cf-uncertainty: epistemic and aleatoric uncertainty in learned car-following

Abstract

Nine strengthening items after the extension of v1.4.0, their code, configurations, analysis tables and the compact tables c25-c31 of the extended manuscript. Plan and reading rules: extension/docs/PLAN.md, section S. Code. The model-spec key sampler of an E3 variant (DiffusionCF.configure_sampler: re-planning interval and DDIM steps at inference); plan sharing between followers (record_plans and the v_plan leader override of the rollout engine, the plan window of the diffusion sampler, simulate_platoon(share_plans=True), the E3 variant key share_plans); the one-step networks fine-tuned in closed loop (cf_unc/models/closed_loop.py: hetero_mdn_cl, lstm_const_sigma_cl, initialised from the stored twins and trained with a 5 s rollout loss); the scripts of the offline items (extension/scripts/state_reweight.py, conditional_coverage.py, string_stability.py, calibrate_idm_modes.py, idm_ref_rerate.py, replan_tradeoff.py, closed_loop_compare.py) and the GPU queues queue2.sh, queue3.sh. Tests: tests/test_sampler_settings.py, test_plan_sharing.py, test_closed_loop.py; the fast suite passes. Results (results/extension/ /tables/): ext_reweight (S2, the state-distribution part of the shift), ext_e3_replan (S3 sampler settings and S8 plan-sharing cells, replan_tradeoff.csv), ext_e3_preview (S4 factorial rows of preview_sweep.csv), ext_coverage (S5 conditional coverage, CQR, interval score), ext_idm_modes and ext_e3_idmref (S6 references per driver mode, h32_mode_reference.csv), ext_cl_highd, ext_cl_pool, ext_cl_lodo, ext_e2_cl, ext_e3_cl, ext_cl_summary (S7), ext_string (S9). The H3.1 and H3.1e verdict tables of E1 are added to results/controls/e1/tables as sources of the scorecard. results/compact_tables/c25_scorecard.tex to c31_closedloop.tex with the generator scripts/compact_tables/make_ext_tables.py; results/TABLE_SOURCES.md extended. Assets. checkpoints_s7.tar.gz: the 30 fine-tuned checkpoints of the three closed-loop roots with their training logs and metrics, and the 10 mode-calibrated IDM references, with a SHA-256 manifest (CHECKPOINTS_S7_MANIFEST.json; unpack into extension/artifacts/). s_platoon_tables.tar.gz: the per-profile platoon tables, configs, specs and re-smoothed profiles of ext_e3_replan, ext_e3_idmref and ext_e3_cl. The assets of v1.3.1 and v1.4.0 are unchanged.