Check what the scene claims.

The sky

Sun, Moon and its phase, the planets and the stars are placed by ephemeris for that instant, that latitude and that longitude. Nothing is a picture of a sky: point at any of it and it names itself — “Venus, mag −4, 8° above the horizon”.

What is drawn from the catalogue stops at magnitude 6.5, and that number is a fact about a observer rather than about the sky: it is how faint a dark-adapted human eye goes. Put something else in front of the same night and the threshold moves with it — see the instrument.

The catalogue follows the same threshold. A naked-eye sky costs 400 kB, down to magnitude 7.5; 57 688 more stars, down to magnitude 9, are fetched only by a recording whose own optics reach that far. Past magnitude 9 it is the data that stops, and the tool reports the limit rather than drawing a sky emptier than the photograph held.

The catalogue is written for the year 2000, and the axis the sky turns about has moved since. Every star is carried to the date of the observation: a degree in 1918, fourteen in 1006, which is the difference between two constellations.

The Milky Way and the zodiacal light are integrated along the line of sight rather than painted as a texture — which is why they move correctly with the season, the hour and the observer's latitude instead of merely being in the right place once.

The light of the sky

The sky itself is not a colour gradient. Sunlight and moonlight are traced through a spherical atmosphere at fifteen wavelengths across what an eye sees, the way light really reaches an observer: scattered once by the air on its way, then again and again, dimmed by what stands in front of it. Three things make up that air. The molecules of the air itself, which scatter blue far more than red and make the sky blue. The haze, which whitens the horizon and puts a bright aureole around the Sun, and which is thicker on a humid day than a dry one: its amount follows the relative humidity of the weather record. And the ozone layer, which absorbs a little orange, invisible at noon and the whole reason a twilight zenith is deep blue rather than a greyish yellow.

Everything a clear sky does then follows without being painted: the reddening and the brighter band along the horizon at sunset, the Earth's shadow rising opposite the Sun with the pink band above it, the slow darkening of twilight toward the east first, the deeper blue seen from an aircraft, the thinner air above a high site, and the moonlit sky, which is the same scattering of a far fainter source. The airglow of the upper atmosphere stays as the floor of every moonless night, brighter near the horizon where a line of sight crosses more of its layer.

The model is checked rather than trusted. A slow Monte Carlo trace of the same atmosphere, with no approximation but noise, puts the sky the scene draws within a quarter of a magnitude and a hundredth in colour from a hazy noon to a Sun sixteen degrees down, from the ground and from an aircraft. One place is known to be too dark, by two thirds of a magnitude: the Earth's shadow low in the sky during nautical twilight. And the zenith meets the twilight photometry measured at the Paranal observatory to within half a magnitude, without ever having been fitted to it.

What reaches the screen is what an eye adapted to that sky makes of it: a night sky is not ten million times darker than a day sky to an observer standing in it, and colour gives way to a dim blue-grey as the rods take over from the cones. A camera does none of this, so a recording made through one keeps the blue a long exposure really shows. How bright a day and a moonless night are shown are the two choices in that chain, and they are stated as choices.

What it does not do: no light pollution, no volcanic haze, no refraction of the Sun below the horizon, and the clouds are drawn on their own rather than dimming the light of the sky around them. How much haze a given town put in the air is unknowable from a record of 1948; only how much the humidity swelled it is modelled. The sources, and the choices behind each of them, are listed under the air.

What else was up there

Each of these is a candidate explanation, so each must answer the same question before it is drawn: could it have been visible from there, then?

Comets appear at their own apparition and nowhere else. The catalogue is generated from JPL Horizons and the orbit is propagated, so a comet is either there on that date or it is not.

Novae and supernovae shine only on the nights their recorded light curve covers, and nowhere outside it: before the first record the star may already have been rising, and nobody can say how bright. Thirty-three eruptions reached the naked eye, from the supernova of 1006 to RS Ophiuchi in 2021. Their brightness is measured by the AAVSO for every nova up to 2006, reconstructed from Tycho's and Kepler's own comparisons for 1572 and 1604, and held to a few chronicled dates for 1054 and 1181; each one says which.

Eclipses are never declared. At every instant the Sun's and the Moon's positions as seen from the observer's own place, parallax included, say how much of the Sun's disc the Moon covers; the scene draws that crescent, the daylight it leaves, the corona when nothing is left, and the stars and planets the darker sky then allows, and for the second or two either side of totality the beads of light that the Moon's own relief lets through (NASA's LOLA map of its surface, with the libration of the day), and the orange ring round the whole horizon of lit air beyond the Moon's shadow. It is why 11 August 1999 is total at Reims and a mere 99 % at Paris. Seen bare-eyed, a crescent Sun is still a dazzling one: the shape shows through a filter, as it does for anyone who looks.

The Moon is drawn at the size it has that day: its orbit is an ellipse, so its disc is 0.56° across at perigee and 0.49° at apogee, the 14 % that “a supermoon” is made of. It is not larger on the horizon, where it is even a little farther: that impression is an illusion of the mind, which a camera does not share, and the scene does not copy it. What is true of a low Moon is its colour, which is the air's: the long path takes out its blue. That is all there is to the “lune rousse” of French lore, the full Moon of the lunation after Easter, which stands low because the Sun opposite it is already high.

Meteor showers come with their radiant and their hourly rate for that night, over the sporadic background that never stops.

Recorded fireballs are events, not rates: since December 2018 the Global Meteor Network's cameras have triangulated where each bright one began and ended and how bright it peaked, and a recording of that date draws every one that burned during it, from where the observer stood, as bright as its distance makes it.

Satellites get the harder question, because being lit is not the same as being seen: the Earth's shadow is computed for the date and hour to say whether one could have caught the Sun at all, and how many objects were in orbit on that date is known, and stated. Since February 2021 the tool goes further: the orbital elements saved once or twice a day by Laurent Chabin (SCEAU) give every satellite's real pass, lit through the Earth's shadow and as bright as it was measured to be, Starlink trains included. Before that date no individual pass is drawn, because none could be drawn truthfully.

Re-entries are placed by an interpretation, not found: a satellite or rocket stage breaking up into burning pieces, on a path somebody established (orbital elements are off by minutes in the last orbit). It is placed on the Earth, 70 to 100 km up, and seen through the Earth's curvature from where the observer stood, each piece as bright as its distance makes it, with its tail along its own path, until it has burned. It glows from its own heating, so the Earth's shadow does not put it out. Which re-entries could have been in a sky is stated from the record (CORDS since 2000, the satellite catalogue's decay days since 1957), each with its precision: a sighting on record near enough in time and place, a prediction whose window meets the observation, or a bare day. An orbit that never comes within sight of the observer's latitude is never listed.

Ice and water

22° and 46° haloes, sundogs, the parhelic circle, tangent, circumzenithal and circumhorizontal arcs: not one of those angles is stored. Each is derived from the refractive index of ice and the geometry of the crystal, which is what makes the whole display move together and stay consistent with the Sun's own height.

Rainbows and moonbows are ray-traced through a spherical drop for the same reason — the order of the colours, the gap between the two bows and the light inside the primary come out of the trace rather than being drawn in.

The weather that day

Cloud cover, cloud base, rain, snow, hail, storms and their thunder, and wind are read from ERA5, the ECMWF reanalysis: hourly, worldwide, from 1940 on (for the last five days, which it has not caught up with, the forecast model's own analysis). They are keyframed along the observation, so a sky that cleared during those four minutes clears in the reconstruction.

A record says how much of the sky each layer covered, never where: the arrangement of the clouds is the reconstruction's own. It is drifted along the day by the record's wind and set apart from one day to the next, so that scrolling through the hours shows clouds that move and change, but it is not a picture of the clouds that were there.

Cloud attenuates every celestial body rather than merely covering it, which is what makes a Moon behind thin cloud read as a Moon behind thin cloud.

The exact query is kept in the recording. That is the part that matters: the claim stays checkable decades later, by someone who does not trust this tool.

The ground

Real relief and aerial imagery are fetched around the observer, along with the decor that got in the way: buildings, trees, streetlights, vehicles, windows and other observers — with their lights, their flash rates and their tracks.

How high the observer stood is looked up from where they stood. The ground under those coordinates comes from a real elevation model, so the altitude on the form is a height above sea level whose floor is the ground itself: nobody can be placed under it. And it is not a detail — at 1500 m the horizon really is 1.2° lower than it is at sea level, which is enough to decide whether something was above it.

The instrument

An eye is not a lens. Naked-eye viewing maps an angle to an angle; a camera maps it to f·tan θ, with a sensor, a focal length, an aperture and an exposure that draws star trails and dots a flashing light. Switch the device and the whole frame changes — the same sighting through three of them.

Only what that device could actually have been set to is offered: an Instamatic had one aperture and one shutter speed, so there is nothing to choose, and a camera that did not exist yet is flagged against the observation's own date.

And the instrument decides how faint a thing could be recorded at all. That same Instamatic stops at magnitude 4.2 — two short of the observer holding it, which is half of why so many “the sky was full of stars” accounts come with an empty black photograph. A 50 mm at f/2 for twenty seconds reaches 9.7, three magnitudes past that observer. Aperture, shutter and focal length settle it against the sky's own brightness.

A longer pose stops helping once the sky has slid further than the lens can resolve, which is why an hour on a tripod is no deeper than five seconds on one — only longer trails.

Something can also be held in front of the instrument. A solar filter — eclipse glasses, or the film that goes over a lens — cuts everything by a hundred-thousandfold: the sky goes black, no star is drawn, and the Sun is left alone, which is the only way the shape of a partly eclipsed Sun can be seen at all (the partial phases of 1999, through one). It changes no angle, only the light, and it is put on and taken off along the recording: the glasses come off for totality.

A picture of the place

Everything above is computed, and a reader looking at the result has no way to tell a faithful reconstruction from a plausible one. A photograph of the same place does. A recording can carry pictures of where it happened — an observer's own, an investigator's, a magazine scan, a street-level capture — each laid over the render at an opacity the reader slides between all picture and all render, with the phenomenon drawn over both. Every tree the thirty-metre relief smoothed away, every ridge, the actual hedge, is then one picture with the reconstruction. Cussac carries its first: the 1968 view from the spot with the sphere and its climb drawn on by hand. With several pictures, the button opens a list above it, where each one is shown or hidden and faded on its own: McMinnville lays two prints over the same scene.

A picture is a field of directions from one point, and it is lined up as one: heading, pitch, roll and the lens's field, angles and nothing else. It stands in the three-dimensional scene as a flat panel at that field — which is what a lens makes — and is drawn through whatever the instrument declares, so the same picture is right under an eye's equidistant view and under a 50 mm's rectilinear one, and stays put when the reader turns. Nothing in the scene hides it and it hides nothing.

Lining it up is a measurement, not a feeling. Name a landmark on the picture and the same one in the render — a church tower, the end of a hedge, the horizon under a tree — and two of them turn the picture to fit; three or more fit its field too. The residual says how well it fits, landmark by landmark, in green, orange or red: one that will not go below a degree is telling you the picture was not taken from that spot, or that the relief is wrong there. And a picture that fits gives back the heading the observer actually faced — adopted into their pose with its provenance written down, where a heading typed in was only their word. Street-level pictures taken nearby (Panoramax, open imagery) arrive already lined up in heading, a full turn arriving as a panorama.

Every source is named where its data is reported, with the attribution its licence requires — and can be swapped for another. The picker is the credit. What is still missing, and what each item is waiting on, is on the roadmap.