quality-control

Quality control for geophysical datasets: invocable directly ("run QC on this dataset") and consulted automatically before analysis. The six-check structure reconstructs the v0.1 enumeration that the specification's acceptance criteria reference (docs/SPECIFICATION.md in open-science-pillars/marketplace).

Consult the bundle for this dataset

Before reporting any check, consult installed knowledge concepts first, as the consult-knowledge skill sets out, by the dataset, variable, and check in play; this skill carries no dataset facts, sentinel lists, screening numbers, or QA bit layouts. The concepts it leans on:

• the sentinel list and detection recipe: conventions/common-fill-values.md;
• the physical-bounds screening table: conventions/physical-bounds-screening.md;
• the satellite QA bit and class layouts: conventions/satellite-qa-flag-decoding.md;
• calendar and time-axis rules: conventions/calendars.md;
• the dataset's own known-issues / gotcha concepts, for named instrument or version transitions.

Numbers, sentinel values, and bit layouts are read from the concepts, never from this file.

The six checks, in order

Order matters: a bounds check on unmasked sentinels reports garbage, so the fill audit precedes bounds.

1. Coordinates and time axis. Latitude within [-90, 90], longitude convention identified, coordinates monotonic after sorting; time axis monotonic, no duplicates, cadence as documented; calendar identified.
2. Fill-value audit. _FillValue/missing_value attributes present and applied; scan minima/maxima for unmasked sentinels and check encoding for packing applied twice or not at all. The sentinel list and the detection recipe are read from conventions/common-fill-values.md, not carried here.
3. Completeness. Missing fraction per variable, overall and per time step; the longest gap, reported in time units, with its dates; spatial holes distinguished from temporal gaps.
4. Physical bounds. Every value inside the hard bounds of its variable; values outside the plausible-but-legal range listed for review rather than auto-rejected. Hard versus plausible is the method; the screening numbers per variable are read from conventions/physical-bounds-screening.md, and regional analyses tighten them. State which bounds were applied.
5. QA flag decoding. Satellite products: decode the bit-packed QA layer with bitwise operations and mask accordingly; never threshold a packed QA integer numerically. The product bit and class layouts (MODIS, Landsat C2, Sentinel-2, and kin) are read from conventions/satellite-qa-flag-decoding.md, not carried here; state which bits or classes were masked and what fraction that removed.
6. Discontinuity detection. Spikes (rolling-median deviation test), steps (change in mean across a break, instrument or version transitions), and suspicious exact repeats (stuck sensor, duplicated granule).

Report each check with a verdict and one line of evidence. Flagged data is masked and reported, never silently deleted; interpolation across a reported gap happens only when stated in the methods.

QA flags are categorical: they are never a quantitative uncertainty (uncertainty-quantification carries that rule). The screening numbers (the physical-bounds table) and the product QA layouts (MODIS, Landsat C2, Sentinel-2) are dataset facts and live in the concepts named in "Consult the bundle for this dataset", read and cited per run, never inlined here.

Discontinuity detection

• Spikes: deviation from a centered rolling median beyond k robust standard deviations (k around 5 for geophysical series); report count and worst offenders with dates.
• Steps: compare means across candidate break dates (known instrument or version transitions first; the dataset concept's known-issues section names them); a persistent shift after a transition date is a finding for the provider, not something to quietly detrend.
• Exact repeats: identical consecutive values or identical granules beyond chance suggest a stuck sensor or duplicated file.

Must NOT (hard refusals and gates: invariant, universal, fire without consulting anything)

• Never run bounds checks before the fill-value audit. (Ordering gate: a bounds check on unmasked sentinels reports garbage.)
• Never threshold or average a packed QA integer as if it were data.
• Never treat QA flags as quantitative uncertainty.
• Never silently delete or interpolate flagged data; mask, report, and state any infilling in the methods.
• Never pass a dataset as "QC clean" without reporting all six checks with evidence.
• Never inline a screening bound, a sentinel value, or a QA bit layout in this skill; read them from the concepts and cite them per run.