The sun that pointed the wrong way: when a number is right and the world is wrong
We had just finished building the world’s clock: the machinery that turns the minutes you spend playing into a time of day, a day of the year, and a season. Days lengthen toward midsummer and shorten toward midwinter; the sun climbs and sets; spring hands off to summer to autumn. Every stage had been checked by an automated test, and it stayed green the whole way through.
Then, the very first time a person actually looked at the rendered sky, the noon sun was below the horizon — shining up, lighting the underside of the terrain like a lamp held beneath a table. Nothing had failed. Every number was correct. And the world was unmistakably wrong.
What the clock had to do
Most of this system is quiet arithmetic. Given a moment in elapsed play time, it answers the same questions the same way every run — what time of day is it, how high is the sun, which of the four seasons are we in — and from those answers fall the things you actually see, chief among them where the directional light points. The full spec for that clock — the day-length curve, the four equal seasons handing off — is a story told elsewhere; what matters here is one detail. We built it in stages, and at each one an automated check confirmed the math matched the numbers we expected. By every signal the tooling could give us, the clock worked — in the only sense those checks could measure. That sense turned out to be narrower than it looked.
The values we left for last
One whole category of the work is not arithmetic at all: it’s how the sky looks. Which direction the sun comes from at any hour, how the light is coloured through the day, what tint each season washes over the world — these aren’t things a test can call right or wrong. They’re things you tune by feel. So they went in early as placeholders: stand-in numbers, plausible enough to keep the machinery running, marked to be set properly at the very end during an aesthetic pass.
That aesthetic pass is also, in this project, the first time the sky is rendered for a human to look at. Everything before it is checked against expected values on paper; the actual picture doesn’t get a viewer until the tuning stage at the close. Two of those early placeholders were quietly, geometrically nonsensical from the start — and because no one had yet looked, both rode all the way through every green stage undetected.
A noon sun shining up from underground
The first was the one that names this post. The placeholder for the sun’s direction had been authored with its vertical sense inverted. The value that was meant to say the sun is overhead at noon instead pointed straight down — which, geometrically, places the sun beneath the world at the height of day. Run through the rule that turns a sun direction into how the world’s main light is actually oriented, that flip pointed the light the wrong way too: at midday the directional light shone upward, lighting the bottoms of hills and the undersides of terrain instead of their tops.
It was, in effect, a noon sun coming from under your feet. There is no subtler way to say it was wrong; it was as wrong as a sky can be.
The fix was small — flip the vertical sense back, so the noon sun sits high overhead, the light falls downward where it belongs, and the dawn and dusk angles lift just above the horizon instead of sinking below it. What’s worth dwelling on isn’t the fix; it’s how a mistake this total survived every test we had.
Why a number can be right and still be wrong
Here is the heart of it. An inverted sun is, to a numbers-only check, completely well-behaved. Feed the same time in and you get the same direction out, every run, byte for byte. The value is stable, continuous, and satisfies every contract the clock makes with itself. The check that verifies the machinery can confirm all of that and find nothing to object to — because the thing it can see is whether the value matches its own definition, and this value matched its definition perfectly. The definition was simply describing a sun that shines the wrong way.
That is the distinction we’d been blurring without noticing: internally consistent is not the same as actually correct. A value can do exactly what its formula says and still mean something the world would never accept. The check tests that the per-frame machinery mutates the light correctly; it cannot test that the mutated values are geometrically right. The only instrument that can tell a sun pointing up from a sun pointing down is a person looking at the rendered world — which, in this project, first happens at the tuning pass. That makes visual review not a final polish but the earliest point where a whole class of correctness even becomes checkable. We logged it not as a process failure but as a structural truth about the work: there is a kind of wrong that no test can see, no matter how green, invisible right up until the world is drawn and somebody looks — and then the most obvious thing in the frame.
The verifier confirms that moving the light works. It does not confirm that the light is pointing the right way. Only an eye on the picture can do that.
The second cousin, found the same day
The inverted sun had a sibling caught in the same sitting, the same shape of mistake. The day-length curve — what makes daylight longest in summer and shortest in winter — has a setting meant to place the longest day of the year, the summer solstice. The value beside it claimed to mark midsummer, but the underlying curve actually peaks a quarter of a year after that mark. So the longest day was landing not in mid-summer but on the first day of autumn — the calendar and the sky a full season out of agreement.
The repair re-anchored the phase so that, across an evenly quartered year, the equinoxes fall on the spring and autumn boundaries and the solstices on the summer and winter ones — calendar and sky finally telling the same time. Like the sun, it was self-consistent the whole way through, honouring every internal contract while quietly meaning the wrong thing, visible only once someone watched the year turn under the rendered light. It’s a recurring trap: a comment confidently asserting a state that reality contradicts. We’ve written about a different face of it — a check that ran green while verifying nothing — in The test that proved nothing; this is the version a test simply cannot reach at all.
Setting the year to the right pace
With both defects corrected, the same pass tuned the clock to feel right — landing a full turn of the seasons at a pace you can sit inside rather than race past, the year-length tuning told in full elsewhere. The four seasonal tints went in here too — a fresh green for spring, a warm gold for summer, an amber for autumn, a cool blue for winter — reviewed across several sample starting worlds before the work was marked done. The sky finally agreed with the calendar, the light fell from above, and the year moved at a pace you could sit with.
Why getting this right is load-bearing
It would be easy to file an inverted sun as a footnote — a one-character flip, fixed in an afternoon. But this is a game whose entire mood rides on light: warm, low-contrast, golden-hour, a world soft enough to want to tend. Light shining up from underground isn’t a small blemish on that; it’s the difference between a world that feels lit and a world that feels lit from beneath, uncanny in a way you’d feel before you could name it. And the same trap waits everywhere there’s a rendered surface to get wrong — terrain shading, the world you preview before you settle in, the creatures, the scatter of plants across a hillside — each able to carry a value that satisfies every test and still looks wrong the instant it’s drawn.
So we came away holding the lesson tight: machine checks lock in behaviour, but only an eye catches a sun that points the wrong way. The visual review at the close isn’t a formality at the end of the real work. It is the first time anyone sees what was built — and the first time the world gets to tell us we were wrong.
