Blame for data/SOURCE.md

# Sources: the Cumberland River record (January-June 2025)

All sources are works of the U.S. federal government (public domain). `envs/flood_ops/tools/curate_cumberland.py`
fetches them, caches the raw responses in `raw/` (git-ignored), and writes `basin.json` and `record.csv.gz`.

## Real (the record)

| what | where | series |
|---|---|---|
| dam release, cfs (hourly mean) | USACE CWMS Data API, office LRN | `<dam>.Flow.Ave.1Hour.1Hour.man-rev` |
| dam storage, acre-feet (hourly) | same | `<dam>.Stor.Inst.1Hour.0.man-rev` |
| pool elevation, ft (hourly) | same | `<dam>.Elev-Pool.Inst.1Hour.0.man-rev` |
| top and bottom of the flood pool, top of the inactive pool | CWMS location levels | `<loc>.Stor.Inst.0.Top of Flood` etc. J. Percy Priest's bottom of flood pool is seasonal: 483 ft to 21 March, 490.5 ft from 21 April to 9 October |
| flow at Burkesville KY | USGS 03414100 (Water Data OGC API) | discharge, 00060 |
| flow at Nashville TN | USGS 03431500 | discharge, 00060 |
| stage at Celina TN | USGS 03417500 | gage height, 00065 (no discharge is published) |
| flood categories, stage-flow ratings | NOAA NWPS, gauges BRKK2, CLAT1, NAST1 | action / minor / moderate / major |

Dams: WLCK2-WOLF_CREEK, DLHT1-DALE_HOLLOW, CEHT1-CENTER_HILL, JPHT1-J_PERCY_PRIEST.

Flood thresholds used (flow, cfs):

| gauge | action | minor | moderate | major | from |
|---|---:|---:|---:|---:|---|
| Burkesville (BRK) | 32,500 | 40,000 | 57,500 | 90,000 | NWS published flows |
| Celina (CEL) | 52,006 | 76,030 | 91,010 | 110,011 | NWS stages 30/40/45/50 ft through the NWS rating |
| Nashville (NAS) | 70,072 | 119,465 | 128,065 | 154,611 | NWS stages 30/40/42/45 ft through the NWS rating |

Gaps: the hourly series are complete except Celina (85 hours) and Nashville (3 hours), filled by linear
interpolation (`basin.json`, `gap_hours_filled`).

| river forecasts NWS issued, Jan-Jun 2025 (live mode) | Ohio River Forecast Center products RVFCMU and RVFCML, from the Iowa Environmental Mesonet's NWS text archive (`mesonet.agron.iastate.edu/cgi-bin/afos/retrieve.py`) | SHEF `.ER ... HGIFF` stage forecasts, 6-hourly to about 5 days, at BRKK2 (30 issues), CLAT1 (219), NAST1 (228); turned into flow through the NWS ratings; `forecasts.csv.gz` (`tools/curate_forecasts.py`) |

| the 3D replay's map: imagery, terrain, rivers, reservoirs | USGS The National Map orthoimagery (USGSImageryOnly, USDA NAIP; the Kentucky side colour-matched to Tennessee's at the state-line seam); USGS 3D Elevation Program (3DEPElevation ImageServer); USGS National Hydrography Dataset (NHDPlus small-scale flowlines and waterbodies); dam locations from CWMS, gauge locations from NWPS | `openenv/flood_openenv/web/twin/basin.json.gz` (`tools/twin/build_basin.py`); display only, never in a cost |

## Derived

- **6-hour steps (UTC):** mean release and flow over the step; storage and elevation at its start.
- **Dam inflow** I[t] = R[t] + (S[t+1] − S[t]) / 0.49587, from the record's own mass balance. It is noisy where the
  storage record is (it can be negative for a step).
- **Local runoff at a gauge** L[t] = Q[t] − Σ routed record releases. With these two definitions the Corps' own
  releases reproduce the record's storage and gauge flows exactly (`core/tests/test_core.py`).
- **Travel times** (hours), from cross-correlating 24-hour-smoothed hourly changes in the 2025 record, reach by reach:

  | reach | lag |
  |---|---:|
  | Wolf Creek → Burkesville | 10 |
  | Burkesville → Celina | 8 |
  | Celina → Cordell Hull (pool) | 0 |
  | Cordell Hull → Old Hickory | 6 |
  | Old Hickory → Nashville | 2 |
  | Center Hill → Old Hickory | 16 |
  | J. Percy Priest → Nashville | 2 |

  Dale Hollow's releases are too steady to identify. Its 7.3 miles to the Cumberland use the 3.5 mph measured below
  Wolf Creek (2 h). The sums give the dam-to-gauge lags in `basin.json`. A fractional lag is split linearly between
  two 6-hour steps.
- **Maximum release:** the largest hourly release in the 2018-2025 record. WC 59,900, DH 10,100 and CH 25,300 (all
  February 2019), JPP 14,900 (February 2020).
- **Ramp:** the 99.9th percentile of the change between consecutive 6-hour mean releases, 2018-2025, rounded: WC
  18,000, DH 4,500, CH 10,000 and JPP 5,000 cfs per step.
- **Storage → elevation tables:** from the hourly record plus the official levels. Used for display only.

- **Forecast errors (live mode):** for each issued forecast, the log ratio of forecast to observed flow, lead by
  lead (observed = the mean of the two 6-hour steps around the valid time). 214 Celina and 222 Nashville forecasts
  reach 5 days. The spread grows from 0.09 (Celina) and 0.13 (Nashville) at 0-12 h to 0.38 and 0.36 at 108-120 h.
  The bias falls to -0.26 and -0.19 at 108-120 h.
- **Forecast scenarios (live mode):** each of the 8 members replays one issued forecast's whole error path,
  multiplied into the true series (positive values only).
  - Upper-basin series take Celina's path: Wolf Creek and Dale Hollow inflow, Burkesville and Celina local flow.
  - Lower-basin series take the Nashville path from the product issued closest in time.
  - There are no forecasts at the four storage lakes, so their inflows borrow the nearest gauge's error.
  - Leads beyond 5 days keep the 5-day error.
  - Members are seeded by task and decision.

## Assumed (modelling choices)

- Cordell Hull and Old Hickory are run-of-river: their pool changes are part of the local runoff.
- Minimum release 0 cfs. In April 2025 Wolf Creek released 20 cfs for three days.
- The cost weights (README) and the 6-hour decision step.
- Drawn tiers, labelled in the task text:
  - `storm`: runoff above the week-before baseflow multiplied by 1.5, at the dams and at the gauges;
  - `full`: every lake starts with half its flood pool in use.
- Plan mode: the agent knows the window's inflows and runoff exactly (hindsight). Live mode replaces that with the
  forecast scenarios above.
- Live mode: the spill penalty (100 per 1,000 acre-feet per step above the top of the flood pool or below the
  inactive pool), 24-hour decision blocks, 8 scenarios.