Solstice Pro AIObservatory · 16544720
RA 06h 45m 09s DEC −16° 42′ 58″ Epoch J2026.4 Plate VI Field Dispatches
Plate VI · Engineering desk

Dispatches from the engineering desk.

Long-form technical writing on what we have actually shipped, what broke, and what the headline metric does not tell you. Dispatches are filed only when the engineering position has survived a hostile internal read; the cadence is irregular by design.

Cadence Irregular — when the work warrants it Filed 2026 · 3 active Editorial Hostile internal review before publication
RA 06h 45m DEC −16°  ·  Field 01 — Current dispatches

What is published right now

Three dispatches are currently filed. Each one describes a single piece of work we have shipped to a partner observatory or operations centre, the metric we are willing to defend on it, and the failure modes we know about. The numbers in each are reproducible from the validation notebook that ships with the corresponding model release on the Data page.

2026-04-22 · Helios mission

M-class flare TSS 0.71 at 30 min

How a spatiotemporal CNN on SDO/AIA EUV imagery plus HMI magnetograms reached TSS 0.71 at the 30-minute horizon for M-class flares, including the calibration step that made the score usable for an operations desk and the failure modes (limb events, multi-flare days).

Open the dispatch →
2026-03-09 · Photon mission

Restoration when the data is a histogram

EMCCD electron-multiplication imaging produces data that is, in the regime that matters, a Poisson histogram rather than an image plus Gaussian noise. The dispatch walks through why classical denoisers fail and what a photon-count-aware variational network buys you (1.7× effective gain over BM3D at g = 200).

Open the dispatch →
2026-01-25 · Sky mission

A real-bogus classifier you can put in front of a broker

97.4% recall at 0.5% bogus contamination on the ZTF 2024-January replay set, plus a cosmic-ray segmenter at F1 0.962 replacing LACosmic. The dispatch is honest about the LSST-era volume considerations that will reset the benchmark.

Open the dispatch →

RA 07h 12m DEC −22°  ·  Field 02 — Editorial standard

Why the cadence is irregular

The engineering desk has a single rule about dispatches: a piece is filed when an internal hostile reviewer cannot kill it, and not before. The rule produces about three dispatches a year. We are aware that this is not how content marketing is meant to work; we are also aware that observatory partners read these pieces carefully before integrating our models into a real pipeline, and we would rather they read three honest dispatches than twelve marketing ones.

Each dispatch carries the filing date, the engineering author byline, and a numeric reference to the validation notebook used to produce the headline figure. Errata, when we find them, are appended to the end of the original dispatch with the date and the correction; we do not silently re-edit numbers.

RA 07h 38m DEC −25°  ·  Field 03 — Archive policy

What about the older work

Work that pre-dates the current dispatch format lives in the engineering notebooks shared directly with partner observatories. Where a finding has held up under continued use, it is rewritten into a current-format dispatch; where it has not, it is archived without ceremony. We do not maintain a public archive of the older pieces because they were written for a specific operational audience and would mislead a general reader.

The full citation pattern for referencing a dispatch in an academic paper is given on the Data page along with each model release.

If a dispatch is relevant to your pipeline

Each dispatch ends with a transit notice for that specific piece of work. The generic path — academic collaboration, operations integration, or instrument-scientist support — is on the Access page; the fastest reply route is the contact form below.

Open contact →