AirOpsEnv v1 (v0.2.0)

AirOpsEnv v1. An agent runs JetBlue's operations desk through a real storm day at Boston Logan or JFK. It cancels, restores and holds flights watch by watch on AgentEnv's virtual clock, and is graded against the plan a MIP proves optimal in hindsight.

This release replays every run as the airport on its storm day:

• The airport in 3D.

  • Logan and JFK are modelled from OpenStreetMap over USGS aerial photos, with the harbour and Jamaica Bay as water.
  • Every airline's real traffic moves on its BTS times: JetBlue flies the run's plan in its colours, with plain-English tags, and the other airlines are grey.
  • Each day's own METARs set the weather: rain, snow, thunderstorms, fog, and day turning to night.

• Around it:

  • a header with the weather in plain English;
  • a desk that explains each watch's FAA and weather news, and what the agent did, in words;
  • an airport-style departures and arrivals board;
  • the score against the references;
  • the US radar map as a swappable inset.

  Today's tail chart, slot caps and raw tool calls sit behind Expert.

• Films. scripts/record.py --director cuts a run into a film: the FAA program arriving, the agent's decisions landing on the board, a held flight finally leaving, and the final score. The five films of GPT-6.1 Sol's runs are attached; the hero is Boston on 16 February 2025.

• v1 structure. The dataset's tables are now v1_tasks, v1_references and v1_episodes, plus results. The README and dataset card take the PortSimEnv v1 card shape.

The env, its 57 days, the grading and the recorded runs are unchanged. Stands, taxi routes and the runway in use in the viewer are inferred for display only.

• Dataset (tag v0.2.0): https://huggingface.co/datasets/earakely-scale/AirOpsEnv-AgentEnv
• Replays in 3D: https://huggingface.co/spaces/earakely-scale/AirOpsEnv-AgentEnv

Airport © OpenStreetMap contributors (ODbL); imagery USGS The National Map; radar NEXRAD via the Iowa Environmental Mesonet; flights BTS; advisories FAA ATCSCC; weather NWS via the IEM. The 3D scene adapts code from PortSimEnv v1 (FineEnvs, Apache-2.0).

AirOpsEnv v0.1.0

AirOpsEnv's first release: an agent runs JetBlue's operations desk at JFK or Boston through a real storm day, watch by watch on AgentEnv's virtual clock, and is graded against the plan a MIP proves optimal in hindsight.

• Env airops-day: per-flight cancel, restore and hold decisions, executed on the day as it really happened (BTS on-time data, the FAA Command Center's ground stops and ground delay programs as hourly slot caps, TAFs and METARs as issued). Tools: get_situation, check_plan, confirm_changes, advance.
• Pack b6-v1: 57 days from 13 months of 2022–2025. 20 eval days (12 JFK, 8 BOS), 3 control days, 34 train days. Every eval day qualifies: doing nothing costs 26–66% over the hindsight optimum and the re-planning follower lands within 15% of it. The references are deterministic, and the pack rebuilds byte-identically from pinned inputs.
• Viewer: the network map with NEXRAD radar, the hub in 3D from OpenStreetMap, the tail chart with the slot caps, and the ops panel with each watch's news and tool calls. agent-env airops view and agent-env airops record.

GPT-6.1 Sol on 5 eval days (sol-v1, one run each, $2.99 in all):

Day	GPT-6.1 Sol	Follower	Blind	Doing nothing	JetBlue
b6-bos-20250216-1500	0.672	0.937	0.937	0.366	0.224
b6-jfk-20230714-1400	0.441	0.782	0.229	0.212	0.200
b6-jfk-20240804-0700	0.382	0.901	0.627	0.213	0.202
b6-jfk-20250709-1500	0.324	0.614	0.743	0.224	0.200
b6-bos-20240107-0600	0.276	0.417	0.436	0.220	0.202
Mean	0.419	0.730	0.594	0.247	0.206

The attached film is Sol's run of the Boston storm of 16 February 2025 (60 s, 1080p).

• Dataset (days, runs, transcripts, episodes): https://huggingface.co/datasets/earakely-scale/AirOpsEnv-AgentEnv
• Replay Space: https://huggingface.co/spaces/earakely-scale/AirOpsEnv-AgentEnv

Basemap © OpenMapTiles © OpenStreetMap contributors via OpenFreeMap; hub twin © OpenStreetMap contributors (ODbL); radar NEXRAD n0q via the Iowa Environmental Mesonet.