FairMedAgent: instability-floor trajectories

Raw agent trajectories behind the per-action instability floor reported for FairMedAgent, an evaluation harness for demographic disparity in the actions of multi-step clinical LLM agents.

Code, protocol and analysis scripts: https://github.com/rohithreddybc/FairMedAgent

Archived software release: https://doi.org/10.5281/zenodo.22165979 (archival identifier, not the citation)

Why this exists

A counterfactual fairness audit holds the clinical content of a case fixed, changes only the patient descriptor, and reports how often the agent's action changes. On its own that number cannot be interpreted, because an agent's action also changes when nothing changes at all.

These files are the measurement of that baseline. Re-running an identical condition ten times over sixteen vignettes, with the same narrative and the same descriptor string and nothing varied, moved the primary model's (haiku-4.5) action in 6.8 percent of replicate pairs over five clinically defined actions (admission, ICU escalation, high acuity, referral, any opioid). The rate differed across actions from 2.2 percent for ICU escalation to 14.2 percent for admission. Across six models (haiku-4.5 and sonnet-5, hosted; llama3.1:8b, mistral:7b, qwen3:4b and phi3:mini, served locally) the pooled five-action floor ranged from 1.7 to 24.5 percent. The controlled-substance caution flag is reported separately as an exploratory, model-generated flag rather than a clinically defined action.

The floor depends on decoding. In logged reruns of the four local models (Ollama 0.35.1, attempt-level logs, a retry uses the same decoding configuration as the first attempt), temperature-0 decoding left no disagreement across five replicates for three of the four models; phi3:mini kept a small floor. A hosted model (openai/gpt-oss-20b on Groq) did not reach zero: its floor was 0.100 at temperature 1.0 and 0.024 at temperature 0 on matched vignettes, a 76 percent reduction, with disagreement remaining at temperature 0. Majority voting over five draws removed 53 percent of the primary model's five-action floor (6.8 to 3.2 percent).

The accompanying manuscript is Instability Floors: Separating Bias from Noise in Fairness Audits of Clinical LLM Agents with FairMedAgent (arXiv:2609.03221, DOI 10.48550/arXiv.2609.03221).

No disparity result is claimed here, and none should be quoted from this dataset. The estimand the harness targets counts only flips between actions a published decision rule admits and a clinician has adjudicated as defensible. That adjudication is under way.

Contents

Path	Contents
experiments/floor16/	Ten repetitions of an identical condition over sixteen vignettes
experiments/floor16_sonnet/	Second hosted model, sonnet-5 (Claude Sonnet 5, identifier claude-sonnet-5 as reported by the calling layer), same design (six replicates). The directory name is unchanged
experiments/floor16v3_llama/, floor16v3_mistral/, floor16v3_qwen3/, floor16v3_phi3/	Logged reruns of the four local models (llama3.1:8b, mistral:7b, qwen3:4b, phi3:mini; Ollama 0.35.1), ten replicates each, default sampling. Each rep*/answers/s*.json carries per-attempt logs (timestamps, durations, parse status); attempt_summary.json totals them. A retry uses the same decoding configuration as the first attempt
experiments/floor16v3_*_t0/	The same four models at temperature 0, five replicates each
experiments/floor16v3_provenance.txt, run_v3_all.sh	Ollama version, model tags and digests, environment, and the driver that produced the v3 runs
experiments/floor16_ollama/, floor16_mistral/, floor16_qwen3/, floor16_phi3/, floor16_*_t0/	The earlier local-model runs, kept for comparison; the reported numbers use the v3 reruns
experiments/floor16_groq_t0/, floor16_groq_t1/	Hosted check: openai/gpt-oss-20b on Groq at temperature 0 and 1.0, five replicates each
experiments/robustness_v3.json, temperature_compare_v3.json, five_action_extras_v3.json, first_attempt_v3.json, panel_summary_v3.txt, rank_tests_v3.txt	Reported analysis outputs (FMA_PANEL=v3), generated by the harness scripts
experiments/robustness.json, temperature_compare.json	The same analyses on the earlier local-model runs
experiments/instability/	Earlier instability probe that motivated the floor study
experiments/pilot, pilot2, pilot3, pilot3ctl	Pilot runs, including the rare-token, re-render and sham-attribute controls
docs/DATASHEET.md	Datasheet for the synthetic cohort
docs/COHORT_PROVENANCE.md	Generator version, seed and command, so the cohort is regenerable
docs/ADJUDICATION_PROTOCOL.md	How an acceptable-action band is adjudicated
docs/BAND_ADJUDICATION_RECORD.md	The adjudicator's reasoning, in full, per band

Each trajectories.json holds complete six-stage trajectories: five model-facing decisions around a deterministic environment step, with the condition identifier, the vignette identifier and the structured action at every stage.

Not included

The development split and the reference leaderboard are published on completion of clinician validation. The held-out test split is sealed and scored under the submission protocol; it is deliberately absent so that it stays a held-out split.

Reproducing the reported numbers

git clone https://github.com/rohithreddybc/FairMedAgent
FMA_PANEL=v3 python FairMedAgent/harness/scripts/verify_paper_numbers.py

FMA_PANEL=v3 selects the logged reruns (floor16v3_*) for the local models. The script recomputes the reported quantities from these trajectories. Pass the path to the manuscript source to also assert that each figure appears in the text as computed.

Data statement

All patient vignettes are synthetic. No real patient data, no protected health information, no human subjects, no IRB required. Cohort provenance is recorded rather than the cohort itself, so it is regenerable from the seed and command in docs/COHORT_PROVENANCE.md.

Citation

Cite the paper, not the archive. The citable reference is the preprint arXiv:2609.03221 (DOI 10.48550/arXiv.2609.03221):

@article{bellibatlu2026instability,
  title   = {Instability Floors: Separating Bias from Noise in Fairness Audits of Clinical {LLM} Agents with {FairMedAgent}},
  author  = {Bellibatlu, Rohith Reddy and Singh, Manpreet and Parashar, Deepak and Joshi, Rahul},
  journal = {arXiv preprint arXiv:2609.03221},
  year    = {2026},
  doi     = {10.48550/arXiv.2609.03221},
  url     = {https://arxiv.org/abs/2609.03221}
}

Zenodo indexes software records but Google Scholar does not, so a Zenodo-only citation does not accrue anywhere a reader or a bibliometric tool will look for it.

10.5281/zenodo.22165979 is the archival identifier for the software itself. It belongs in a data or code availability statement, which is where journals ask for it, and it is what makes a specific version of this harness retrievable years from now. It is not the reference to put in a bibliography.

See CITATION.cff for the machine-readable form.