v0.1.12

Data and analysis release for the third revision of the manuscript.

• Fix: in v0.1.11 the replicate directories of the counterbalanced end-to-end runs (experiments/audit_cb_) and of the interleaved temperature runs (experiments/temp_il_) were published incompletely (a single misnamed directory per run instead of rep01..rep12 / rep01..rep06). This release contains all 120 replicate directories; use this release, not v0.1.11, for those runs.
• New: harness/scripts/audit_sim.py and experiments/audit_sim.json, a simulation of the audit statistics with known truth (dense, opposing, sparse and population scenarios; vignette-level and replicate-paired sign-flip tests; excess detection; interval coverage). It replaces the injection into observed draws, which is no longer run.
• audit_validation.py now reports the number of non-zero vignette-level shifts, the smallest attainable sign-flip p-value and a post hoc replicate-paired sign-flip test.
• verify_paper_numbers.py checks 681 claims against these artifacts.

v0.1.11

Adds the data and scripts for the second revision of the manuscript.

• End-to-end check rerun with a counterbalanced condition order: two local models (llama3.1:8b, qwen3:4b), reference, byte-identical null control and one demographic contrast, R=12. Each of the six condition orders is used twice, fixed before any call (audit_cb_schedule.py, audit_cb_schedule.json), and every step is time-stamped (FMA_ORDER_LOG in run_ollama_floor.py; audit_cb_order_log.jsonl). Results in audit_validation.json; order and elapsed-time permutation check in audit_order_check.py / audit_order_check.json.
• The earlier fixed-order run (R=10, audit_val_*) is kept as superseded; its results moved to audit_validation_audit_val.json (audit_validation.py --prefix audit_val).
• Interleaved default versus temperature-0 check on the four local models (temp_il_*, run_temp_il.sh, temp_il_compare.py, temp_il_compare.json), replacing reliance on the block-ordered comparison alone.
• verify_paper_numbers.py checks 597 claims against these artifacts.

Note (2026-10-14): the audit_cb_* and temp_il_* replicate directories in this release are incomplete. Use v0.1.12.

v0.1.10

Adds the analyses and data for the revised manuscript: an end-to-end check of the audit procedure on two local models (llama3.1:8b, qwen3:4b; reference, byte-identical null control and one demographic contrast; R=10, interleaved) with injected known effects (audit_validation.py, audit_render_diff.py, experiments/audit_val_); a same-window interleaved comparison of isolated and batched calls for claude-haiku-4-5 (claude_cli_batched_answerer.py, haiku_dispatch_compare.py, experiments/floor16_haiku_il_); dependence-robust zero-discordance bounds (zero_bounds_v4.py); correlated-action coverage scenarios (coverage_sim.py --v4); range and applicability analyses (range_v4.py, applicability_v4.py). verify_paper_numbers.py checks 572 claims; 91 tests pass.

v0.1.9

Repository cleanup with no change to any reported number. The logged v3 reruns are now the only local-model panel (FMA_PANEL=v3 is the default); the superseded Ollama 0.34.3 local runs and their analysis outputs are removed and remain available in v0.1.8. Unused build, review and pilot-reconstruction scripts and internal planning documents are removed; method notes (CROSSED_BOOT.md, COVERAGE_SIM.md) are rewritten as neutral documentation; panel_summary.py reports the five-action pooled floor as the headline; figure scripts take --outdir; temperature_compare.py no longer degrades its shipped output when attempt logs are absent; README, harness README and docs are aligned with the current manuscript. verify_paper_numbers.py checks 407 claims; 91 tests pass.

v0.1.8

Adds the crossed (vignette by replicate) bootstrap used for every interval in the current manuscript (crossed_boot.py), its coverage simulation (coverage_sim.py), exact one-sided bounds for zero-observed floors (fixed_case_bounds.py), and an isolated-call sensitivity arm for endpoint A (claude-haiku-4-5, one claude -p call per prompt, ten replicates; claude_cli_answerer.py, haiku_iso_floor.py, experiments/floor16_haiku_iso). README results updated to five-action floors with crossed intervals and the new paper title. verify_paper_numbers.py checks 407 claims.

FairMedAgent 0.1.7

Adds logged reruns of the four local models (Ollama 0.35.1; ten default-sampling and five temperature-0 replicates each) with attempt-level logs, where a retry uses the same decoding configuration as the first attempt. Adds five-action analyses, first-attempt sensitivity and attempt-summary scripts, and the FMA_PANEL=v3 switch that selects the reruns for the reported numbers. README results summary updated to the current manuscript.

FairMedAgent 0.1.6

Citation metadata release. CITATION.cff and the README now cite the paper as the preprint arXiv:2609.03221 (DOI 10.48550/arXiv.2609.03221). No code or data changes since 0.1.5 (commit 3982974). The updated supplementary material is attached.

v0.1.5: hosted temperature-1.0 arm

Adds the temperature-1.0 arm of the hosted check (openai/gpt-oss-20b on Groq, five replicates). Temperature 0 lowers the hosted floor from 0.100 to 0.024 over the 15 vignettes complete in both arms, a 76 percent reduction that leaves it above zero. The verifier checks 193 claims. Supplementary material attached. Commit 3982974.

v0.1.4: hosted temperature-0 check

Adds a hosted-endpoint check: openai/gpt-oss-20b on Groq at temperature 0 (five replicates) still disagreed with itself in 22 of 960 comparisons, across 48 distinct system fingerprints. Adds groq_answerer.py and selects answerers with FMA_ANSWERER. The verifier checks 188 claims. CITATION.cff cites the manuscript Instability Floors: Separating Bias from Noise in Fairness Audits of Clinical LLM Agents with FairMedAgent. The supplementary material is attached. Commit a6fd6d6.

v0.1.3: temperature-0 runs and robustness analyses

Adds temperature-0 replicates for the four locally served models, robustness analyses (paired bootstrap differences, leave-one-vignette-out, results excluding the undefined caution output, a bootstrap interval for the majority-vote reduction, a null-simulation sensitivity check, and a signed paired-test simulation), and single-sourced intervals at 5000 resamples. python harness/scripts/verify_paper_numbers.py checks 183 claims. The supplementary material is attached. Commit 80f6549.