2026.RA.Fairness-Counterfactual-Pairs

Action-level contrastive pairs from five-party private-information negotiations: at every turn an LLM took, what a computable ideal agent would have done at that same decision point.

Each row is one (episode, turn, counterfactual_type). rejected_action is what the LLM actually did; chosen_action is the counterfactual agent's action. Both are structured actions ({atype, offer_id, deal}), never prose.

72,192 rows, 58,168 of them (81%) divergent — i.e. the counterfactual agent would have done something different. The agreeing rows are shipped too, because their base rate is what tells you how hard the decision was.

What counts as a row here — read this before comparing the number to any other pair count

A row is one (decision point, counterfactual type), not one training pair. Each decision point contributes up to one row per requested type, so the 72,192 rows above stand on 18,048 distinct (episode_id, turn_idx) decision points. Selecting a single counterfactual_type is what gives you one row per decision point.

This matters because a different convention is used elsewhere in this program: the p4_pairs tooling emits one regret-weighted DPO pair per best-response divergence turn, which is a smaller number computed over a narrower set of turns. The same cell can therefore have two entirely defensible pair counts — the Qwen3-32B companion cell measures 519 under p4_pairs and 10,952 here — and neither is wrong. They are not the same quantity, they will not agree, and a factor-of-twenty gap between two numbers both called "pairs" is the sort of thing that gets read as a discrepancy. Filter and count for yourself before quoting a size.

Episodes come from two policies, and which one matters enormously for what you may do with the rows:

• 41,132 rows from Claude Opus (anthropic:claude-opus-5, thinking on) — the original five-arm study.
• 10,952 rows from Qwen3-32B (Qwen/Qwen3-32B, thinking off) — added later, and on-policy for that model. See On-policy vs off-policy below.

Both sets negotiate five-party private-information deals over the same frozen 24-game bank, so they pair one-to-one on (instance_id, episode_seed). The study these come from found that at one seat out of five, a computable agent's objective — own-surplus versus table-welfare — decides what it does to the table, while the fairness seat leaves every welfare aggregate (Nash welfare, Gini, worst-off share) indistinguishable from an all-LLM table. The originally published size of that objective swap (+0.389 utilitarian score, +0.425 deal rate) is inflated and must not be quoted — see the erratum below. Sources: research note 0036 (with its erratum), 0039, and the ten-arm results hub.

⚠ ERRATUM (2026-08-10) — the one_oracle and all_oracle source arms carry a spoiled ballot

The omniscient best-response policy seated in the one_oracle and all_oracle arms 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 rejected that as a legality error, the seat spent its single retry repeating itself, and the turn was recorded as a pass — a silent abstention. It consumed 114 of 334 forced-final turns in one_oracle and 94 of 107 in all_oracle. Fixed in commit ca20157 (2026-08-10), which postdates every episode in this dataset.

What this means for these rows. The rows are unaffected as data — they are a faithful record of what each agent did — but two things follow. (1) Any headline campaign number involving those two arms is spoiled vintage: on a preregistered fresh-bank replication with the ballot repaired, one_oracle − all_llm on utilitarian score is −0.068 [−0.136, −0.001] rather than −0.412, all_oracle closes 1.000 rather than 0.875, the +0.389 objective-swap contrast is inflated by the same defect, and the "information amplifies motive" claim is withdrawn (repaired interaction +0.019 [−0.052, +0.091], contains zero). (2) The chosen_action for the oracle_omniscient counterfactual type on forced-final vote turns in the five_seat_private_v1__one_oracle and __all_oracle splits is produced by the same pre-fix policy, so treat those turns as suspect for training or evaluation; every other arm here (all_llm, one_rational, one_fairness_oracle, one_fairness_algorithmic, and the Qwen3-32B cell) is clean — 0 mismatches over 9,589 re-derived turns.

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

On-policy vs off-policy (read this before training)

This is the single most important distinction in the dataset, and it splits the rows in two.

The Opus rows are OFF-POLICY for every model except Claude Opus. rejected_action is what Opus did. Training another model against them teaches it to move away from behaviour it never produces, while saying nothing about the failure modes it actually has. Methods that assume on-policy negatives — and any implicit-reward reading of the pair — are mis-specified on these rows. The rows are state-complete, so the decision point can be re-rendered and re-sampled on-policy for whatever model you actually care about.

The 2026.RA.Qwen3-32B-OnPolicy-FiveSeat rows are ON-POLICY for Qwen3-32B. They exist specifically to close that gap: rejected_action on those rows is Qwen3-32B's own verbatim emission at that decision point, generated on the same frozen bank under the same protocol. If you are doing preference learning on Qwen3-32B, filter to this experiment-name rather than training on the corpus as a whole.

Do not mix the two arms without labelling them. A mixed corpus confounds "learned from its own mistakes" with "learned to imitate Opus", which are different experiments with different conclusions.

One deliberate protocol difference: the Qwen arm runs thinking off, matching the GRPO rollout configuration it is meant to be on-policy for, whereas the Opus arm runs thinking on. Private reasoning is not removed either way — the scaffold's scratchpad field carries it.

Protocol versions — do not pool across them, and protocol_version alone is not enough

Matching protocol_version is necessary and not sufficient: match turn_token_budget too. Filtering to one protocol version feels like it yields a poolable subset and does not — inside the frozen version alone, rows span an 8× per-turn budget difference (16,384 against 2,048), which is the very failure this section exists to prevent. Both columns, every time.

And do not read the unresolved contrast below as "the versions agree, so pooling is harmless". That null is one endpoint at bounded precision. On the quantities feeding most other analyses the two versions differ resolvedly: for the thinking-on arm, placeholder turns 0.2444 → 0.0000, at-cap 0.2713 → 0.0000, non-action 0.4103 → 0.2245, and output tokens per turn +1,442 [+1,348, +1,525] (notes 0048, 0041). The mechanism that makes this bite: non-action rows are exactly what a divergence statistic excludes, and the two versions differ by ~18 points in how many of them there are — so in a pooled corpus the exclusion rate becomes a function of the version mix, and any drop-rows statistic silently depends on how many rows came from each. That is not a claim that an exclusion biased some estimate; it is that pooling makes the estimand depend on the mix.

protocol_version records the turn-budget protocol the episode ran under. five-seat-private-v1 is the frozen protocol: the scenario's own per-phase caps, 2,048 tokens for an action turn and 2,560 for a forced-final vote, with no override. A five-seat-uncapped-turn<N>-v1 value means the cell ran --turn-max-tokens <N>.

protocol_version	rows	per-turn budgets (turn_token_budget)	experiments
five-seat-private-v1	52,084	2,048, 2,560, 16,384	five_seat_private_v1__all_llm, five_seat_private_v1__one_rational, five_seat_private_v1__one_oracle, fivepriv_opus_fairness_v2__one_fairness_oracle, fivepriv_opus_fairness_v2__one_fairness_algorithmic, 2026.RA.Qwen3-32B-OnPolicy-FiveSeat
five-seat-uncapped-turn32768-v1	20,108	32,768	2026.RA.Qwen3-32B-Uncapped-ThinkOn, 2026.RA.Qwen3-32B-Uncapped-ThinkOff

Rows from different protocol versions do not pool. A raised cap changes how much of each turn survives, so a contrast computed across versions differences a cap change and reports it as a model, objective or information effect. That is definitional, not an empirical finding, and it is the whole of the rule: a turn truncated at 2,048 tokens and the same turn allowed 32,768 are not the same measurement, so the arms are not comparable however the numbers come out. The concrete, uncontested case is in turn_token_budget below.

Two figures that were once cited here are withdrawn — quote neither. Note 0048 reported the thinking-on-minus-thinking-off divergence contrast as −0.029 [−0.080, +0.023] under the frozen caps and +0.049 [+0.011, +0.090] uncapped, and read the pair as the cap reversing the sign. Note 0052 re-measured the uncapped side at 24 instances × 5 seeds: +0.005 [−0.016, +0.025], not resolving, with 0048's point estimate outside the powered interval. The two original estimates were one-seed-per-cluster excursions on opposite sides of a powered near-zero — on this bank seed variance dominates (ICC 0.255) — and they do not disagree with each other either. So do not quote +0.049, and do not quote the generalisation built on it: the frozen protocol did exclude non-action rows 2.11× more often in one arm than the other (note 0041) and reporting per-arm exclusion rates remains cheap and worth doing, but there is no evidence in this programme that the asymmetry biased any estimate, and the worked example that purported to show it was the noise it claimed to diagnose. This corpus contains more than one protocol version. Filter to one before differencing anything.

turn_token_budget is a separate field because the version does not determine it. It is the budget that actually applied to that turn: the scenario cap on the turn, raised by the hosted-API turn_token_floor where one was set. Inside the single frozen version the Opus arms therefore ran at an effective 16,384 while the open-weight arm ran at 2,048/2,560 — so a frozen-protocol Opus arm and a frozen-protocol local arm are still not cap-matched, which is invisible in the version label and is the correction note 0048 records. Check this column, not just the version, before pairing arms.

Read this before training on it

• Per-turn imitation of the self-interested oracle ANTI-TRACKS outcomes in this program. Research notes 0004 and 0016 measured agents that matched the omniscient best-response oracle more closely at the turn level while doing worse on episode outcomes. oracle_omniscient is included for measurement, not because it is a good target.
• information_leaky = true rows cannot be honestly imitated. oracle_omniscient and fairness_oracle read every party's hidden score sheet. The acting model's prompt does not contain those sheets, so a model trained to reproduce these actions is being trained to act on information it does not have — it will learn to guess confidently. Filter on information_leaky unless you specifically want that.
• fairness_private is the novel target and it is UNTESTED for training. It sees exactly what the LLM saw (own sheet + public moves) and optimizes the table's normalized Nash welfare instead of its own surplus. No result in this program yet shows that training on it helps. It is offered as a hypothesis with a matched control (rational_private, same architecture and information, self-interested objective), which is the comparison worth running.
• No rationales, by construction. The P4 lane measured that 95.5% of the naive per-turn signal here was style rather than substance. There is no field in this schema for synthesized reasoning text; adding one would reintroduce exactly that contamination.

Counterfactual types

type	rows	information-leaky
fairness_oracle	13021	yes
fairness_private	13021	no
oracle_omniscient	13021	yes
rational_private	13021	no

rational_private and fairness_private differ only in objective; *_oracle variants differ from their private counterparts only in information. That 2x2 is deliberate.

Fields

Rows are state-complete: issues, own_sheet, offers, standing_offer_id, round, min_accept, deadline and n_parties are enough to re-render the decision point into any model's prompt format. The original rendered prompt is referenced by view_ref and embedded only when built with --include-views, so the dataset is not implicitly tied to one model's formatting.

counterfactual_value is the deal's true full-information table-welfare objective and table_optimum is the best any deal could score, so table_optimum - counterfactual_value is a regret in fairness units. A fairness_private row is decided under its posterior but priced on the truth — that gap is a direct measurement of how badly private-information belief recovery mistargets.

protocol_version and turn_token_budget say which turn-budget protocol produced the row and what budget actually applied to that turn. Both are filtering keys, not metadata — see Protocol versions — and both must match before rows pool.

diverges marks whether the counterfactual action differs from the model's. Non-divergent rows are kept deliberately: their base rate is the denominator for any divergence-rate claim, and dropping them silently turns every such rate into 100%.

Experiments

Every row carries experiment-name, so a later campaign appends to this repo rather than forking it. An all-policy arm (every seat computable) contributes no rows by design — there is no LLM action to contrast — so it is absent from this table even though it is part of the ten-arm study.

experiment-name	rows	policy	what the table was
five_seat_private_v1__all_llm	8660	Opus	five Claude Opus seats — the campaign's reference table, no computable agent present
five_seat_private_v1__one_oracle	8536	Opus	four Opus seats + one omniscient self-interested seat (enumerate-and-score)
five_seat_private_v1__one_rational	8512	Opus	four Opus seats + one private-information self-interested Bayesian seat
fivepriv_opus_fairness_v2__one_fairness_algorithmic	7984	Opus	four Opus seats + one private-information table-welfare (fairness) seat
fivepriv_opus_fairness_v2__one_fairness_oracle	7440	Opus	four Opus seats + one omniscient table-welfare (fairness) seat
2026.RA.Qwen3-32B-OnPolicy-FiveSeat	10952	Qwen3-32B	five Qwen3-32B seats (thinking off) on the same frozen bank — on-policy for Qwen3-32B, the counterpart to five_seat_private_v1__all_llm
2026.RA.Qwen3-32B-Uncapped-ThinkOff	10896	Qwen3-32B	five Qwen3-32B seats, thinking off, same frozen bank but run uncapped (--turn-max-tokens 32768) — the placebo arm of research note 0052; different protocol_version, does not pool with the frozen-cap arms
2026.RA.Qwen3-32B-Uncapped-ThinkOn	9212	Qwen3-32B	five Qwen3-32B seats, thinking on, uncapped (--turn-max-tokens 32768) — viable only uncapped (0 placeholder turns of 2,303, against 24.4% at the frozen 2,048 cap); different protocol_version, does not pool with the frozen-cap arms

About the Qwen3-32B arm

120 episodes (24 bank instances × seeds 0–4), 2,738 turns, run on one rented H100 for ≈$11.40. Rollout gates: fabrication 0 of 2,738 attempted turns, 120/120 episodes done, 95.80% parse_ok, zero failed or empty-content turns, 120 unique (instance_id, seed) keys.

Sizing caveat, stated plainly. Under the p4_pairs definition this arm is 519 pairs — roughly 1.2% of the Opus rows. It is the correct on-policy corpus and it is small. Realistic uses are an evaluation set, or a DPO arm with heavy regularization; it will not support a large training run on its own. The cheap lever for more volume is more seeds on the same frozen bank (~6.1 min per 6 episodes on one H100), which preserves one-to-one pairing with the Opus arm.

Do not read the per-arm divergence rates as a model-quality comparison. This arm is 83.1% divergent against the Opus arms' 77.7–78.8%, and that gap is confounded: the Qwen arm runs thinking off (to match the GRPO rollout configuration it is on-policy for) while the Opus arms run thinking on. Model and thinking mode vary together across that comparison, so nothing in it separates "Qwen diverges from the counterfactual more than Opus does" from "a thinking-off policy diverges more than a thinking-on one." Within-arm comparisons and per-counterfactual-type contrasts are unaffected; only the cross-policy read is compromised. Disentangling it needs a labelled thinking-on Qwen chunk, which has not been run.

A measurement caveat worth inheriting. The bestresponse divergence rate over the full 120 episodes is 23.2% (519 of 2,238 scored turns — a narrower denominator than the row-level 83.1% above, which counts all four counterfactual types). An early six-episode sub-chunk measured 36% and was extrapolated to a projected "~800 pairs" — 35% above the truth. Two further six-episode samples read 25.2% and 23.2%, so the per-chunk rate swings across roughly 23–36% on ~110 scored turns. Sub-chunk previews are for gating a pipeline, never for sizing a dataset.

The stronger form of that lesson is not about variance. Two behavioural flags (ir_violation, below_threshold_propose) occur only at full-cell scale and appear in no six-episode chunk. Variance is at least quantifiable — you can put an interval around a rate estimated from 110 turns and see that it is wide — but a preview is structurally blind to events rare enough not to occur in it, so it can be perfectly unbiased about the rates it does observe and still misrepresent the phenomenon.

Source artifacts on the cluster

• Extension run dirs (episodes, transcripts, annotations_fairness/, per-episode pages): /juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__<arm>/
• The five published Opus arms, byte-frozen and never written to: /nlp/scr/siddharth/ii_mats/rational_agents/five_seat_private_hf_snapshot_v1/runs/<arm>/
• Their fairness re-annotation, in a symlinked working copy so the snapshot stays frozen: /nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/<arm>/annotations_fairness/
• Qwen3-32B cell (merged, 120 episodes) and its annotations: /nlp/scr/siddharth/ii_mats/rational_agents/five_seat_qwen3_32b_all_llm/ (+ annotations_fairness/, quality_report.json)
• Qwen3-32B per-sub-chunk originals (untouched execution record, 20 dirs, transcripts per chunk): /nlp/scr/siddharth/ii_mats/rational_agents/five_seat_qwen3_32b_all_llm_b{0..3}s{0..4}/
• Qwen3-32B rollout and streaming-pull logs (first line is the exact invocation): /nlp/scr/siddharth/ii_mats/rational_agents/_qwen32b_rerun_logs/
• Campaign control dir, manifest and per-cell launcher logs: experiments/rational_agents/five_seat_fairness_control_v2/
• Ten-arm analysis output (summary.json, analysis.md, episode_rows.csv): .claude/products/five-seat-opus-ten-arm-complete/analysis/

No Weights & Biases runs are associated with this dataset: annotation and packaging are CPU-only and the rollouts are inference, not training.

Regenerating

Annotate with the fairness counterfactuals (CPU only, no API cost), then package. --include-fairness is required — without it the annotator emits the dual-counterfactual schema, which the packager rejects.

uv run --no-sync python annotate_five_seat_campaign.py \
  --episodes <run>/episodes --instances <run>/instances \
  --out <run>/annotations_fairness --include-fairness

uv run --no-sync python package_hf_counterfactual_pairs.py \
  --run five_seat_private_v1__all_llm=/nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/all_llm \
  --run five_seat_private_v1__one_rational=/nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/one_rational \
  --run five_seat_private_v1__one_oracle=/nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/one_oracle \
  --run five_seat_private_v1__all_rational=/nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/all_rational \
  --run five_seat_private_v1__all_oracle=/nlp/scr/siddharth/ii_mats/rational_agents/fairness_pairs_workdir/all_oracle \
  --run fivepriv_opus_fairness_v2__one_fairness_oracle=/juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__one_fairness_oracle \
  --run fivepriv_opus_fairness_v2__one_fairness_algorithmic=/juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__one_fairness_algorithmic \
  --run fivepriv_opus_fairness_v2__all_fairness_oracle=/juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__all_fairness_oracle \
  --run fivepriv_opus_fairness_v2__all_fairness_algorithmic=/juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__all_fairness_algorithmic \
  --run fivepriv_opus_fairness_v2__all_selfish_dp_oracle=/juice2/scr2/siddharth/ii_mats/rational_agents/fivepriv_opus_fairness_v2__all_selfish_dp_oracle \
  --run 2026.RA.Qwen3-32B-OnPolicy-FiveSeat=/nlp/scr/siddharth/ii_mats/rational_agents/five_seat_qwen3_32b_all_llm \
  --annotations-dirname annotations_fairness --out 2026.RA.Fairness-Counterfactual-Pairs

Regenerating the Qwen3-32B episodes themselves (≈2.6 GPU-hours on one 80 GB H100, ≈$11):

# --instance-batch 6 is the GPU-memory dial: the co-stepping wave width IS the memory knob.
# 6 fits 74.5 GB of 81 GB on an H100; use 12 on ~96 GB, 24 on 141 GB.
uv run --no-sync python run_five_seat_local_cell.py run \
  --run-name five_seat_qwen3_32b_all_llm --instances instances_five_seat_private_v1 \
  --model Qwen/Qwen3-32B --seeds 0 1 2 3 4 --instance-batch 6 --model-thinking off

uv run --no-sync python run_five_seat_local_cell.py merge \
  --chunks /nlp/scr/siddharth/ii_mats/rational_agents/five_seat_qwen3_32b_all_llm_b?s? \
  --out    /nlp/scr/siddharth/ii_mats/rational_agents/five_seat_qwen3_32b_all_llm

Related

• Campaign and arms: research note 0036 (fairness extension), 0030 (five-seat prereg).
• Qwen3-32B on-policy arm and the RL sizing that motivated it: research note 0037.
• Training-signal warnings: notes 0003, 0004, 0016 (per-turn oracle imitation), 0023 and 0028 (fairness-GRPO).
• Source episodes, transcripts and interactive per-episode pages live in the five-seat frontier bundle.
• Models that generated these rows: Qwen/Qwen3-32B (on-policy arm); Claude Opus (anthropic:claude-opus-5) via API for the five Opus arms.