version: "1.0.0" name: load-smap description: "Load SMAP L3 soil moisture with the volume gate: product line and version parsed, retrieval-quality flags decoded, applicable concepts consulted."

load-smap

Bring SMAP L3 into the session safely: product line resolved, gated on volume, quality flags decoded, applicable concepts consulted. Works by slash command or conversationally ("load SMAP soil moisture for the Southern Great Plains, 2016").

Behavior, in order

1. Parse and show back: product line, version, region, window, and

AM/PM overpass choice. The model-assimilated L4 family is offered only on explicit request and labeled as model output; which L3 product lines exist and how they differ is read from the bundle (step 2), not listed here.

1. Consult the bundle for this dataset first. Consult installed

knowledge concepts first, as the core consult-knowledge skill sets out, by product line, region, and window; the smap skill lists the concepts this plugin resolves to. Read the dataset concept for the product lines, resolutions, ~5 cm depth semantics, and flag meanings, and the gotchas the window or intent triggers (a window touching 2015 pulls in the radar-loss lineage break). Restate what each changes about the plan and cite it by path; do not carry these facts in this skill. A concept added since you last ran is found this way.

1. Search before fetching: granule count and estimated volume BEFORE

any download.

1. The volume gate. Threshold from the project local config. At or

below: state count, size, destination, proceed. Above: STOP, present count, total size, destination, and a smaller alternative (the coarser-posting product, a shorter window, fewer granules), and wait for explicit confirmation.

1. Load with flags decoded: apply the retrieval-quality gate the

dataset concept defines and report it (gated fraction and the dominant flag reasons); keep AM and PM retrievals distinct.

1. Summary as provenance: product line and version, granule count,

window, the quality accounting, the depth semantics stated (read from the concept), cache location, and the concepts consulted.

Must NOT

• Never fetch above the gate threshold without explicit confirmation, on

any surface. (Hard gate: fires without consulting anything.)

• Never merge the model-assimilated L4 family with observational L3

retrievals, or present the one as the other. (Hard refusal: the observation-vs-model epistemic class is universal; which SMAP line is which is read from the dataset concept.)

• Never average AM and PM retrievals into one field silently. (Load

procedure: they are distinct overpass populations.)

• Never present retrieval statistics without the quality gate applied

and reported. (Load procedure; the flag semantics are the concept's.)

• Never restate a dataset rule this skill could consult: the L3 product

lines and resolutions, the ~5 cm depth semantics, the flag meanings, and the 2015 radar-loss lineage break live in the SMAP concepts (datasets/smap-l3.md and gotchas/smap-radar-loss.md) and are read from them per load. Consulting them is how a corrected or new concept changes this skill without editing it.