How it works
One photograph, taken apart
How Next Light turns a photograph you already made into a forecast for the light inside it — followed the whole way through on a single evening, the one where thin cloud took the color after the sun had already gone.
You already know what you're waiting for. You stood out that evening, watched it happen, and got the shot. Recognizing the light was never the problem. Knowing when it's coming back is.
A weather app can't answer that, because a weather app doesn't know what you're asking. It forecasts the sky in general. You're asking about one specific sky — the one in your photograph. Next Light starts from the photograph and works backward. Here is the whole machine, a stage at a time.
Stage 1
Reconstruction — the app time-travels to your photograph
When you import the photograph, Next Light doesn't ask what it looks like. It asks what the sky was doing. The file carries its own moment, and from that the server rebuilds the atmosphere of that exact minute: archive weather — the measured past, not somebody's old forecast — and the geometry of the sun, computed to the minute.
Reconstructed · July 11 · 9:20 PM
- Sun elevation
- −1.9°13 min after sunset
- Sun azimuth
- 308°northwest
- Twilight
- Civil twilight
- High cloud
- 100%the canvas color can land on
- Mid cloud
- 0%
- Low cloud
- 0%what sits between you and the horizon
- Precipitation
- 0.0 mm
- Air clarity
- 0.13haze and smoke mute color
- Moon
- 98%of the disc lit
Why this way
We could try to guess from the pixels why the photograph is beautiful. We don't. A guess can't be explained, audited, or corrected. A lookup can. Everything downstream inherits that honesty.
Stage 2
Intent — you tell it the one thing no model could know
The reconstruction says what the sky was doing. It can't say what you cared about. The color overhead? The fog in the valley? The still water? For this photograph you'd tap one answer — Colorful sky — and that single tap is very nearly all the system asks of you.
- Clear open sky
- Colorful sky
- Fair-weather cumulus
- Falling snow
- Fog & mist
- Freezing fog
- Frost
- High cirrus sky
- Low cloud blanket
- Rain & wet mood
- Soft overcast light
- Still water reflection
- Valley fog
- Windswept & dynamic
Why no AI vision
This is a deliberate absence. A vision model could list everything in the frame — ridgeline, cloud, trees — but it can't know the color was what you were chasing and the ridge was only the anchor. One tap of intent carries more than a thousand detected objects, and it keeps every number that follows explainable.
Stage 3
The fingerprint — your photograph becomes a testable definition
Intent and reconstruction snap together into a fingerprint: a precise, checkable definition of this photograph's conditions at this spot. Not a mood — a specification, with hard requirements, graded preferences, and a time window.
The fingerprint, in plain terms
Colorful sky is one of 14 condition types in a hand-built registry. Each one is written down, held to fixed test cases, and tuned by replaying real archived weather against real photographs before it ships — and the verdict on whether it worked comes from someone who was standing there.
Why a registry, not a formula
Fog is won and lost on humidity and wind. Color is won and lost on cloud geometry and clean air. One universal "good conditions" score would flatten exactly the differences that make your photograph your photograph. So every look gets its own definition — and its own evidence.
Stage 4
The watch — every future twilight, scored against your photograph
From the moment the fingerprint exists, a watcher re-scores every upcoming twilight at your spot against it, run after run, as far ahead as the forecast reaches.
For a sky that colors, the question is mostly geometry. After sunset the light that paints cloud arrives from below the horizon, traveling up a corridor of sky along the sun's path. So the watcher doesn't only check the weather over your head — it fetches cloud conditions along that corridor, far out past the horizon you can see, toward where the sun went down. A single forecast cloud bank sitting in that tunnel is the difference between a grey dusk and the photograph you took.
The watch is also honest about time. Color forecasts age badly, so this look isn't allowed to raise a first alert at the long end of the range at all — not because the arithmetic can't produce a number out there, but because that number wouldn't deserve your drive.
Why alerts are rare on purpose
An alert costs you something real: an early alarm, a drive, a cold hour in a field. Every threshold in the system is tuned against that cost — which is why silence is the default, and an alert is meant to be news.
Stage 5
What the alert says
A later evening scored above the bar, and the window ran 8:00 PM–10:00 PM. Here is what the alert that went out actually told you.
Match strength
96 / 100
How closely that evening's forecast resembled your photograph's sky. This photograph's alert bar sits at 90 — under it, nothing is sent.
Confidence
Medium
How much that forecast deserved to be trusted at that range. A word, not a score.
Confidence is where the system argues with itself. It decays as the window sits further out. It discounts itself when the score hinges on exactly where the forecast drew a cloud edge along the sun corridor — the single most fragile thing a weather model does. And some looks are capped outright: where the evidence behind one doesn't yet support the word High, the tier refuses to say it.
Alongside the numbers, the alert explains itself in plain language — the reasons on one side, the named risks on the other:
Risks, spelled out
- The horizon walls overLow cloud is forecast to build in the west — the light may never get under it, however good the sky overhead looks.
- The canvas seals shutThe thin cloud may thicken from a canvas into a ceiling: grey rather than glow.
- The air goes hazySmoke or haze is drifting in, and it greys the color down from behind.
- Rain smothers the windowWeather nearby could bury the whole window while it is open.
Why confidence rides along
"The sky should look like your photograph — but the forecast is only moderately trustworthy this far out" is a claim you can hold us to. You can act on it, and afterwards you can grade it. A promise of great light with nothing attached to it is one you could never check.
Stage 6
The loop — your verdict adjusts the machine, within limits
After a window passes, Next Light asks how it actually went — not a thumbs up or down, but a graded verdict, from missed it to nailed it, with a tap for what specifically differed: the color, the cloud, the haze.
Your answer adjusts that fingerprint's tolerances by bounded, written rules. Tell it the sky came out too grey twice and the canvas requirement tightens — by a capped amount, in a direction a person decided in advance. No retraining. No black box. Every adjustment is a rule you could read, and every adjustment has a floor: the system is never allowed to become blindly sure of itself.
You're not feeding a model. You're calibrating an instrument.
Why bounded rules, not learning
A learned model would quietly drift, and you'd never know why this month's alerts differ from last month's. Bounded rules mean your feedback has visible, reversible consequences — and the system stays yours to steer.
The spine
Three rules the whole system is built on
You're told how much to trust it.
Every alert carries how confident the forecast deserves to be at that range — so a long-range maybe never arrives dressed as a sure thing.
Nothing without the photograph.
Every feature faces one test: would it still make sense with your photograph deleted? If it would, it doesn't belong here. That's the line between this and a weather dashboard.
Explainable all the way down.
No computer vision, no machine learning, no vibes. Every alert can show its reasons, name its risks, and point at the rule that produced it.
Request an invite
Next Light is invite-only while the circle is still small. Leave your email and we'll be in touch as it widens.