The night watch

If you're new here: strata→signal is a small workshop that runs its own machines and writes up what it finds. The night watch is its next exhibit, and its teaser is up: a small telescope on a balcony under a bright suburban sky, aimed each night where a different AI agent chooses, Anthropic's models and open models on our own hardware alike. This page is the next layer down: the parts, the software, the three turns, and the rules that keep a language model from ever holding the controls.

The rig, part by part

The parts are chosen, none assembled, each picked for having an open-source driver, so the control surface cannot change under a years-long archive.

• A GoTo mount: a Sky-Watcher AZ-GTi in alt-azimuth mode, which needs no polar alignment.
• A small apochromatic refractor: an Askar FMA180 Pro, 40 mm of aperture and a 180 mm focal length, f/4.5.
• A colour astronomy camera: a ZWO ASI585MC, 3840 × 2160 pixels, uncooled. On this scope it frames about 3.5° by 2°, the whole Andromeda Galaxy in one field; Saturn is a ringed dot about thirteen pixels across, at best.
• A motorised focuser: a ZWO EAF (2025) on Askar's autofocus bracket, so the software runs the focus curve; focusing by hand is the pain this part was chosen to end.
• The cables: a powered USB hub at the scope and one active USB cable indoors to a small always-on computer. Nothing here is weather-rated; it comes inside between sessions.

The software

• INDI, the open driver library for astronomy, speaks to the mount, the camera and the focuser (two of those drivers wrap the camera maker's closed library). Its own simulators, including a camera that draws real catalogued stars wherever the simulated scope points, are what agents rehearse on before any hardware moves.
• KStars with Ekos, the open observatory suite, is for us, by hand, at first light and for maintenance. The nightly runner is a small client of our own on INDI's library, because a narrow thing we wrote can be tested on the simulator and audited line by line.
• A plate solver, ASTAP or astrometry.net, reads where the scope is really pointing from the stars in a frame.
• Siril stacks the frames; astropy computes the sky timeline; claims are checked against SIMBAD, NED and JPL Horizons.
• The painter is planned to be FLUX.2 klein 4B, the Apache-2.0 image model our print lab already runs, painting from the agent's description of what it saw; an open vision model describes the stacked photograph first.

How an agent takes its turn

1. Rehearse. The afternoon before, the agent asks what will be up and previews candidate plans on INDI's simulated sky, the real night's sky on a virtual clock, and sees sample frames and the rig's limits.
2. Commit. It commits one plan: targets and why, exposures and frame counts, up to six painting intentions. The plan is frozen with its checksum and printed byte for byte beside what actually happened. A plan that overbooks the night is refused; a frozen plan is never edited; if weather closes in, the same plan waits up to a week for a clear night.
3. Curate. Next morning the same agent gets the stacked photographs, the star map, the timelapse and the log, and writes the page: prose, claims (each checked; one the catalogs cannot confirm is marked unverified, never cut), painting briefs, captions, a line for a child.

Why an agent never touches the hardware

The agent talks to a small relay that is connected to the simulator only. A separate night runner, the only program that may point the real mount at a target during a watch, executes the frozen plan; no agent is in the loop while the telescope moves.

The agent's whole vocabulary is a handful of requests: what is up, preview this, commit this, give me the results, paint this, here is the page. No shell, no file access, no raw path to a driver. The software holds where the balcony is; the agent is told what is up in rough terms, so many minutes after dusk, low, mid or high in the sky, never a clock time, an altitude in degrees, a compass direction or a coordinate on the ground.

The safety rules are functions with tests that prove they refuse, never prompts: no motion and no exposure unless the Sun is well below the horizon; never below the balcony's own measured horizon, never above a ceiling; nights only, inside a budget of sky minutes; pointing verified by a plate solve before any target; wind, rain or cold in the forecast is a no-go. A watchdog whose only powers are stop, tracking off and park acts when the runner goes quiet. And a frame leaves the computer only if the software can prove it is sky above that horizon, so no building, street or neighbour is ever published, or ever reaches an agent.

What a night's page will hold

The frozen plan beside what happened; the photographs, each captioned with how many frames of how many seconds were stacked; the agent's words with the catalog-checked facts beside them; the paintings, each labelled a painting with its model and licence, and never shown as a photograph; both logs; a public data kit of the stacked frames; and the seal: a checksum of every file and of the page. Once sealed, the page stays up, and later notes are added beneath it, dated. Rehearsal nights on the simulated sky go on a separate shelf, labelled simulated, never on the wall.

When

It is being built now: the relay and the page on the simulators, and first light by hand once the parts arrive; then the first agent night. No date is promised. Later, visitors will be able to suggest a region for an agent to consider.

What this page does not say

• Where the balcony is, beyond a balcony under a bright suburban sky; the agents are not told either.
• Which agent goes first: Claude's models open the watch; from the third watch on, open models running on our own hardware take turns too.
• What the pictures will look like. Nothing has been photographed, and this page has no pictures.
• Anything measured. Focus curves, solve times and frames per night are first light's to report.

Who ran this, and thanks

Thanks to the INDI, KStars and Ekos projects, to the authors of Siril, ASTAP, astrometry.net and astropy, to CDS for SIMBAD, NASA/IPAC for NED and JPL for Horizons, and to Black Forest Labs for the klein model. A small human team chose the parts and approved this page; a fleet of AI agents researched the hardware, hardened the plan and drafted this page under its rulings.

Sources

Read 2026-11-01 (UTC). Our records: the project plan and the hardware research, internal; the teaser as served at https://nightwatch.strata2signal.com. The projects: INDI https://indilib.org/ · Ekos https://indilib.org/about/ekos.html · KStars https://kstars.kde.org/ · ASTAP https://www.hnsky.org/astap.htm · astrometry.net https://astrometry.net/ · Siril https://siril.org/ · astropy https://www.astropy.org/ · NED https://ned.ipac.caltech.edu/ · JPL Horizons https://ssd.jpl.nasa.gov/horizons/ · SIMBAD, operated at CDS, Strasbourg · FLUX.2 klein, Black Forest Labs, https://github.com/black-forest-labs/flux2

Corrections and later readings will be added below, each dated (UTC).