# Geo Desk

> Paste the layer manifest, the validation confusion matrix and the methods text of one finished
> land-cover or earth-observation mapping study, and find out whether its accuracy claim is
> defensible. Twenty-two named checks run free in the browser; three metered lanes then work the
> study.

URL: https://geo-desk.skillsafe.ai/
API: https://geo-desk.skillsafe.ai/api.html
Model: gpt-terra (resolves to gpt-5.6-terra) · publisher markup 1000 bps · price_credits 0
Sources: a derived work of two SkillSafe skills, both credited —
  @k-dense-ai/geopandas (coordinate reference systems and their units, reprojection, geometry,
  overlay, extent intersection, per-class area arithmetic) and
  @k-dense-ai/geomaster (raster grids, resampling, remote sensing, spatial statistics, spatial
  machine learning, validation design, accuracy assessment).

## What problem it solves

A published map accuracy figure is an estimate produced by a sampling design, and most of the ways
it goes wrong are invisible in the number itself. The reference sample was split at random over
pixels drawn from inside single training polygons, so the folds share neighbours. The overall
accuracy was never weighted by the area each class actually covers, so it describes the sample
rather than the map. A predictor was upsampled from 500 m onto a 100 m grid, so the model saw blocks
of identical values. A DEM in EPSG:4326 has 20 m east-west and 31 m north-south ground pixels at
49 degrees north, so every slope taken from it is stretched. Geo Desk finds these before a reviewer
does.

## The free read: 22 named checks, no account, no model

Each check reports `pass`, `attention`, `fail` or `not_assessable`. A check whose input is missing
reports `not_assessable` and is never counted as a pass; the header always reads "N of 22 ran".

Coordinate systems and the grid — every layer declares a CRS; the layers share one; each pixel size
is plausible in its own CRS's units (a degree-sized pixel in a metric system is blocking, and so is
the reverse); same-resolution rasters sit on a common pixel grid, with any offset reported as a
fraction of a pixel; no layer is upsampled onto a finer grid than it has; every layer covers the
mapped area, by real extent intersection, refused across differing coordinate systems rather than
guessed; every raster declares its nodata value, with an absent key distinguished from a declared
absence.

The matrix — it agrees with itself, and its orientation is declared; the reference layer's feature
count matches the matrix total; every class carries enough reference units, in three tiers, with an
empty class separated from a merely thin one and a rare class judged against its own area weight.

The design — the sampling design is stated; it supports the inference the numbers are used for (a
purposive or reused-training sample supports none, by any estimator); the validation folds are
spatially separated; the spatial block is at least as large as the stated autocorrelation range;
training and validation units are independent of each other.

The numbers — the reported overall accuracy and kappa are recomputed from the matrix's own margins
and compared with what the text claims; overall accuracy is compared with the no-information rate,
so a high figure on an unbalanced reference sample is recognised as a property of the sample; the
area-weighted (Olofsson et al. 2014) estimator is computed where the mapped areas are supplied,
with the weighted overall and producer's accuracies, the error-adjusted areas and 95% intervals per
class; an accuracy reported without an interval is flagged, with the interval supplied; and the
reported metrics are matched against the kind of product, so kappa on a continuous surface is
caught.

## Three lanes over one work object

- `align` — Layer and grid conformance. Coordinate systems, the pixel grid, coverage, what to
  resample and in which direction, nodata, and whether the stack is ready to extract from.
- `validate` — Validation design and leakage. The sampling design, spatial separation of the folds,
  allocation across classes, independence of the units, and what to change in order.
- `report` — The area-weighted accuracy pack. The estimator the design requires, the defensible
  accuracy statement, the error-adjusted areas, whether the metrics fit the product, the
  limitations, and a drafted methods-and-accuracy paragraph.

Every lane reads the same three pasted blocks. Each lane's verdict follows only the areas that lane
owns; a serious fact from another lane that the answer depends on is surfaced as a binding caveat
rather than silently moving the verdict, and the whole study's verdict is shown separately in the
free read.

## Input and output contract

The run input object IS the POST body and is never wrapped in an `input` key. Fields: `task`
(`align`, `validate` or `report`), `layers`, `matrix`, `methods`, `product`, `design`, `claim`,
`context`, `prescan_facts`. The `product` and `design` selectors only ever fill a silence in the
methods text; they never override what the prose says, and which source was used is recorded.

The reply is one JSON object: `lane`, `title`, `verdict` (`defensible`,
`defensible_with_changes`, `needs_rework`, `not_defensible`), `headline`, `summary`, `body`,
`findings[]`, `reconciliation[]`, `context_notes[]`, `unassessable[]`. Every free-read flag is
reconciled one-for-one by uid, and a flag the reply never addresses is displayed as unanswered.

## What it deliberately does not do

It never reprojects, so two extents in different coordinate systems are reported as not comparable
rather than compared. It does not know the true areas: the mapped areas are whatever was pasted, and
the error-adjusted areas inherit any error in them. The ground size of a pixel in a geographic CRS
is a spherical approximation at the layer's mid-latitude. The unweighted interval assumes simple
random sampling. It cannot tell whether a reference label is correct, so reference error is
invisible to it. A ratio with no denominator is reported as undefined, never as zero.

## Free, in the browser, before sign-in

The whole read, both worked examples, all six saved model runs (one per lane per example), and every
export — Markdown, JSON, and CSV of the checks, the flags, the per-class accuracy table, the layer
table and the confusion matrix. Nothing you paste leaves the browser until you run a metered lane.