| --- |
| title: FloodGate v1 |
| emoji: 🌊 |
| colorFrom: blue |
| colorTo: green |
| sdk: gradio |
| sdk_version: 6.29.1 |
| python_version: "3.12" |
| app_file: app.py |
| license: mit |
| pinned: false |
| short_description: Run the Army Corps' Cumberland dams through real 2025 floods |
| tags: [openenv, simulation, rl-environment, flood-control, water-resources, operations-research, beam] |
| --- |
| |
| # FloodGate v1 · OpenEnv |
|
|
| The agent is the U.S. Army Corps of Engineers' water manager for the four flood-control dams above Nashville: Wolf |
| Creek, Dale Hollow, Center Hill and J. Percy Priest. It runs them through the real February and April 2025 Cumberland |
| River floods, rebuilt from the public record (USACE dam operations, USGS gauges, NWS flood stages). Every 6 hours it |
| sets each dam's release. The water takes 2-26 hours to reach the flood gauges at Burkesville, Celina and Nashville, and |
| every acre-foot it holds back is risk for the next storm. |
|
|
| **This Space serves live mode.** The agent commits releases one day at a time, without knowing the future: |
| - it sees 8 forecast scenarios, each replaying the error path of a real 2025 National Weather Service river forecast; |
| - committed water cannot be recalled; |
| - each decision is graded against the scenario-optimal decision, a two-stage stochastic LP. The grade is the expected |
| cost the decision added, so luck does not count. |
|
|
| | page | what | |
| |---|---| |
| | `/web` | OpenEnv's web UI. **Explore**: the 3D replay and every committed run's results. **Playground**: reset, step and the tools by hand | |
| | `/twin/` | the 3D replay: any plan for any task on the basin's terrain, rivers and reservoirs (USGS), with the hydrographs | |
| | `/reset`, `/step`, `/state`, `/ws`, `/mcp` | OpenEnv | |
| | `/api/tasks`, `/api/tasks/{id}/plans`, `/api/tasks/{id}/replay?plan=` | read-only JSON | |
|
|
| ## Tools (live mode) |
|
|
| | tool | returns | |
| |---|---| |
| | `get_status()` | the clock and the block to commit, storage, recent flows, every forecast scenario | |
| | `check_plan(plan)` | a draft for the remaining steps simulated under every scenario: cost spread, gauge peaks, lake levels (10 per episode) | |
| | `commit(releases)` | commits the day's releases, e.g. `{"WC": 12000, "DH": [3000, 3500, 4000, 4000]}`; the last commit ends the episode | |
|
|
| ## Results on the 24 sample tasks (live mode) |
|
|
| | model or policy | mean reward (95% CI) | real | storm ×1.5 | half full | solved | |
| |---|---|---:|---:|---:|---:| |
| | scenario-optimal decisions (the reference) | 1.000 | 1.000 | 1.000 | 1.000 | 24 | |
| | a perfect solver planning on the mean forecast (no model) | 0.552 | 0.517 | 0.504 | 0.636 | - | |
| | DeepSeek-V4.1-Flash | **0.364** (0.306-0.428) | 0.414 | 0.334 | 0.344 | 0 | |
| | Beam-1.0-501B-A23B, default effort | **0.236** (0.217-0.261) | 0.226 | 0.263 | 0.219 | 0 | |
| | Beam-1.0-501B-A23B, reasoning_effort high | **0.237** (0.217-0.265) | 0.216 | 0.260 | 0.234 | 0 | |
| | hold every release (the lazy operator) | 0.215 | 0.215 | 0.215 | 0.215 | - | |
| |
| - **Live mode is hard for every model:** none solves a task, and none reaches a perfect solver that simply plans on |
| the mean forecast. |
| - **Beam over-releases into a river already in flood.** It opens dams to their maximum while Nashville is above |
| major flood stage, and on half the tasks its decisions add more expected cost than doing nothing. |
| - **More reasoning effort does not change Beam's score.** DeepSeek-V4.1-Flash beats Beam on 22 of the 24 tasks. |
| |
| ## Data |
| |
| All U.S. federal works, public domain: |
| - dam release, storage and pool levels: USACE CWMS; |
| - river flow and stage: USGS; |
| - flood categories and ratings: NOAA NWS; |
| - the river forecasts NWS issued in 2025: the Ohio River Forecast Center, via the Iowa Environmental Mesonet archive; |
| - the replay's terrain and hydrography: USGS 3DEP and NHD. |
| |
| Each rule's source (real, derived or assumed) is in `data/cumberland/SOURCE.md`. Tasks, data and every run's plans: |
| the [FloodGate dataset](https://huggingface.co/datasets/bpHigh/FloodGate). |
| |
| ## How Beam was used |
| |
| Beam (`reflection/Beam-1.0-501B-A23B`) played every task through the Hugging Face router (Baseten), at default and |
| high reasoning effort: |
| - its `reasoning_content` was sent back on every turn, as Beam's providers require; |
| - its 1M-token context holds a whole live episode (6 decisions, 8 scenarios each) without truncation; |
| - its plans are in the replay and the dataset. |
|
|
| The environment is also an RL environment for Beam: a deterministic grader with no LLM in the reward, and an exact |
| reference for every task. |
|
|
