One-year evaluation output check

Use a one-year cold-start run to establish what the selected executable actually writes. A successful process exit or a variable name in source code is insufficient.

Select targets from observations

1. Inventory actual nonmissing observations for the requested site, including QC, dates, units, depths, organs and uncertainty. Use the BIF timeseries skill for irregular field records and inspect tower values, not just CSV headers. Distinguish observations, partition-derived GPP/respiration, forcing inputs, calibration data and unavailable series. Do not invent grain yield from a Total or Foliage production record.
2. Choose one calendar year with required climate and management coverage, preferably overlapping useful observations. Record why it was selected. This cold-start year tests output plumbing; it is not an equilibrated prediction.
3. Write a manifest using references/manifest.md and the mappings in references/target_mapping.md. Include every relevant observation target, marking unresolved mappings blocked with the reason instead of dropping them. Trace units, sign, cumulative/instantaneous semantics, depth support and tissue pools through the active history registration and update code. Do not assume that defaults include requested fields or that relative saturation equals volumetric water.

Prepare, run and verify

Run from the ecosim-agent root (Python, NumPy and netCDF4 required):

python .agents/skills/ecosim-evaluation-output-check/scripts/check_outputs.py prepare \
  --namelist result/SITE/run_dir/SITE.namelist --manifest /path/to/manifest.json \
  --scratch result/SITE/output_check/unique-run
python .agents/skills/ecosim-evaluation-output-check/scripts/check_outputs.py run \
  --scratch result/SITE/output_check/unique-run --exe /path/to/ecosim.f90.x

prepare copies selected inputs into a new scratch case, requests the union of existing history selections and required manifest fields, fixes daily h0 output, and sets one selected forcing/model year, fresh initialization and annual stop. The source case and existing runs/restarts stay untouched. It checks registry names and year axes; the full climate-readiness checker runs before execution. Required fields registered with default='inactive' are explicitly enabled via hist_fincl1; do not change their Fortran registration defaults. Required fields are removed from conflicting hist_fexcl1 entries in the scratch namelist, while unrelated selections/exclusions remain. plan.json and report.json record these changes. The subsequent NetCDF check must confirm the fields were actually emitted; adding a name alone does not establish successful output. Unsupported namelist syntax, coupled/daily-only forcing or restart initialization must be resolved explicitly rather than silently approximated.

Before launch, run enforces the shared case preflight on both the source production request and the one-year scratch case. A missing later production year blocks even if the selected short year is complete. Source namelist changes require a new prepared case. Unsupported model namelist keys are errors even when another model version or a template supports them.

Select/build the intended executable using repository AGENTS.md toolchain instructions. run rejects an executable older than selected Fortran sources; --allow-stale-executable is an explicit diagnostic exception, retained in the report, not evidence that the source and executable agree. The default wall-time limit is two hours; adjust --timeout-seconds for the selected case. A failure or timeout is reported without automatic parameter changes or continuation. Use a new scratch path for each retry; preserve previous reports/logs.

To recheck already generated scratch output without rerunning EcoSIM:

python .agents/skills/ecosim-evaluation-output-check/scripts/check_outputs.py check \
  --scratch result/SITE/output_check/unique-run

Review report.json and report.md. The checker requires the expected number of hourly steps, full-year daily history bounds, each required field in every h0 file, compatible metadata, valid dimensions, and finite non-fill data in all records on at least one consistent cell/PFT/layer. It scans all written numeric history fields for unmasked NaN/Inf. Explicit manifest checks can require nonzero values or variation when physically appropriate. Inactive/fill-only PFT slots are not themselves errors; all-fill target output is an error.

The current history uses day-offset time_bounds without units. The checker accepts exactly contiguous [0,1] ... [days-1,days] only for its controlled daily scratch namelist, reporting that inferred convention. Other bounds must be handled explicitly; the filename alone does not prove year coverage.

Resolve gaps and preserve the outcome

• A requested field absent from the registry needs a history mapping or model instrumentation, not a guessed alias. Follow model-repository instructions for any authorized registration/allocation/update work, build and rerun the check.
• Missing output, units, finite data or depth/pool support prevents an overall pass. Target-specific blocked issues yield an incomplete result even if the other history fields pass. Preserve observation-unit/depth uncertainty in the manifest; the script cannot establish scientific equivalence from strings.
• Refresh inherited manifests against current registry/output and observation metadata before reuse. A previously blocked net-radiation mapping or unknown sensor depth may already have been resolved in a later site evaluation. Record the evidence for each revision; do not repeat stale missing-data claims or mark a conditional conversion as exact. Report process/history success, mapped-target passes and unresolved scientific mappings separately: an overall incomplete report is not necessarily an execution failure.
• VSPore_vr * rWatFLP_vr made from separately averaged daily fields is not necessarily mean volumetric water content. Prefer a directly accumulated volumetric-water field or demonstrate the needed instantaneous/time-weighted conversion and pore basis. Do not hide this gap by requesting two names only.
• Once coverage is verified, carry the reviewed history selections into the production namelist within the requested preparation scope, preserving its dates, forcing cycles, restart choices and other settings. Save the exact diff and input/executable identities; a scratch pass alone does not update production output controls. Rerun the relevant check if the mapping changes. Include all required inactive-by-default fields in production hist_fincl1 as well, and reconcile their hist_fexcl1 entries. Preserve reviewed :A/:I averaging qualifiers; a conflicting existing qualifier must be resolved using the manifest's temporal semantics, rather than appending a duplicate name.
• Record this as an evaluation-output check in the existing site tracker/guide; do not mark the two-year plant-trait smoke test, full simulation or scientific evaluation complete. Explain unsupported targets and what remains to enable.

Validation

python -m unittest discover -s .agents/skills/ecosim-evaluation-output-check/tests -v