Tonight

I · Darkness

Astronomical darkness is a resource, and it is rationed.

The sun has to fall 18° below the horizon before the sky stops glowing. Tonight that happens at 22:11, and it un-happens at 04:12. Six hours and one minute — that is the whole budget.

Stars arrive in order of brightness, because that is the order in which a darkening sky releases them. The limiting magnitude readout is the real one.

“Twilight time never counts as imaging time.”services/scoring — observable-hours rule

II · The site

The sky you can actually use is smaller than the sky.

A horizon is not a circle. It is a roofline, a birch, a neighbour’s conifer, and a balcony rail. AstroBrain lets you draw yours — up to 720 points — and then refuses to plan through it.

It also models something no capture software can express: an overhead ceiling. My roof blocks everything above 55° to the south. The exported horizon file stays byte-identical either way.

III · The rig

Then the question narrows to glass.

A 130 mm apochromat at f/7, pulled to 728 mm by a 0.8× reducer, in front of a 26-megapixel APS-C sensor with 3.76 µm pixels. Those five numbers decide most of what follows.

Assembled from 2,831 source-backed components, then frozen as an immutable rig version — so changing your camera next month does not silently rewrite last month’s plan.

IV · The verdict

Thirteen hundred candidates. One night. Show your work.

Every other planner hands you a score. AstroBrain hands you the derivation: seven weighted components, twenty-six reason codes, and a confidence label that drops when an input is missing rather than pretending it is not.

Feasibility precedes ranking. A target that never clears your treeline is not a low score — it is not a candidate.

“Do not render black-box scores.”ADR-011 — explainability boundary

V · The distance

The light in tonight’s frames is older than the species collecting it.

M51 is 23 million light-years away. The photons that will land on the sensor tonight left before the Alps finished rising.

Two galaxies, mid-collision. The small one, NGC 5195, has already passed through the disc of the large one — twice.

VI · The frame

A galaxy is a shape. A sensor is a rectangle. They have to agree.

Field of view uses the exact 2·atan(d ÷ 2f), not the small-angle shortcut — the shortcut is fine at 2 000 mm and wrong at 200. Rotate the camera and the numbers move with it.

When a target overflows the frame, AstroBrain lays mosaic panels on a proper gnomonic tangent plane. The flat cos(δ) approximation misplaces a panel by 6% of a frame at declination 80° — and a clamped mosaic points somewhere you did not ask for.

VII · The photon budget

How long should one frame be? There is an actual answer.

Long enough that the sky itself swamps the camera’s read noise, and short enough that the stars do not saturate. Everything between those two walls is arithmetic.

Read noise is charged once per read, as RN², not scaled with exposure time. Most calculators get that wrong, and it is why they recommend twenty-minute subs to people who do not need them.

“Never thousands of nights.”spec 252 — the depth-verdict rewrite

VIII · The night runs

Signal adds. Noise adds in quadrature. That difference is the entire craft.

One 180-second frame of M51 is an insult. Stack N of them and the signal-to-noise ratio climbs as √N — which is why four hours looks twice as good as one, and sixteen hours looks twice as good as four.

Something crosses frame 61 — a satellite, most likely. Sigma-clipping deletes the streak without deleting the frame, because those pixels disagree with the other ninety-three and the threshold does not care what made them.

IX · Morning

The sky closes in the same order it opened.

Astronomical dark ends at 04:12 and nautical twilight follows at 04:58. The faint stars go first, then everything under third magnitude, then the target. The mount parks itself long before the sun is anywhere near the horizon.

In the morning you drag the FITS files in. AstroBrain reads seven header keys and nothing else — by design — then tells you the truth: four hours forty-two is a first look, and you are eighteen minutes short of calling it solid. Deep is fifteen hours. That is three more nights like this one.

“…and you find that out in PixInsight three weeks later, when two nights will not co-stack.”services/sequence — why the Night Script exists

X · The machine

None of that was an animation.

The sun altitude, the moon phase, the sidereal rotation, the airmass, the sub-exposure length — all of it was computed from the date and the coordinates while you scrolled. Change either one and the night changes with it.

That is also how the product works. AstroBrain is deterministic at runtime: pure functions, no clock reads inside the domain, no random numbers, no model inference. Same inputs, same numbers, forever.

“AI output must not enter deterministic score vectors or acquisition recipes.”ADR-016 — the AI trust boundary

Enter

AstroBrain

A planning instrument for deep-sky astrophotography. Your site, your horizon, your rig, tonight’s sky — turned into one decision you can audit, and a sequence file your telescope will actually run.

Private beta. The planner runs in the browser against a real ephemeris.