Metforge Analysis

MetForge Analysis

Turn atmospheric data into a traceable scientific diagnostic. Make the mathematical definition, sampling assumptions, units, execution mode, and validation visible.

Analysis contract

Before computing, establish:

• scientific quantity or hypothesis
• exact formula, published definition, or package function
• variables, units, coordinates, domain, levels, period, season, and baseline
• weighting, anomaly, standardization, missing-data, and event-selection rules
• output variables, dimensions, units, metadata, and acceptance checks

Do not silently choose among competing index definitions. For a named index, recover the authoritative definition and identify any required reference EOFs, climatology, or external coefficients.

Computation workflow

1. Inspect files, coordinates, units, calendar, chunking, missing values, and existing project code.
2. Choose an established implementation before writing formulas:

• use MetPy with unit-aware inputs for supported meteorological diagnostics
• use xarray for labeled reductions, grouping, rolling, resampling, and alignment
• use SciPy/statsmodels for filters and statistical models
• use xclim for supported climate indices and xeofs/eofs for EOF analysis
• use xskillscore or established statistical libraries for supported metrics

1. Read [references/indices-and-diagnostics.md](references/indices-and-diagnostics.md), [references/trends-and-filtering.md](references/trends-and-filtering.md), or [references/statistical-methods.md](references/statistical-methods.md) as the task requires.
2. Validate the definition on a small subset or synthetic case with a known result.
3. Select direct execution or a scheduler using the rules below, then run the smallest valid calculation before scaling out.
4. Validate dimensions, coordinates, units, numeric ranges, sample counts, missing-data propagation, and at least one independent or limiting-case check.
5. Save self-describing outputs and append an execution record after each completed computation stage.

Scientific safeguards

• Preserve units and coordinate metadata; quantify any loss during format conversion.
• Apply area or mass weighting where the diagnostic requires it.
• Keep preprocessing order explicit. Detrending, anomaly removal, filtering, standardization, and compositing are not interchangeable.
• Do not concatenate separated seasons and then apply a continuous time filter.
• Distinguish regression maps from correlation maps and regression coefficients from composite anomalies.
• Account for serial dependence, multiple testing, and event overlap when they affect inference.
• Treat EOF sign as arbitrary and document weighting, normalization, missing mask, training period, and projection method.
• Never replace a package function with a hand-written approximation only to avoid understanding its assumptions.

Execution mode and processing record

Before running code, inspect the current environment: available CPUs and memory, data volume and chunking, expected runtime, whether sbatch/Slurm is available, whether the session is already inside a Slurm job, and existing project conventions.

• Run directly for metadata inspection, subset validation, interactive exploration, and calculations that fit safely within the current session.
• Use Slurm or the site's scheduler for expensive multi-file calculations, parameter sweeps, memory-heavy operations, or work that should not run on a login node.
• Do not invent account, partition, walltime, or module settings. Infer them from existing scripts/configuration or request the missing value.
• For submitted work, capture the job script, job ID, resources, environment activation, stdout/stderr paths, and terminal status.

After every concrete compute call or completed batch stage, append a record to the project's existing log/provenance location. If none exists, use output/logs/YYYYMMDD-HHMM-.md. Record inputs, software environment, direct/Slurm decision, commands or script, parameters, outputs, validation results, deviations, and status. Never record credentials or secrets. If files cannot be written, return the same record explicitly in the response.

Handoff

Return the calculation definition, code or script, outputs, validation evidence, statistical caveats, and the processing record. Use metforge-figure only when a visual result is needed and metforge-model-diagnose only when the question concerns model failure or numerical behavior.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yuanruichen
• Source: yuanruichen/MetForge
• License: MIT

Install and usage instructions live in the source repository linked above.