siddharthmb/2026.RA.Five-Seat-Majority-Quorum

Five-Seat 4-of-5 Majority Quorum Arms

This is the public evidence bundle for two arms run under a 4-of-5 majority agreement rule instead of the frozen campaign's unanimity: all_llm_quorum_4 (five anthropic:claude-opus-5 seats) and one_rational_quorum_4 (four anthropic:claude-opus-5 seats plus one computable private-information Bayesian-rational seat). Each seat sees only its own private preference sheet and its own whole-number walk-away threshold, exchanges public messages, proposes complete packages, and votes. Seats use explicit API thinking.

Why the arms exist. They are the protocol control for the central result of the five-arm campaign in 2026.RA.Five-Seat-Frontier-Negotiation: replacing one of five frontier seats with a computable rational agent costs the table normalized score, almost entirely through closure. Every episode of that campaign ran under unanimity, which is the rule that gives the single rational seat a veto. So the published effect cannot distinguish two readings — the seat's lone REJECT blocks a deal the other four would sign, or the seat changes what the other four do. At 4-of-5 its rejection no longer blocks, and the two readings come apart. The estimand is a difference in differences against the frozen unanimity twins.

Designed to pair, not to stand alone. Both arms run the same 24 parameter sets, the same seeds 0–4, the same model, thinking mode, datacenter framing, canonical scaffold, deadline and seed-offset rotating proposer as the frozen campaign. The only thing that differs is min_accept. Each of the 240 episodes therefore joins a frozen five-arm episode exactly on (instance_id, episode_seed).

The bank is a derived copy, and that is deliberate. The vote threshold is a k-of-n field of the game, so these arms run on a re-materialized bank (five-seat-private-quorum-bank-v1) whose parameter sets, sheets, thresholds and instance_ids are identical to the frozen bank's and whose min_accept is 4. Each derived instance records the SHA-256 of the frozen instance it came from, so the correspondence is checkable rather than asserted. Instance hashes therefore differ from the sibling dataset's by construction; the parameter sets do not.

A guarantee that unanimity was providing for free does not survive this change. The frozen campaign reports a below-threshold accept rate of exactly 0.000 across its closed deals. That holds because of the rule, not because of the agents: a party below its walk-away threshold refuses, and under unanimity its refusal blocks. Under any k-of-n with k < n, four seats can close over the fifth's objection and nothing prevents the resulting deal leaving that seat below its threshold. Anyone reusing these episodes should treat individual rationality as a measured quantity here, not an invariant, and the analysis tables carry the outvoted seat's own outcome for exactly that reason.

The generating model is anthropic:claude-opus-5. The campaign name is fivemaj_opus_v1, 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 carries both a private-information rational counterfactual and an omniscient oracle counterfactual. Unparseable emissions that commit no move remain visible in the raw episode and transcript but have no annotation turn of their own.

Results are not restated here, so this card cannot go stale against the analysis. The paired tables, their cluster-bootstrap intervals and the verdict live in research note 0049-majority-quorum-arms.md (executing note 0047 arm C) in the ii_mats repository and on the published results hub linked from it. analysis/ in this tree carries these arms' own machine-readable numbers.

Experiment-name mapping

Every CSV and JSONL table uses experiment-name=five-seat-majority-quorum-v1.

experiment-name	description
five-seat-majority-quorum-v1	Two arms, all_llm_quorum_4 and one_rational_quorum_4: 4-of-5 majority closure, 24 parameter sets × five seeds each, private sheets, whole-number thresholds, explicit thinking, datacenter framing, dual per-turn counterfactuals.
five-seat-private-frontier-v1	The unanimity arms these pair against. Not in this repository — see 2026.RA.Five-Seat-Frontier-Negotiation.

The logical-table column names the condition, exactly as in the sibling dataset, so rows from both repositories concatenate without a schema change.

Contents

• bank/: the derived 4-of-5 bank and its difficulty measures — same parameter sets as the sibling dataset's bank, min_accept rewritten, each instance carrying its source hash.
• campaign/: the campaign manifest and the deterministic selection of complete immutable attempts.
• runs/all_llm_quorum_4/ and runs/one_rational_quorum_4/: selected run manifest/log, exact saved instances, episodes, Markdown/HTML transcripts, dual annotations, and self-contained interactive visualizers.
• 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 summaries and narrative summaries, including the difference-in-differences against the frozen unanimity arms and the outvoted-seat endpoints. The within-campaign paired-effects forest plot is absent by construction: this campaign contains no all_llm reference of its own, and its contrasts are computed against the frozen five-arm campaign instead.
• writeup/: reproducible LaTeX source and compiled PDF for the wider program.
• artifact_manifest.json: byte sizes and SHA-256 hashes for every other packaged file.

This is a separate repository rather than an append to the sibling dataset on purpose. A Hugging Face folder upload is a read-modify-write with no compare-and-swap, and an earlier append in this program silently clobbered another lane's concurrent write. A new repository plus cross-links carries the same discoverability with none of that risk.

Exact regeneration commands

From the ii_mats repository root. The source bank is the frozen one already published with the sibling dataset; the 4-of-5 copy is derived from it automatically by the runner, so there is nothing extra to generate.

uv run python experiments/rational_agents/launch_five_seat_campaign.py \
  --bank experiments/rational_agents/instances_five_seat_private_v1 \
  --out experiments/rational_agents/five_seat_majority_v1 \
  --campaign-name fivemaj_opus_v1 \
  --model anthropic:claude-opus-5 \
  --majority-only \
  --budget-usd 450 \
  --episode-concurrency 4 \
  --cell-concurrency 1 \
  --api-turn-token-floor 16384 \
  --worst-case-usd-per-request 6.0 \
  --artifact-root "$LARGE_ARTIFACTS_DIR" \
  --execute --resume

The two analyses reported on the hub, against the frozen unanimity runs:

uv run python experiments/rational_agents/analyze_quorum_did.py \
  --runs /path/to/combined_runs \
  --bank experiments/rational_agents/instances_five_seat_private_v1 \
  --out /path/to/analysis

uv run python experiments/rational_agents/analyze_quorum_outvoted.py \
  --runs /path/to/combined_runs \
  --arms all_llm_quorum_4 one_rational_quorum_4 \
  --out /path/to/analysis

--runs is a directory holding one directory per arm, named by the arm: the two arms of this dataset beside the frozen all_llm and one_rational runs from the sibling dataset. Symlinks are fine.

Original artifact paths

• Frozen source bank: /juice2/u/siddharth/ii_mats/experiments/rational_agents/instances_five_seat_private_v1
• Campaign manifest: /juice2/u/siddharth/ii_mats/experiments/rational_agents/five_seat_majority_v1/campaign_manifest.json
• Persistent run root: /juice2/scr2/siddharth
• Analysis: /juice2/scr2/siddharth/ii_mats/rational_agents/majority_v1_release_analysis
• Writeup: /juice2/u/siddharth/ii_mats/experiments/rational_agents/writeup/majority_quorum

Selected immutable run directories:

• all_llm_quorum_4: /juice2/scr2/siddharth/ii_mats/rational_agents/fivemaj_opus_v1__all_llm_quorum_4
• one_rational_quorum_4: /juice2/scr2/siddharth/ii_mats/rational_agents/fivemaj_opus_v1__one_rational_quorum_4

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

Related public datasets

• 2026.RA.Five-Seat-Frontier-Negotiation — the unanimity arms these pair against. Same parameter sets, same seeds, same model; 5-of-5 closure.
• 2026.RA.Five-Seat-MovesOnly-Control — the sibling protocol control that varies who may talk rather than how many must agree, against the same frozen campaign.
• 2026.RA.Negotiation-Campaigns — earlier rational-agent negotiation campaigns and controls.
• Anthropic model documentation — 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.