2026.RA.Fairness-GRPO-v2 — the complete λ-frontier of a fairness-trained LLM negotiator

What this is. The full evaluation record of the fairness-GRPO v2 campaign (experiments/rational_agents/ in the ii_mats repo): GRPO training of Qwen3-8B (LoRA r32/α64) on an engine-computed, text-blind clipped log-Nash-welfare reward over 6-party negotiation games, at two reward mixtures (λ=1.0 pure table welfare; λ=0.5 half own-outcome), 50 steps each against a frozen population-opponent zoo, with every checkpoint evaluated on a preregistered 8-cell held-out basket (abstract primary bank, prose-only sheets, three story framings, rational-agent table, ultimatum, divide-the-dollar).

Headline (preregistered gates, note 0028): every interpretable rung of both arms fails the POSITIVE gate — among-IR Nash welfare is null or resolved adverse, never favorable; held-out closure degrades at every λ (monotonically to non-play at λ=1.0, deal Δ −0.746; oscillating to −0.369 at λ=0.5); the prose transfer cell never resolves favorable on fairness; worst-off share is the lone consistent positive until it dies at λ=1's collapse endpoint; λ=1 additionally learns exact boundary-extraction (100/0 ultimatum splits at checkpoint 25) before collapsing to refusal.

Files

• contrasts.csv — 1,048 rows: every (arm × rung × cell × metric) paired contrast vs the untrained baseline, with cluster-bootstrap CIs, pair/cluster counts, and the preregistered VOID (closure-viability) flag. Long format with experiment_name for future appends.
• telemetry/steps.csv — 118 per-step training rows across all five run segments (reward terms, deal rate, below-threshold rate, GPU/host memory incl. the host/rss_gb leak-diagnosis columns).
• verdicts/ — the 17 raw analyzer JSONs the CSV is derived from.

Experiment names

experiment_name	description
grpo_v2_lam1	λ=1.0 arm, steps 1–39 (RunPod B200; died in a cgroup-memory livelock at step 40)
grpo_v2_lam1_resume25	λ=1.0 resume from checkpoint-25, steps 26–50 (fresh Adam; Amendment 6)
grpo_v2_lam05	λ=0.5 arm, steps 1–15 (RunPod B200, groups=22; pod died of balance exhaustion)
grpo_v2_lam05_resume15	λ=0.5 steps 16–31 (Lambda 1×H100, groups=14; OOM-crashed at step 32 — Amendments 9–10)
grpo_v2_lam05_resume30b	λ=0.5 steps 31–50 (Lambda 1×H100, groups=12, expandable-segments allocator)

Reading caveat: λ=0.5 rungs 20 and 35 are each 5 steps after a fresh-Adam reset and are anomalously deep on held-out closure — annotate any trend claim with steps-since-optimizer-reset.

Regeneration

Training (per arm; see Amendments 9–10 for the λ=0.5 width history):

python -m grpo_v2.train --lam {1.0|0.5} --steps 50 --groups 22 --k 8 \
  --bank instances_grpo_v2_train --out <run-dir> --checkpoint-steps 5 10 15 25 30 35 40 45 50 \
  --wandb-run-name <name> --wandb-group fairness-grpo-v2

Eval fleet per checkpoint (17 cells) + analysis:

python -m grpo_v2.launch_eval --checkpoint <label>=<adapter-dir> --include-prose --out-dir <sbatch-dir>
bash grpo_v2/submit_fleet.sh <sbatch-dir>
python -m grpo_v2.analyze_eval --cell <family>:baseline=<dirs> --cell <family>:<label>=<dirs> --out <json>

Cluster paths and runs

• Verdict JSONs: experiments/rational_agents/results/fairness_grpo_v2/ (repo) — preregistration + amendments + final tables: research-notes/0028-fairness-grpo-v2.md; ops log: self_benefit/LOG.md.
• Checkpoints (17, COMPLETE-gated, 349,243,752-byte adapters): /nlp/scr/siddharth/ii_mats/rational_agents/grpo_v2/{lam1,lam1_resume25,lam05,lam05_resume15,_lambda_final/lam05_resume30b}/checkpoint-*.
• Rollout transcripts (CoT saved): same tree, */transcripts/; eval episode JSONs: /nlp/scr/siddharth/ii_mats/rational_agents/grpov2eval_*.
• wandb (entity siddharth-stanford, project rational_agents_fairness_grpo, group fairness-grpo-v2): runs grpo_v2_lam1, grpo_v2_lam1_resume25 (glv9z8qc), grpo_v2_lam05 (vk8ktsdj), grpo_v2_lam05_resume15 (edvxz8sg), grpo_v2_lam05_resume30b (337q90b6).

Base model: Qwen/Qwen3-8B. Related: 2026.RA.QKV-Attention-Interface, 2026.RA.ToM-Hidden-Preference-Probe.

λ=0 arm — the pure-self-interest falsification control

What it is. A third arm of the same v2 campaign, trained with --lam 0.0 --shaping-weight 0.0: the reward is pure own-outcome self-interest, with no table-welfare term and no fairness shaping. Qwen/Qwen3-8B LoRA r32/α64, 15 GRPO steps, groups=12, k=8, lr=5e-05, thinking off, checkpoints at [5, 10, 15], trained on a Lambda 1×H100. It is preregistered in research-notes/0050-lam0-selfish-control-prereg.md as the falsification control for the λ=1.0 "walk-neutrality" collapse mechanism: if the closure collapse seen at λ=1.0 were an artifact of GRPO-on-this-harness rather than of the welfare reward, a self-interest reward should collapse closure too. Owning campaign record: research-notes/0028-fairness-grpo-v2.md.

Result (held-out primary bank, paired vs the untrained base). Rungs 5 and 10 are complete and interpretable, and neither collapses: deal Δ −0.028 [−0.054, −0.001] at rung 5 and −0.075 [−0.111, −0.036] at rung 10 (cluster-bootstrap CIs over 48 instance clusters), both inside the campaign's 0.10 closure-viability floor — against λ=1.0's −0.143 at the matched rung 10. The collapse is therefore not a generic consequence of training this harness with GRPO. (The λ=1.0 comparison figure and the 0.10 floor are external constants owned by research-notes/0028-fairness-grpo-v2.md; the λ=0 figures above are computed from lam0/contrasts_lam0.csv at build time.)

Rung 15 is a PARTIAL, NON-RANDOM STOP and is not a completed rung. The eval was terminated for cost while the slowest episodes were still running, leaving 346 scorable episodes of an intended 576 (352 episode files were written, 6 of them caught mid-flight with status = "running" and therefore unscorable — the termination is visible in the data itself). Termination-by-wallclock censors the slow episodes, so the surviving sample is biased toward fast-finishing (deal-closing) episodes; its rates are not a completed rung, are not comparable to rungs 5 and 10, and its closure-conditional metrics are VOID under the campaign's viability rule. Its rows are kept separate by experiment_name = grpo_v2_lam0_ckpt15_partial_terminated and by the _PARTIAL cell suffix, and it must be labelled a partial non-random stop wherever it is quoted.

λ=0 files

rung	status	shards	episode files	scorable episodes	episode file
5	complete	10	960	960	lam0/episodes/grpov2eval_lam0_checkpoint-5_primary_s*.jsonl.gz
10	complete	10	960	960	lam0/episodes/grpov2eval_lam0_checkpoint-10_primary_s*.jsonl.gz
15	PARTIAL / non-random stop	6	352	346	lam0/episodes/grpov2eval_lam0_checkpoint-15_primary_s*_partial_terminated.jsonl.gz

• lam0/contrasts_lam0.csv — 52 rows in the same long schema as the top-level contrasts.csv (same columns, same meaning), so the two concatenate directly. Kept as a separate file so this append cannot overwrite another lane's copy of the shared table.
• lam0/verdicts/ — the 5 raw analyzer JSONs those rows are derived from.
• lam0/episodes/*.jsonl.gz — 2272 held-out eval episodes, one gzipped JSONL per seed shard, each row the complete episode JSON including every turn's content, raw and reasoning.
• lam0/instances.jsonl.gz — the 48 held-out game instances (deduplicated; shared bank).
• lam0/runs/<shard>/ — run.log for all 26 shards; manifest.json (with the exact run.py invocation) and transcripts.tar.gz (the rendered Markdown/HTML views of the same episodes) for the 20 shards that have them. The 6 shards that do not are exactly the rung-15 partial ones: a run writes its manifest and renders its transcripts on completion, and these never completed. The episode records under lam0/episodes/ are unaffected and carry every model turn verbatim for all 26 shards.
• lam0/telemetry/steps_lam0.csv — 15 per-step training rows; summary.json, train.log.

λ=0 experiment names

experiment_name	description
grpo_v2_lam0	λ=0 arm, 15 steps; complete held-out evals at rungs 5 and 10 (10 + 10 seed shards)
grpo_v2_lam0_ckpt15_partial_terminated	rung 15 only — partial, non-random stop, 346/576 scorable episodes, biased toward fast-closing episodes, closure-conditional metrics VOID

λ=0 regeneration

python -m grpo_v2.train --lam 0.0 --shaping-weight 0.0 --steps 15 \
  --groups 12 --k 8 --lr 5e-05 --lora-r 32 --lora-alpha 64 \
  --bank instances_grpo_v2_train --out <run-dir> --checkpoint-steps 5 10 15 \
  --wandb-run-name grpo_v2_lam0 --wandb-group fairness-grpo-v2

# per rung, per seed shard (the invocation is also stored verbatim in each runs/<shard>/manifest.json)
python run.py --run-name grpov2eval_lam0_checkpoint-<rung>_primary_s<seed> \
  --instances instances_grpo_eval_v1 --table all_llm --arm moves_chat moves_only --seeds <seed> \
  --models <ckpt-dir> --model-thinking off --oracles threshold acceptance bestresponse --sequential
python analyze_lam0_arm.py ... --out results/fairness_grpo_v2/eval_lam0_ckpt<rung>.json

python package_hf_grpo_v2_lam0.py --out-dir $TMPDIR/hf_lam0 --push

λ=0 cluster paths

• Eval runs (episodes, instances, rendered transcripts, manifests, logs): /nlp/scr/siddharth/ii_mats/rational_agents/grpov2eval_lam0_checkpoint-{5,10}_primary_s{0..9} and .../grpov2eval_lam0_checkpoint-15_primary_s{0..5}_partial_terminated.
• Training run (checkpoints 5/10/15, telemetry, rollout transcripts): /nlp/scr/siddharth/ii_mats/rational_agents/grpo_v2/lam0/.
• Analyzer JSONs in-repo: experiments/rational_agents/results/fairness_grpo_v2/eval_lam0_*.json.
• wandb: entity siddharth-stanford, project rational_agents_fairness_grpo, run grpo_v2_lam0 (group fairness-grpo-v2).