscientBestResponsePolicy) cast its forced-final vote on whichever live offer it valued most rather than on the one offer under the up/down vote. The protocol rejects that as a legality error, the seat repeats itself on its single retry, and the turn is recorded as a pass — a silent abstention, because a pass parses cleanly. Reach, measured by re-deriving each seat's move offline from the exact view it was conditioned on: 94 of 107 of all_oracle's forced-final turns, touching every one of its 15 no-deal episodes, and 114 of the omniscient seat's 116 forced-final ballots in one_oracle (92 intended ACCEPTs, 18 REJECTs, 4 WALKs), touching 57 of its 59 no-deals. The all_llm, one_rational and all_rational arms are unaffected — their seats read the standing offer, and re-derivation finds no vote of theirs disturbed.

Both arms were repaired (commit ca20157) and re-run on the identical 120 games:

arm	deal rate	normalized score	paired score vs all_llm
all_oracle as published here	0.875	0.791	−0.082 [−0.146, −0.022]
all_oracle repaired	1.000	0.907	+0.034 [+0.001, +0.074]
one_oracle as published here	0.508	0.461	−0.412 [−0.495, −0.321]
one_oracle repaired	0.917	0.829	−0.044 [−0.104, +0.011] (null)

The repaired one_oracle run is published in this dataset under runs_ballot_repaired/one_oracle/; the repaired all_oracle runs are in the sibling dataset 2026.RA.Seeded-Optimal-Opening as its *_ballot_fixed arms. The arms as published here are kept unchanged as the record of that vintage.

This changes the campaign's headline shape. Its published reading — computable agents collapse, and worse the more of them there are — does not survive the repair. The corrected ordering is all_oracle > all_llm > one_oracle > one_rational > all_rational on both deal rate and score: only the private-information Bayesian agents collapse, the omniscient lineups close as often as the LLM table or more at both compositions, and the gradient these five arms trace is information, not computability. What does survive at every composition is that every computable arm splits worse than the all-LLM table — the repaired one_oracle's worst-off party is −0.060 [−0.092, −0.033] and its Gini +0.043 [+0.025, +0.065], now measured on 106 paired closed games instead of 56.

Full account: research notes 0045 (the gate that caught it) and 0043 (errata 1 and 2).

Experiment-name mapping

Every CSV and JSONL table uses experiment-name=five-seat-private-frontier-v1.

experiment-name	description
five-seat-private-frontier-v1	Five primary arms, each using 24 parameter sets × five seeds, with private sheets, whole-number thresholds, explicit thinking, rotating treated seats, dual per-turn counterfactuals, and datacenter framing.

The logical-table column separates the five conditions without fragmenting the matched experiment into incompatible tables.

Contents

• bank/: the frozen, hash-verified 24-parameter-set bank and its difficulty measures.
• campaign/: the campaign manifest and the deterministic selection of complete immutable attempts.
• runs/<logical-table>/: selected run manifest/log, exact saved instances, episodes, Markdown/HTML transcripts, dual annotations, and self-contained interactive visualizers. The visualizer hover card places the actual action, private-rational action, and omniscient-oracle action on the same frontier mini-graph.
• tables/parameter_sets.csv: append-friendly parameter-set index.
• tables/episode_index.csv: one row per episode with paths to all paired public artifacts.
• analysis/: analysis rows, machine-readable summary, narrative summary, and every plot declared in summary.json.
• writeup/: reproducible LaTeX source and compiled PDF.
• artifact_manifest.json: byte sizes and SHA-256 hashes for every other packaged file.

Exact regeneration commands

From the ii_mats repository root, regenerate the frozen bank:

uv run python experiments/rational_agents/generate_five_seat_bank.py \
  --out /path/to/five_seat_private_bank_v1 \
  --n-instances 24 \
  --first-seed 52000 \
  --n-issues 4 \
  --n-options 4 \
  --rounds 4

Render and execute the campaign on the designated CPU RunPod (the generated durable worker performs the API runs, annotation, and analysis):

uv run python experiments/rational_agents/launch_five_seat_campaign.py \
  --bank /path/to/five_seat_private_bank_v1 \
  --out /path/to/five_seat_private_frontier_v1 \
  --campaign-name five_seat_private_opus_v3 \
  --model anthropic:claude-opus-5 \
  --budget-usd 2700 \
  --episode-concurrency 8 \
  --cell-concurrency 1 \
  --api-turn-token-floor 16384 \
  --artifact-root /workspace/large_artifacts \
  --include-personas \
  --execute

Regenerate final analysis and package the public tree without uploading:

uv run python experiments/rational_agents/analyze_five_seat_campaign.py \
  --manifest /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_campaign_manifest.json \
  --out /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_analysis \
  --bootstrap-samples 10000 \
  --seed 20260804

uv run python experiments/rational_agents/package_hf_five_seat.py \
  --manifest /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_campaign_manifest.json \
  --analysis /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_analysis \
  --writeup /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_report \
  --out /path/to/2026.RA.Five-Seat-Frontier-Negotiation

After reviewing the packaged tree and its artifact_manifest.json, publish it as a public dataset (these commands are intentionally not run by the packager):

uv run hf repo create siddharthmb/2026.RA.Five-Seat-Frontier-Negotiation \
  --repo-type dataset \
  --exist-ok

uv run hf upload-large-folder \
  siddharthmb/2026.RA.Five-Seat-Frontier-Negotiation \
  /path/to/2026.RA.Five-Seat-Frontier-Negotiation \
  --repo-type dataset \
  --num-workers 8

For a fast fixture, append --smoke; it relaxes the 120-per-arm release size and permits missing final analysis/writeup while retaining episode/annotation/transcript/visualizer matching and all integrity checks.

Original artifact paths

• Frozen bank: /workspace/projects/mats-ii-five-seat-20260804/experiments/rational_agents/instances_five_seat_private_v1
• Campaign manifest: /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_campaign_manifest.json
• Persistent run root: /workspace/large_artifacts
• Analysis: /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_analysis
• Writeup: /workspace/large_artifacts/ii_mats/rational_agents/control_five_seat_private_opus_v3/primary_report

Selected immutable run directories:

• all_llm: /workspace/large_artifacts/ii_mats/rational_agents/five_seat_private_opus_v3__all_llm
• all_oracle: /workspace/large_artifacts/ii_mats/rational_agents/five_seat_private_opus_v3__all_oracle
• all_rational: /workspace/large_artifacts/ii_mats/rational_agents/five_seat_private_opus_v3__all_rational
• one_oracle: /workspace/large_artifacts/ii_mats/rational_agents/five_seat_private_opus_v3__one_oracle
• one_rational: /workspace/large_artifacts/ii_mats/rational_agents/five_seat_private_opus_v3__one_rational

No Weights & Biases runs are used by this API campaign.

Related public datasets

• 2026.RA.Negotiation-Campaigns contains the earlier rational-agent negotiation campaigns and controls.
• 2026.RA.Divergence-DPO-Pairs contains preference pairs derived from related negotiation divergences.
• Anthropic model documentation documents the hosted model family used to generate the LLM turns.

This repository is designed for later append-only additions: compatible future tables should retain their schema and use a new experiment-name value documented above.

Capability ladder — Sonnet 5 and Haiku 4.5 all-LLM tables (note 0060) — capability_ladder_v1/

240 new episodes: the five-seat all-LLM table re-run with five Sonnet 5 seats and with five Haiku 4.5 seats on the bank-v2 protocol, paired against the matched Opus-5 reference and the same-bank computable arms. Headline: the ladder's big step is Sonnet → Opus (paired deal rate −0.383), not Haiku → Sonnet (−0.075, unresolved), and fairness orders with capability. Stamped experiment-name = five-seat-capability-ladder-v1; bank v2, so never pooled with the bank-v1 arms at the root. See capability_ladder_v1/README.md.

Advised advocate (wave 1, arm A2) — five_seat_advised_v1/

120 episodes of rational_advised_llm: five Opus-5 seats, one of which is privately handed the exact action the Bayesian policy of the one_rational arm would take, and told to play it and add a persuasive public message. Same bank, seeds, framing and protocol as runs/, so it pairs against them episode for episode; it separates the published one_rational deficit into a channel component and a rule component. Stamped experiment-name = five-seat-advised-advocate-v1. See five_seat_advised_v1/README.md.

The stratified generality bank — model-free arms (note 0069) — generality_bank_v1/

360 episodes ($0, model-free) on a new 36-instance bank stratified by feasibility — 22 feasible, 8 knife-edge, 6 with no individually-rational package at all, the first five-seat bank on which a no-deal can be the correct answer. Two findings: the Bayesian arm never walks (0/180), collecting the correct-refusal score on impossible games by running the clock out; and the omniscient arm closes 30/30 below-threshold deals there, a localized agent defect (all action values tie at 0.0 and the alphabetical tie-break picks accept) shipped with its diagnosis rather than as a result. deal_rate inverts meaning between strata — never pool it. Stamped experiment-name = five-seat-generality-free-arms-v1. See generality_bank_v1/README.md.

Capability × intervention grid — one computable seat at Sonnet 5 and Haiku 4.5 (note 0067) — capability_grid_v1/

480 new episodes: the one_rational and one_oracle differential arms re-run at the two mid-capability points of note 0060's ladder, paired against their own all-LLM tables and the Opus bank-v2 references. Headline: the omniscient seat's paired Δ score flips sign with capability — +0.121 [-0.011, +0.263] at Sonnet and +0.176 [+0.077, +0.281] at Haiku, against −0.068 at Opus. Stamped experiment-name = five-seat-capability-grid-v1; bank v2, never pooled across capability points. See capability_grid_v1/README.md.

Advised advocate (wave 2, arm A5 — LLM-as-interpreter) — five_seat_advised_v1/wave2_interpreter/

120 episodes of interpreter_advised_llm, the closed-loop advised seat: it talks, it parses the other seats' public language into structured preference evidence (a separate Opus call that no score sheet can reach, with a verbatim-quote provenance guard), feeds it to a min-consensus planner, and is re-advised every turn. Wave 1 priced the OUTPUT channel; A5 - A2 prices the INPUT channel at matched output. Parser hallucination rate 0.00042, which is what makes the outcome interpretable either way. Same bank, seeds, framing and protocol as runs/, so it pairs episode for episode. Stamped experiment-name = five-seat-advised-interpreter-v2. See five_seat_advised_v1/wave2_interpreter/README.md.

Talking rational agent, Phase 1 — five_seat_talking_v1/

The mute Bayesian rational seat given a truthful templated voice (commit / narrate / hint / listen ladder + babble control), five-seat all-rational tables on the frozen v2 bank; decision rule unchanged (gate G1), every message re-derived offline from policy state (gate G3-msg). Headline T4 (talk + listen) verdict: "starvation not schedule-shaped (<=0.40)" at deal rate 0.3. Stamped experiment-name = five-seat-talking-rational-v1. See five_seat_talking_v1/README.md.

Advised advocate (wave 3, arm A6 — ALL FIVE seats interpreter-advised) — five_seat_advised_v1/wave3_all_interpreter/

120 episodes of all_interpreter_advised_llm, the all-parties version of wave 2's A5: every seat parses the public channel into verified preference evidence, feeds a min-consensus planner and is re-advised each turn, with its own parser, planner and claim history and nothing shared between seats. Wave 2 found decision support to be a public good with ONE donor and four free-riders; this arm asks what the equilibrium is when every seat is the donor. Per-seat override verdict: 5 donors, 0 holdouts. Parser hallucination rate 0.00056 with five simultaneous readers, which is what makes the outcome interpretable either way. Same bank, seeds, framing and protocol as runs/, so it pairs episode for episode. Stamped experiment-name = five-seat-all-interpreter-advised-v3. See five_seat_advised_v1/wave3_all_interpreter/README.md.

Omniscient-oracle tie-break guard — errata E1 re-scoring (note 0080) — oracle_tiebreak_guard_v1/

840 episodes: the all-algorithmic omniscient arm re-scored on three banks after a repaired tie-break, plus the same three cells re-run with the repair disabled on the same host and commit as the control. On a game with no individually-rational package every legal action tied at 0.0 and accept won alphabetically, so five omniscient seats signed a deal each was certain could not pass. Guarded: 30/30 below-threshold agreements and 130 IR violations become 30 correct no-deals and zero violations; on feasibility-screened banks 11 of 240 episodes also change, which the errata had predicted would be none. Stamped experiment-name = five-seat-oracle-tiebreak-guard-v1. See oracle_tiebreak_guard_v1/README.md.

Advised advocate — narrate-only CONTROL (rational_advised_llm_narrate_only) — five_seat_advised_v1/narrate_only_control/

120 episodes. Wave 1's A2 with a byte-identical prompt and the Bayesian advisor's move enforced by the harness, so the Opus seat contributes only its public message; its move sequence is one_rational's by construction. Splits A2's restoration into speech (narrate_only - one_rational) and move departures (A2 - narrate_only). The model would have overridden on 62 of 530 turns; every such turn is recorded in run/enforced_advice/. Stamped experiment-name = five-seat-advised-narrate-only-control-v1. See five_seat_advised_v1/narrate_only_control/README.md.