Guided analysis
Anyone can be a
cat modeler.
Pick a peril, describe what is at stake and where, and read the curve an open catastrophe model computes for it. Every figure comes from the run document itself, its label and basis beside it.
Backdrop: Tokyo Bay at night from the ISS. NASA, public domain.
Step 1 · Analysis
What do you want to run?
A guided catastrophe loss analysis lives here; everything beside it is built on the same platform, in its own home.
Catastrophe loss analysis
Point an open catastrophe model at your own exposure and read the losses it computes: an annual expectation, an exceedance curve, the largest events. Every figure comes from the run document itself, with the model’s label and basis beside it.
Runs in the model lab under a free sandbox key this page makes for you on the way: no form, no email, nobody in the loop.
Code a schedule
OED occupancy and construction codes for the messy rows you already have.
SOV coder →Find multi-building rows
The rows that bundle a campus into one point, found before you model.
Detector →Spreadsheet → structured file
Ingest a schedule into one documented shape, with every guess and gap recorded.
Docs →Compose your own model
Validated parts from an open registry, assembled into your own view of risk.
Docs →Accumulation across a portfolio
How much sits behind each peril, each region, each name, across every object you hold.
Docs →Pinned open data
Every number traces to a dated snapshot with its digest, and an archive of open risk datasets sits behind the same key.
The data →Step 2 · Model
Pick a peril and model
Every limit on these cards is the lab’s own answer, not this page’s.
Asking the lab what it runs…
Where these cards come from
The lab’s model listing names each model, its pinned version and its caps. This page draws one card per model the lab serves, adds the human copy it keeps for the ones it knows, and marks the ones it knows about that the lab does not serve yet as coming. A model the lab serves that this page has no copy for still gets a card: the listing decides, the copy only describes.
Raw request and response
Nothing sent yet.
Nothing back yet.
Step 3 · Exposure
Say what is exposed
Put points on the map, paste lines, send a file, or reuse an object you already stored.
Each point carries a value: what a total loss there would cost you. Print the figures in as a label for this page only. Nothing is converted: the model works in whatever amounts you enter, and answers in the same money.
Basemap: Natural Earth 110m land, public domain.
Click or tap the map to add an exposure point; every point lands in the list below with the latitude and longitude it resolved to, ready to fine-tune.
At the world view a click lands within about a third of a degree; the closer views place proportionally finer.
Each line becomes a row in the map tab's list; the two stay one set of points.
- queued
- running
- succeeded
No file sent yet.
Raw request and response
Nothing sent yet.
Nothing back yet.
Objects this browser has ingested before:
No exposure yet.
What leaves your browser
The map is a committed drawing served with this page: no tile host, no geocoder: typed search reads a list shipped in the page’s own files, and placing a point sends nothing anywhere. Points go nowhere until the run is sent; they travel inside the run’s spec. A file becomes one request to the API’s ingest verb, which answers with a job; when the job succeeds, the file is a stored exposure object and the run names its id instead of carrying the points. A stored object’s schedule is read by the lab with your own key at run time.
What the run will carry
No points yet.
Step 4 · Parameters
Set the model’s parameters
Each cap shown beside a field is the lab’s own.
Pick a model first.
The worked example
The same deterministic spec Craton’s agents page documents: two recorded hurricanes, 2004 to 2005, against three Gulf coast points. Pressing the button picks the historical tropical-cyclone model and fills the basin, the years, the storms and the three points, so the first run you send is one whose behavior is documented. Equivalent specs share one content-addressed run, so it usually answers instantly from the store.
The spec so far
No spec yet.
Step 5 · Run
Send it to the lab
Send the run and watch it move.
Ready when you are.
Why a run may take a moment
The lab executes one run at a time on a modest machine, so a busy moment means a short queue. Every run is content-addressed: the same spec is the same run id, and a spec the lab has answered before returns its stored result instantly instead of recomputing. The shareable link on the results screen is that address.
Run document
Raw request and response
Nothing sent yet.
Nothing back yet.
Step 6 · Results
The run’s own figures
Every figure below is the run document’s own, and its label and basis stand beside it.
Headline
No run yet.
Your key
This ran under your own free sandbox key. This page made it for you and this browser is holding it.
Take it anywhere the product answers: the quickstart from a terminal, the model runner in a browser, or over MCP from an agent:
Endpoint: https://mcp.oncraton.com/mcp claude mcp add --transport http craton \ https://mcp.oncraton.com/mcp
Exceedance curve
Largest events
By exposure point
Raw run document
Nothing back yet.