analysis-review

The post-computation review: invocable directly ("review this analysis") and run before results are reported. Review what was computed (the code and outputs), not what the prose claims.

Consult the bundle (standing step)

This review carries NO expected numbers, dataset facts, or gotcha rules of its own. Before flagging anything numeric or product-specific, consult installed knowledge concepts first, as the consult-knowledge skill sets out, and cite what each match constrains. In particular: the cross-cutting sanity anchors live in conventions/smell-test-ranges.md; a product's native uncertainty or quality layer lives in its dataset concept's ## Uncertainty section; a quantity's expected range lives in its domain recipe concept and the computation that recipe cites; a trap's rule lives in its gotcha or convention concept. Never work from a remembered number where a concept exists.

Verdict format

Each check gets an inline flag and one line of evidence:

• 🔴 wrong result likely; must be fixed before reporting
• 🟡 suspicious or incomplete; investigate or justify explicitly
• 🟢 clean

The review ends with an overall verdict; any unresolved 🔴 blocks the report. The review proposes fixes; it never silently applies them.

Methodology checklist

1. Area weighting. Any spatial mean over a lat-lon grid without cos-lat or cell-area weights is 🔴 (xarray-fundamentals has the rule and the fix).
2. Autocorrelation in trends. Original Mann-Kendall or uncorrected OLS significance on monthly or daily series is 🔴; check the method against basic-statistics' decision tree, and per-cell map claims against its FDR rule.
3. Baselines. Climatology, anomaly, or percentile results without a stated baseline are 🔴; two compared series on different baselines without comment is 🔴; a trended record used as its own full-record baseline earns 🟡 unless the midpoint-centering is acknowledged.
4. Projections and figures. Per cartography: jet or rainbow 🔴; diverging colormap without a pinned physical center 🔴; missing transform= 🔴 (misplaced data); polar fields on PlateCarree 🟡; stippling, hatching, or bands undefined in the caption 🟡; comparable panels without a shared colorbar 🟡.
5. Budget grid (hard refusal: invariant, universal). Property budgets computed on regridded fields are 🔴, full stop; budgets close only on the native grid. This flag fires without consulting anything; the domain plugin's budget skills and gotcha concepts carry the formulation details, cited when present.
6. Calendar and seasonal handling. DJF spanning the year boundary via groupby("time.season") is 🔴 for cross-year seasons; annual means of monthly data without month-length weighting 🟡. The rule and the fix live in the calendars convention concept (conventions/calendars.md); read and cite it, do not restate it.
7. QC lineage. Statistics computed before a fill-value audit are 🔴 if sentinels were present; no QC run at all on a new dataset is 🟡. The sentinel list and detection recipe live in the fill-values concept (conventions/common-fill-values.md), consulted per the standing step, not carried here.

The three UQ checks

1. Is uncertainty reported alongside the headline number? Every headline quantity carries an interval, a spread, or a native product uncertainty, or an explicit one-line waiver. Absent: 🔴.
2. Is an ensemble mean shown without spread? 🔴; with fewer than about 10 members the spread must also be called a likely underestimate (uncertainty-quantification has the caveat).
3. Does the product carry a native uncertainty or quality layer that the analysis ignored? Consult the dataset concept's Uncertainty section to know what exists. Ignored and material to the conclusion: 🔴; ignored and plausibly immaterial: 🟡 with a justification asked.

Smell-test ranges

The procedure: a headline value outside its anchor is 🔴 until explained, inside but odd is 🟡. Do NOT carry or remember the anchor values here. Read the cross-cutting anchors from the smell-test concept (conventions/smell-test-ranges.md) per the standing step, and read a domain quantity's expected range from its own domain recipe concept (never hardcode a number that a recipe or the anchors concept states, and never invent one when a concept exists). Restate and cite the anchor you used.

Unit smells resolve most out-of-anchor values (invariant unit method, not a dataset fact): a 273-ish offset is K vs C; a factor ~30 on precipitation is mm/day vs mm/month; a factor 100 on fractions is percent vs fraction; a factor 1000 on transport is TW vs PW.

Must NOT (hard refusals and gates)

• Never pass an analysis with an unresolved 🔴.
• Never soften a missing-uncertainty finding to stylistic advice; it is 🔴 by the house rule.
• Never invent an expected range, or remember one, when a concept states it; read it from the recipe or the smell-test anchors concept and cite it.
• Never review prose alone when code and outputs are available.
• Never silently fix what the review found; propose, with the flag and the evidence.