Version 1 record

This page describes software that has been retired. It is accurate about Version 1 and is not a description of what runs now — see the current site and the archive.

C-13 Model runner

Status: built

Code: services/model_runner/ — kernel.py is the port, analytic_kernel.py the implementation behind it, and ensemble.py the perturbed members and their spread
Delivered by: specs/009-control-loop
Covered by: services/model_runner/tests/, including test_kernel_port.py for the boundary and test_member_failure.py for a member that dies, plus tests/integration/test_runner_publisher_handoff.py

Responsibility: analytic advection and noise; ensemble member runs. Owns the failure mode of: being irreplaceable.

What it does

It takes an initialisation state and produces a gridded forecast field. It does this by advecting the seeded features forward analytically and adding noise. It runs a small ensemble with perturbed initial conditions and emits the ensemble spread as an uncertainty field alongside the forecast.

It does not implement real numerics, and it is not trying to. The numbers are deliberately fake.

One property of that spread is worth stating here rather than leaving a reader to infer it: it does not grow with the forecast horizon. Measured over the runner's own fixture, the domain-mean spread is 0.221, 0.209 and 0.199 °C at t+0 for 4, 8 and 200 members, and 0.220, 0.210 and 0.199 °C at t+6h for the same three. There are no dynamics here to amplify an initial-condition error, and no growth law is authored in their place. FR-29 asks for a spread emitted alongside the forecast, which this is; it does not ask for one that widens, and this one does not. The derivation says where the number comes from.

Why "being irreplaceable" is the failure mode

This is the one component in drogna that is certain to be thrown away. Any real version of this architecture would put an actual ocean model here. So the whole value of building it is in the shape of the hole it leaves behind: initialisation state in, gridded field out, and nothing else crossing the boundary.

A model runner that quietly acquired a dependency on the observation store, or that wrote its output somewhere only it knew about, or that needed to be told about the scenario in a way the interface did not express, would be a component that could not be swapped. That is what the failure mode names. The interface here is one of the four genuine ports, and it is genuine precisely because the second implementation is not hypothetical.

The ensemble

The ensemble is small — this is a harness, not a forecasting centre — and its purpose is not accuracy. It is to produce an uncertainty field that behaves like a real one: larger where the initial state was less constrained, larger further into the forecast, and structured rather than uniform.

Requirements: FR-28, FR-29. Feature: 009.