Eve library

Phase 175 locally actionable completion evidence

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Date: 2026-07-15

Completion audit#

The Phase 175 checklist now has seven unchecked entries. Every one is an external training, benchmark, accelerator, or publication outcome; none is a remaining repository-local implementation task:

Task Required completion evidence Why it remains open
175.2.2.1 300k-step 1° ERA5 run plus staged multi-step fine-tuning and published-skill comparison Requires the full ERA5 archive and a sustained accelerator run.
175.2.2.2 226M-parameter 0.25° checkpoint trained with FSDP2/pjit on an H100/TPU pod Requires an accelerator pod and the full 0.25° archive.
175.3.1.3 1979–2018 training run and CRPS improvement over IFS-ENS on 97% of 1,320 targets Requires ERA5/IFS-ENS archives, accelerator execution, and benchmark outputs.
175.4.1.3 End-to-end trained checkpoint and operational landfall-skill win over HAFS/HWRF Requires operational forecast archives and empirical NHC verification.
175.6.1.1 CMIP6-trained 100-year rollouts with GMST drift at or below 0.05 K per decade Requires CMIP6 runs and long accelerator rollouts.
175.7.1.1 Trained 25 km to 2 km CorrDiff checkpoint validated on HRRR and CERRA Requires paired regional reanalysis archives and accelerator training.
175.9.1.5 Published scientific papers and public open-weight releases Requires completed major-model training, release artifacts, licensing review, and publication authority.

The code components needed to execute those external runs are present in the corresponding packages: trainable model architectures, objectives, rollouts, stability/budget diagnostics, and benchmark metrics. Marking any of the seven items complete from synthetic local tests would falsely claim the specified empirical result.

This audit found two genuinely local gaps and closed both:

  1. Observation transforms existed for tasks 175.1.2.2–6, but provider request clients did not. ingest.observation_clients now supplies cadence-validated, tier-aware requests and archive keys for EUMETNET ORD OPERA composites; GOES-16/18, MTG-I1 and Himawari-9 imagery; GPM IMERG Early/Late/Final; METAR, SYNOP, BUFR and ASOS station feeds; and GOES GLM/EUCLID lightning.
  2. The ten-member deterministic baseline in 175.2.2.3 had no implementation. gaia_graphcast.ensemble now supplies reproducible unique seeds, leave-one-fold-out subset plans, checkpoint-compatible member execution, full member products, ensemble mean, population spread, and rollout aggregation.

The phase-wide verification audit also found two gate defects in already checked Phase 175 packages. The GRIB/Zarr codec did not pass its canonical cargo fmt --check, and forecast-serving lacked the Node type declaration and TypeScript 6 deprecation setting needed by its declared typecheck/declaration build. The codec has been mechanically formatted, and forecast-serving's package/tsconfig metadata now makes both TypeScript gates pass.

Requirement-to-evidence map#

Checklist scope Authoritative repository evidence Automated proof
175.1.2.2 weather-data/ingest/observation_clients.py::OPERAClient Request, cadence, product, and key tests
175.1.2.3 GeostationarySatelliteClient plus existing SatPy calibration seam GOES/MTG/Himawari and 1–16 band tests
175.1.2.4 IMERGClient plus IMERGGrid Early/Late/Final CMR, 0.1°, half-hour tests
175.1.2.5 StationNetworkClient plus METAR/SYNOP decoders All four feed and decoder tests
175.1.2.6 LightningClient plus flash-density gridding GLM/EUCLID semantics and gridding tests
175.2.2.3 graphcast/ensemble.py Plan uniqueness, subset, aggregation, rollout tests

Verification#

All applicable gates for the two changed Python packages pass:

bash
cd libs/gaia/weather-data
python3 -m ruff check gaia_weather_data tests
python3 -m mypy gaia_weather_data
python3 -m pytest tests/ -q                       # 58 passed
python3 -m build                                  # sdist + wheel

cd libs/gaia/graphcast
python3 -m ruff check gaia_graphcast tests
python3 -m mypy gaia_graphcast
python3 -m pytest tests/ -q                       # 22 passed
python3 -m build                                  # sdist + wheel

The complete Python model-package test surface also passes, sequentially:

text
weather-data       58 passed
graphcast          22 passed
gencast            23 passed
cyclone-forecast   13 passed
nowcast            13 passed
climate-emulator   12 passed
downscale           8 passed
                  ---
                  149 passed

Forecast serving passes its test, type, and declaration-build gates:

bash
pnpm exec vitest run --root libs/gaia/forecast-serving  # 29 passed
pnpm exec tsc --noEmit -p libs/gaia/forecast-serving/tsconfig.json
cd libs/gaia/forecast-serving && pnpm exec tsup

The sovereign GRIB2/Zarr codec passes its full Rust gates:

bash
cargo fmt --manifest-path libs/gaia/grib-codec/Cargo.toml -- --check
cargo clippy --manifest-path libs/gaia/grib-codec/Cargo.toml --all-targets -- -D warnings
cargo test --manifest-path libs/gaia/grib-codec/Cargo.toml --all-targets  # 7 passed
cargo build --release --manifest-path libs/gaia/grib-codec/Cargo.toml

In total, 185 Phase 175 automated tests passed (149 Python, 29 TypeScript, and 7 Rust), with no failures or ignored cases.

Verification ran on a 15 GiB host with 31 GiB swap and about 12 GiB available memory. A BFF Nx development watcher and other agents were active, so no broad workspace sweep was run. These changes are headless Python ingestion and model orchestration; they do not alter web, PWA, mobile, Unreal, or other rendered UI, so Playwright, mobile automation, and Unreal editor automation are not applicable.

Remaining external outcomes#

Tasks 175.2.2.1, 175.2.2.2, 175.3.1.3, 175.4.1.3, 175.6.1.1, 175.7.1.1, and 175.9.1.5 intentionally remain unchecked until their specified real training, benchmark, licensing, release, and publication evidence exists. There are no remaining open locally actionable Phase 175 tasks.