# AGENTS.md — machine-readable guide to the TORRENT flash-flood benchmark

This dataset is designed to be driven by autonomous agents and scripts. Everything
you need is a single flat index (`manifest.jsonl`) plus precomputed download URLs.
No SDK, no auth, no HTML scraping.

## One-paragraph model
Three levels. **L1** = 42,466 reported NOAA/NCEI flash-flood episodes (1996–2025).
**L2** = 5,424 recent CONUS episodes (2021–2025) with observation availability
annotated. **L3** = 806 episode–gauge–watershed evaluation testbeds (the benchmark
unit). A conservative, human-reviewed **primary benchmark of 21 testbeds
(14 episodes, 21 gauges)** is the default set for model intercomparison. Funnel:
806 → 175 strict → 115 deduplicated → 54 NCEI-confirmed → 31 (return period ≥ 2 yr)
→ 21 primary. See the companion paper (TORRENT, ESSD).

## The one file you need
```
BASE=https://huggingface.co/datasets/skyan1002/flash-flood-benchmark-data/resolve/main
curl -sL "$BASE/manifest.jsonl" -o manifest.jsonl
```
Each line is one JSON object. `record_type=="record"` rows are catalog entries;
`record_type=="artifact"` rows are downloadable files with a `resolve_url`.

## Agent-facing filter keys (L3 records)
| key | meaning |
|---|---|
| `primary_benchmark` | true → in the 21-testbed conservative default set |
| `is_strict_flash_flood` | true → passed the strict flash-flood screen (175) |
| `dedup_selected` (`selected`) | true → deduplicated representative (115) |
| `ncei_confirmed` | true → in-basin NCEI report point (54) |
| `rp_ge_2yr` | true → LP3 return period ≥ 2 yr (31) |
| `nldi_recovered` | true → watershed restored by the hardened NLDI retrieval |
| `regulation_excluded` / `postfire_excluded` | curated out of the primary set (kept in L3) |
| `needs_curation_review` | true → auto-appended by the annual pipeline, awaiting human review |
| `water_balance_flag` | non-empty → RC > 1, use caution for volume metrics |
| `benchmark_tier` | `A` (in-basin NCEI point) or `B` (evidence label, not an exclusion) |
| `lp3_return_period_yr`, `peak_value`, `drainage_area_km2`, `unit_peak_q_cms_km2` | numeric |
| `primary_state_abbrev` / `all_states_abbrev`, `year`, `month`, `begin_date` | space/time |

## Canonical agent recipes
```bash
# The default benchmark set (21 testbeds) as clean JSON
jq -c 'select(.level=="L3" and .primary_benchmark==true)
  | {testbed_id, station_name, lp3_return_period_yr, drainage_area_km2}' manifest.jsonl

# Larger sample for robustness testing: deduplicated tier (115)
jq -c 'select(.level=="L3" and .selected==true) | {testbed_id, benchmark_tier}' manifest.jsonl

# Everything the full L3 pool offers in one state + season
jq -c 'select(.level=="L3" and .primary_state_abbrev=="TX"
  and .begin_date>="2025-07-01" and .begin_date<="2025-07-31")
  | {testbed_id, is_strict_flash_flood, primary_benchmark}' manifest.jsonl

# Download all files for one testbed (resolve URLs precomputed; -L follows the redirect)
jq -r 'select(.testbed_id=="FF_2025_07_TX_ep002__08165500" and .record_type=="artifact").resolve_url' \
  manifest.jsonl | xargs -n1 curl -L -O
```

## Server-side filter (no manifest download)
```bash
curl -sG https://datasets-server.huggingface.co/filter \
  --data-urlencode dataset=skyan1002/flash-flood-benchmark-data \
  --data-urlencode config=l3_testbeds --data-urlencode split=train \
  --data-urlencode "where=\"primary_benchmark\"=true" --data-urlencode length=100
```

## Tool schema
Machine-readable tool/function definitions for LLM agents are in
`agent/tools.json` (JSON Schema per tool: query_testbeds, get_testbed,
list_primary_benchmark, expand_l3). They map onto the recipes above and the
CLIs in `tools/`.

## Extending the benchmark (the annual pipeline)
`pipeline/expand_l3.py` discovers newly qualifying testbeds from a given NCEI
year and appends them to `catalog/raw_csv/l3_testbeds.csv` as light rows
(`needs_curation_review=true`, no forcing). `pipeline/rebuild_catalog.py`
regenerates the Parquet + manifest. `pipeline/.github/workflows/annual_l3_update.yml`
runs both once a year and pushes the refreshed catalog. Forcing for a new light
testbed is fetched on demand with `tools/download_forcing.py`.

## Hard rules
- Always `curl -L` (LFS/CDN 302 redirect; without `-L` you get a pointer stub).
- Never `awk -F,`/`cut -d,` the raw CSVs — fields contain commas. Use `jq` on
  `manifest.jsonl`, the Parquet configs, or the `/filter` API.
- `primary_benchmark` is the paper's default set; `needs_curation_review` rows are
  machine-appended candidates that have not yet had the human curation review.