Output data

A project is successful only after ensemble execution, assimilation, compact grid export, benchmarking and every configured plot, map and report pass their checks. The atomic run manifest is the authoritative completion record.

Run manifest

results/run_manifest.json records configuration and input hashes, the openAMUNDSEN-DA software version, stage/member state, output inventory and cleanup results. The optional image and image_digest fields are populated only when the runtime supplies OPENAMUNDSEN_DA_IMAGE and OPENAMUNDSEN_DA_IMAGE_DIGEST; an empty value does not invalidate a run. Interrupted or failed runs keep restart state. A mismatched configuration or input inventory cannot resume an existing run.

Compact gridded result

Every successful data assimilation project writes:

results/grids/da_output_grids.nc

This compressed NetCDF contains the configured scientific variables and metrics, including the open loop, ensemble statistics and event-weighted analysis fields. Typical names are open_loop_<var>, ens_mean_<var>, ens_std_<var>, increment_<var>, analysis_mean_<var> and analysis_increment_<var>.

When compact variables are explicitly configured, output completeness is a hard contract. The setup must configure the matching openAMUNDSEN grid output name before propagation, every open-loop and member NetCDF must contain that source and the compact file must contain every requested metric-variable pair. The same validation is repeated for subdomain leaf files and their merged NetCDF, so a consistently incomplete set of leaves cannot be accepted.

The model propagation grids feeding this builder can be NetCDF or GeoTIFF as selected in the setup YAML. The compact data assimilation result itself is always NetCDF.

Compact projects also retain two compressed, all-member time-series files:

results/points/ensemble_points.nc
results/forcing/ensemble_forcing.nc

Their dimensions are time, member, point and time, member, station, respectively. They contain the open loop and every ensemble member, so standard Python and R NetCDF tooling can plot or analyze the retained time series after member CSV cleanup. Projects with satellite observation events also retain results/grids/da_map_support.nc, which contains the event fields needed by the configured map renderer.

Weighted propagation means, standard deviations and increments use the persistent PF prior ledger for each step. Configured ensemble minima and maxima describe the unweighted member collection; compact grid quantiles are not currently generated. Event-analysis fields use the normalized pre-resampling posterior weights. Unweighted member traces and extrema describe the materialized member collection and are not interchangeable with the weighted analysis.

Particle-weight and stochastic provenance

Every step stores assim/prior_weights.csv with member_id, log_weight and weight, plus prior_weights_manifest.json. Event tables add prior_log_weight, prior_weight, log_likelihood, posterior log_weight and weight, prior/posterior ESS, threshold and resampled. Paired manifests hash the configuration, observation/model inputs and ledger ancestry.

Resampling indices and manifests distinguish a mirrored posterior from an actual systematic resample. rejuvenate_manifest.json records the keyed-v1 RNG contract, configured seed and event-specific perturbations. Incomplete old chains without these contracts remain readable but cannot resume under the corrected method.

Tables and diagnostics

The standard result tree includes:

results/
  benchmark/
  grids/da_output_grids.nc
  maps/
  misc/
  plots/assim/
  plots/perf/
  plots/points/
  plots/results/
  reports/project_report.pdf

Benchmark tables and their manifest live under results/benchmark/. Weight, effective sample size and score plots live under results/plots/assim/. results/plots/perf/ is refreshed during every project run. The CSV retains raw filesystem telemetry, while the plot focuses on CPU, RAM, project-directory size and optional CPU temperature. For successful single-domain runs, exact project-size samples immediately before and after restart-state cleanup make the storage reduction visible. The reviewed Rofental walkthrough explains the files in Results and Diagnostics.

Cleanup and retention

After all required outputs validate, retention: compact removes package-owned restart checkpoints, member point and forcing CSVs, replaceable member grids and step-local forcing PNGs after their configured render completes. The compact forcing NetCDF remains the source for rerendering project-wide forcing plots. Fresh compact projects place those disposable files below .openamundsen-da/runtime/<generation>/; durable weights, member and PF manifests, logs, benchmarks, compact NetCDFs, maps and reports stay outside. After consumer and render validation, retention-ledger schema v6 records the generation root identity, producer authority, retained-consumer hashes and cleanup state without enumerating the raw tree. Cleanup pins and validates the retained consumers, atomically renames the generation to the same-filesystem .openamundsen-da/quarantine/ tree and physically deletes it with bounded directory parallelism. The generation is complete only after neither live nor quarantined bytes remain and that fact is durably recorded. A crash before or after rename, during deletion or before final ledger completion resumes the same generation safely. Disposable raw files are not rehashed after compact consumer validation.

Existing projects with member-local raw artifacts retain the schema-v5 batch contract: every planned source path is bound to its retained consumer and producer evidence, and interrupted batches resume only when those identities still match. Completed v5 ledgers remain readable; an interrupted v5 cleanup is never silently migrated into a v6 tree. The retained NetCDFs, weights, benchmarks, plots, maps, reports, logs and scientific configuration remain.

retention: full skips this artifact cleanup and preserves the raw member forcing, points, grids and restart states for reanalysis.

For an older single-domain project, preview and then apply the same safe cleanup:

openamundsen-da clean PROJECT_DIR
openamundsen-da clean PROJECT_DIR --apply

Single-domain and subdomain leaf projects use the same cleanup lifecycle. A successful compact subdomain leaf is finalized after its own benchmark, configured render outputs and report validate. Its raw member forcing, point, grid and final restart artifacts are then removed immediately, while the leaf compact grid, point/forcing stores, DA map support, weights and report remain. For fresh projects, step-local forcing PNGs and all other disposable member outputs are removed with the generation tree after stable render-completion evidence exists. The evidence and producing member manifests are bound into the cleanup ledger; the report may subsequently be refreshed with the final performance snapshot without invalidating cleanup provenance. leaf_finalization_manifest.json binds those retained inputs before the next leaf wave is admitted. An interrupted or low-disk run resumes existing work non-destructively; rebuilding requires an explicit overwrite request.

Every executing project also retains results/storage/storage_reservation.json. This operational audit artifact records the immutable conservative plan, lifecycle phase, per-leaf admission progress, upward-only observed byte high-water marks, latest filesystem snapshot, source-catalog read counters, materialized/removed byte counters and estimate/check timing. It does not replace scientific provenance or the retention ledger. Normal step boundaries update this compact record from producer summaries and call the filesystem usage check once; they do not rescan configs, source data, accumulated artifacts or sibling project trees. Ledger serialization scales with the immutable prepared leaves and steps, not with materialized data-tree size. Resume reconciles the plan and authoritative workflow manifests before admitting new work. Explicit overwrite archives the superseded storage ledger generation beside the active ledger. Spawned leaf workers exchange bounded atomic request/response manifests under results/storage/ipc/<generation>/. Normal shutdown removes an empty IPC generation; abrupt process death retains pending transport evidence for retry. These are operational coordination files, not scientific output. Lifecycle request IDs remain deterministic while fresh transport nonces prevent orphaned responses from satisfying a new filesystem check.

Performance monitoring uses the ledger counters after one baseline directory reconciliation, so the five-second CPU/RAM/temperature sampler never walks the growing project tree. project_perf_phases.csv and project_perf_phases.png retain the storage-ledger phase sequence and mark a restart gap as unmonitored_downtime. The terminal performance snapshot is the next allowed full project-size reconciliation.

The current compact lifecycle is deliberately boundary-based. It shortens every generated forcing copy to its consuming step and incrementally removes obsolete restart checkpoints, but keeps a leaf’s forcing/point CSVs and raw grids until that leaf’s final compaction and rendering succeed. Subdomain admission uses bounded waves: current filesystem usage already includes measured retained compact leaves, while the additional reserve covers only the active wave, rolling checkpoints, the projected compact outputs of queued leaves and unfinished parent merge/render output. Active model members are not killed mid-propagation. Per-step compact fragments and cooperative mid-member disk stops remain future work. This is a safe increment, and its audited prepared-Euregio estimate fits a clean 3.6 TB filesystem. That estimate is not a substitute for the first complete production-run acceptance, and unrelated resident runs must be archived before relying on the envelope.

Compact point and forcing stores collapse overlapping timestamps from adjacent steps by their numeric mean, matching the established raw-series readers. Cleanup validates every retained value against that collapsed raw source before deletion. Subdomain cleanup retains each leaf’s da_output_grids.nc, so leaf snow-depth and SWE maps remain rerenderable without raw member grids. Compact grid, point, forcing and map-support writes use validated, filesystem-synchronized same-directory temporaries. Raw grid cleanup additionally requires complete configured metrics and member sources, while satellite map support is checked against its ROI, probability domain and raw source values. Before a generation is marked complete, every active batch’s source inventory, removed-path state, retained consumers and producer evidence are revalidated. The supported single-domain run API repeats complete-ledger validation after the final performance/report refresh and refuses to publish a successful run manifest when retention evidence is incomplete or invalid. Full retention keeps raw member grids and forcing/point files. SCF and wet-snow projects in this mode must successfully rebuild their configured event fields from that raw render support and do not require the compact-only da_map_support.nc archive.

Subdomain outputs

Data assimilation subdomains merge into the same compact NetCDF and project-level summary/report tree. Plain-model subdomains merge the exact configured NetCDF or GeoTIFF model grids. Both modes use hard mosaics: no interpolation, blending or particle exchange occurs across boundaries.