other/cycling_time_trial_pacing - reference output and evaluation dependencies (HIDDEN: do not publish).

expert_output/output/             the reference agent's output (engine/, plans/, results.json), made by
                                  reference_solution/oracle/run_reference.sh on the task's input/ and software/
expert_output/evaluator_report.json   the evaluator's report on it (score 1.0)
reference/                        hidden task data staged only for evaluate(): protocol.json (limits, thresholds,
                                  weights, hidden and development race days with reference times, calibration
                                  record), hidden/<race day>/ (scenario.json, course.csv, weather.csv),
                                  reference_plans/ (the plan behind every reference time)
evaluation/                       main.py (load/start/evaluate), task_card.json, scripts/grader/ (VM-side evaluator:
                                  run_eval.py, sandbox.py, integrity.py, ttsim_pkg/ = verbatim copy of the public
                                  simulator), env_package_python3-numpy-scipy/ (image installer)
reference_solution/oracle/        the reference engine (pacer), its offline best-known search (bestknown) and the
                                  agent workflow run_reference.sh
reference_solution/data_build/    how the task data were built (OSM routing, DEM, ERA5, scenarios, protocol)
reference_solution/tests/         local evaluation, negative controls, calibration engines, docker end-to-end test
