Three doors into a working pipeline.

Door I — Academic collaboration

A research group bringing a science question — a flare lead-time benchmark, a survey-cadence trade-off, a denoising baseline that needs honest competition — engages us as a quiet methods collaborator. We supply the model artefact, the validation notebook, and a careful baseline that we will defend in writing. We co-author when we have contributed at the methods level; we do not co-author when our contribution is operational.

Typical engagement is six to twelve weeks. Outputs are versioned releases on the Data page, tagged by paper. We do not embargo our own follow-up work behind your publication; we will, however, hold release of any joint figure to the publication date.

Door II — Operations integration

A space-weather operations centre, a survey alert broker, or an autonomous-observatory scheduler integrates a Solstice model into a working pipeline. We supply a deterministic CUDA build (we pin to a specific NVIDIA container revision and reproduce the binary on request), the validation notebook for the model's headline metric, and a small on-call window during the first thirty days of production traffic.

The integration target is a Docker image you run on your own infrastructure. We do not host the model; the engineering desk is too small to be a managed-service provider, and observatory partners have always preferred to keep inference on the same iron as the rest of their pipeline. Latency budgets are negotiated up front and written into the runbook.

Door III — Instrument-scientist support

A specific instrument scientist needs a specific by-hand step retired: a cosmic-ray cleanup that LACosmic has stopped winning at the cadence the instrument now operates at, a real-bogus filter that the pipeline's current heuristic gets wrong on the new EMCCD detector, an EUV exposure denoiser that is failing on the post-upgrade gain regime. Door III is the one-off door, sized to the problem rather than the calendar.

We scope these in a one-page brief, agree the metric we will defend, deliver the artefact with a notebook the scientist can run themselves, and step away. We do not extend Door III into Door II without an explicit handshake; the engineering desk has been bitten by scope drift and we will not pretend otherwise.

The doors we keep shut

We do not build space-weather forecasting products for retail subscribers. We do not white-label models for resale by data brokers. We do not chase venture-style growth contracts that require the engineering desk to triple in size; the size of the desk is the product. We also do not take on commercial work where the headline metric was chosen by the marketing department before the engineering team had seen the data — we will tell you that politely and decline.