--- tags: - negotiation - multi-agent - game-theory - llm-agents language: - en --- # 2026.RA.Seeded-Optimal-Opening — does anyone sign the best deal when it is already on the table? Episode-level data for the **seeded-optimal-opening** experiment in the rational-agents programme: five-party scorable negotiation under private information, where a neutral **non-voting facilitator tables one complete package before any seat speaks**, and the only thing left for the table to do is judge it. Two seeded deals are crossed with two protocols: - **ceiling** — the feasible deal maximizing the sum of normalized surpluses `z_i = (u_i − τ_i) / c_i` (`c_i` = seat *i*'s best surplus over the deal space). This is exactly the deal the campaign's normalized score calls 1.0, so "the optimum was on the table" is true by construction, not by assertion. - **placebo** — a mediocre-but-signable deal drawn uniformly from the all-individually-rational set excluding the ceiling and the top decile by `Σ z`. The control that separates *recognizing* a good deal from *accepting whatever is standing*. - **V-vote** — one forced up/down vote on the tabled package by every seat, then done. - **V-continue** — the package stands while the ordinary rotating-proposer game runs to its deadline. Every episode is on the same frozen 24-game bank (`instances_five_seat_private_v1`, shipped in `bank/`) with seeds 0–4, private score sheets, datacenter framing, and unanimity (`min_accept` 5). Any LLM seat is `anthropic:claude-opus-5` with explicit thinking; the model-free lineups (`all_rational` = private-information Bayesian seats, `all_oracle` = omniscient best-response seats) contain no model text and cost nothing to reproduce. ## Headline Immediate acceptance = every seat's first move of the episode is an ACCEPT of the facilitator's package. Intervals are 95% cluster bootstraps over the 24 parameter sets with each set's five seeds resampled together. Rows with `seeded = none` are the frozen unseeded arms of the parent campaign, carried for reference. | arm | lineup | seeded | variant | episodes | immediate acceptance | deal rate | P(final == seeded) | normalized score | |---|---|---|---|---:|---:|---:|---:|---:| | `all_llm__ceiling_continue` | all_llm | ceiling | continue | 120 | 0.000 [0.000, 0.000] | 0.958 [0.908, 1.000] | 0.383 [0.242, 0.533] | 0.875 [0.814, 0.929] | | `all_llm__ceiling_vote` | all_llm | ceiling | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | | `all_llm__placebo_vote` | all_llm | placebo | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 0.678 [0.640, 0.713] | | `all_oracle__ceiling_continue` | all_oracle | ceiling | continue | 120 | 0.325 [0.217, 0.433] | 0.783 [0.700, 0.867] | 0.325 [0.217, 0.433] | 0.716 [0.638, 0.792] | | `all_oracle__ceiling_continue_ballot_fixed` | all_oracle | ceiling | continue | 120 | 0.325 [0.217, 0.433] | 1.000 [1.000, 1.000] | 0.450 [0.325, 0.575] | 0.924 [0.901, 0.946] | | `all_oracle__ceiling_vote` | all_oracle | ceiling | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | | `all_oracle__placebo_continue` | all_oracle | placebo | continue | 120 | 0.075 [0.033, 0.117] | 0.700 [0.567, 0.825] | 0.075 [0.033, 0.117] | 0.628 [0.506, 0.740] | | `all_oracle__placebo_continue_ballot_fixed` | all_oracle | placebo | continue | 120 | 0.075 [0.033, 0.117] | 1.000 [1.000, 1.000] | 0.167 [0.075, 0.275] | 0.877 [0.842, 0.907] | | `all_oracle__placebo_vote` | all_oracle | placebo | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 0.678 [0.640, 0.713] | | `all_rational__ceiling_continue` | all_rational | ceiling | continue | 120 | 0.000 [0.000, 0.000] | 0.233 [0.117, 0.367] | 0.025 [0.000, 0.075] | 0.177 [0.089, 0.279] | | `all_rational__ceiling_vote` | all_rational | ceiling | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | | `all_rational__placebo_continue` | all_rational | placebo | continue | 120 | 0.000 [0.000, 0.000] | 0.258 [0.142, 0.384] | 0.008 [0.000, 0.025] | 0.201 [0.115, 0.297] | | `all_rational__placebo_vote` | all_rational | placebo | vote | 120 | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 1.000 [1.000, 1.000] | 0.678 [0.640, 0.713] | | `all_llm__unseeded` | all_llm | none | continue | 120 | 0.000 [0.000, 0.000] | 0.958 [0.925, 0.992] | 0.000 [0.000, 0.000] | 0.873 [0.833, 0.911] | | `all_oracle__unseeded` | all_oracle | none | continue | 120 | 0.000 [0.000, 0.000] | 0.875 [0.817, 0.925] | 0.000 [0.000, 0.000] | 0.791 [0.735, 0.843] | | `all_oracle__unseeded_ballot_fixed` | all_oracle | none | continue | 120 | 0.000 [0.000, 0.000] | 1.000 [1.000, 1.000] | 0.000 [0.000, 0.000] | 0.907 [0.880, 0.933] | | `all_rational__unseeded` | all_rational | none | continue | 120 | 0.000 [0.000, 0.000] | 0.233 [0.133, 0.350] | 0.000 [0.000, 0.000] | 0.189 [0.110, 0.279] | | `one_oracle__unseeded` | one_oracle | none | continue | 120 | 0.000 [0.000, 0.000] | 0.508 [0.417, 0.600] | 0.000 [0.000, 0.000] | 0.461 [0.376, 0.545] | | `one_rational__unseeded` | one_rational | none | continue | 120 | 0.000 [0.000, 0.000] | 0.767 [0.683, 0.842] | 0.000 [0.000, 0.000] | 0.686 [0.611, 0.761] | Rows with `seeded = none` are the **frozen unseeded arms of the parent campaign**, carried so the seeded arms have a control on identical games. Their episodes are **not** shipped here — they live in the parent dataset linked at the bottom — and their acceptance columns are zero by construction because nothing was tabled. Full tables, paired contrasts, the fairness basket and the per-seat vote-versus-share curve are in `analysis/results.md` and `analysis/summary.json`. The narrative writeup is research note [0045](https://github.com/Sid-MB/ii_mats/blob/main/experiments/rational_agents/research-notes/0045-seeded-optimal-opening.md); the published hub is [here](https://rational-agents-runs.sidmb.com/seeded-optimal-opening/). ### Known defect in every arm carrying an omniscient seat — read before using their closure columns The omniscient seat casts its forced-final vote on whichever live offer it values most rather than on the one offer actually 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 an **abstention**. It affects only the forced final vote, so first-move acceptance is unaffected, but deal rate and everything conditional on closing are measured on an agent that sometimes fails to cast a vote it intended to cast. This defect **pre-dates this experiment and is present in the parent campaign** (94 of 107 forced-final turns in its `all_oracle` arm), which is why the arms here deliberately run the agent as published — changing it mid-experiment would have made the seeded arms incomparable with their own control. The per-turn evidence is in every episode record: look for `parsed_action.syntax_error` containing "The final vote is only on". The same defect is in the **`one_oracle__unseeded` reference row** shipped here, which is the frozen five-arm campaign's own `one_oracle` arm: 114 of that seat's 116 forced-final ballots were dropped (92 intended ACCEPTs, 18 REJECTs, 4 WALKs). Its 0.508 deal rate and 0.461 score in the headline table above are therefore a spoiled vintage. Both oracle arms have since been repaired and re-run on the identical games; the corrected figures are `all_oracle` 0.875 → 1.000 deal rate and `one_oracle` 0.508 → **0.917** (score 0.461 → 0.829, paired against all-LLM −0.412 → −0.044 [−0.104, +0.011], i.e. null). The repaired `all_oracle` cells are in this dataset as the `*_ballot_fixed` arms; the repaired `one_oracle` run belongs to the parent campaign and is published with it, in `siddharthmb/2026.RA.Five-Seat-Frontier-Negotiation`. Use the repaired figures for any claim about closure. ## What is in here | path | contents | |---|---| | `episodes/.jsonl` | one record per episode: the complete stored episode (every turn, its parsed action, the view the seat was conditioned on, the outcome) plus the cell's design coordinates | | `tables/episode_rows.csv` | one row per episode with every analysis column | | `tables/seat_votes.csv` | one row per seat decision, carrying that seat's own normalized share `z` of the tabled deal | | `analysis/` | `summary.json` (every estimate and interval), `results.md`, the figure | | `seeded_offer_manifest.json` | the frozen sidecar: which package each `(instance, seed)` was shown, and its geometry | | `bank/` | the frozen 24-instance bank, byte-identical to the one the campaign ran on | | `campaign_manifest.json` | the campaign's own record: cells, attempts, spend ledger, reproducibility fingerprint | Every table and every JSONL record carries an **`experiment-name`** column (`seeded-optimal-opening-v1`) so a later campaign with the same columns can be appended to this dataset rather than forking a new one. ## Cells | cell | lineup | seeded | variant | episodes | run | |---|---|---|---|---:|---| | `all_rational__ceiling_vote` | all_rational | ceiling | vote | 120 | `fiveseed_opus_v2__all_rational__ceiling_vote` | | `all_rational__ceiling_continue` | all_rational | ceiling | continue | 120 | `fiveseed_opus_v2__all_rational__ceiling_continue` | | `all_rational__placebo_vote` | all_rational | placebo | vote | 120 | `fiveseed_opus_v2__all_rational__placebo_vote` | | `all_rational__placebo_continue` | all_rational | placebo | continue | 120 | `fiveseed_opus_v2__all_rational__placebo_continue` | | `all_oracle__ceiling_vote` | all_oracle | ceiling | vote | 120 | `fiveseed_opus_v2__all_oracle__ceiling_vote` | | `all_oracle__ceiling_continue` | all_oracle | ceiling | continue | 120 | `fiveseed_opus_v2__all_oracle__ceiling_continue` | | `all_oracle__placebo_vote` | all_oracle | placebo | vote | 120 | `fiveseed_opus_v2__all_oracle__placebo_vote` | | `all_oracle__placebo_continue` | all_oracle | placebo | continue | 120 | `fiveseed_opus_v2__all_oracle__placebo_continue` | | `all_oracle__unseeded_ballot_fixed` | all_oracle | none | continue | 120 | `oraclefix_none_continue` | | `all_oracle__ceiling_continue_ballot_fixed` | all_oracle | ceiling | continue=all_oracle__unseeded_ballot_fixed | 120 | `oraclefix_ceiling_continue` | | `all_oracle__placebo_continue_ballot_fixed` | all_oracle | placebo | continue=all_oracle__unseeded_ballot_fixed | 120 | `oraclefix_placebo_continue` | | `all_llm__ceiling_vote` | all_llm | ceiling | vote | 120 | `fiveseed_opus_v3__all_llm__ceiling_vote_attempt03` | | `all_llm__placebo_vote` | all_llm | placebo | vote | 120 | `fiveseed_opus_v3__all_llm__placebo_vote_attempt03` | | `all_llm__ceiling_continue` | all_llm | ceiling | continue | 120 | `fiveseed_opus_v3__all_llm__ceiling_continue_attempt02` | ## Regenerating it ```bash # 1. the frozen seeded-offer sidecar (ceiling + per-seed placebo deals) uv run --no-sync python build_seeded_offer_manifest.py \ --bank instances_five_seat_private_v1 --out seeded_offer_manifest_v1.json # 2. the campaign (model-free cells first at $0, then the Opus cells under a hard cap) uv run --no-sync python launch_five_seat_campaign.py \ --bank instances_five_seat_private_v1 --out five_seat_seeded_offer_v3 --campaign-name fiveseed_opus_v2 \ --seeded-offer-only --seeded-offer-manifest seeded_offer_manifest_v1.json \ --budget-usd 350.0 --artifact-root $LARGE_ARTIFACTS_DIR --execute # 3. gate G3 — re-derive every computable seat's move offline and check the record uv run --no-sync python gate_seeded_offer_votes.py --run # 4. analysis uv run --no-sync python analyze_seeded_offer.py \ --campaign-manifest five_seat_seeded_offer_v3/campaign_manifest.json \ --artifact-root $LARGE_ARTIFACTS_DIR --bank instances_five_seat_private_v1 \ --reference-runs /runs --out # 5. this bundle uv run --no-sync python package_hf_seeded_offer.py --campaign-manifest ... --out ``` All commands run from `experiments/rational_agents/` in the project repository. ## Provenance - Preregistered design and gates: `proposals/2026-08-10-seeded-optimal-opening.md` (fixed before any code was written; every deviation is logged in the results note). - Campaign fingerprint: `a67629cb8118cf2a78b21e569971fcb5e853ba4ddafd0c5274dc90593b3c8c6d` — pins the bank hashes, the code hashes, the seeded-offer sidecar hash, and both repository revisions. - Seeds: [0, 1, 2, 3, 4]; 24 parameter sets; information `private`; framing `datacenter`. - Cluster paths for the source runs, transcripts and per-episode visualizations are recorded in `packaging_index.json` and in the campaign manifest. ## Related - Parent campaign dataset: — the same bank and seats with nothing tabled, which is this experiment's control.