What does the accuracy percentage actually measure?

Agreement between the Night Engine's recommendation and an independent rule-based expert baseline computed from raw cloud, Moon and darkness numbers for the same night and location. It is not a measurement against observed ground truth, and we do not present it as one.

Which observing goal is the engine best at?

ISS passes, at 99% agreement, because pass geometry is deterministic and only cloud can spoil it. Planets follow at 97%.

Which observing goal is weakest?

Meteor showers, at 66%. The engine does not yet score radiant altitude or shower activity, which is a missing input rather than a mis-tuned threshold.

How often are these numbers updated?

The harness is re-run whenever the scoring code changes, and the results published here are replaced with that run. The date of the current run is shown on the page.

SnapCosmos Night Engine v1.0 · Validation

Night Engine accuracy

We publish how often the engine is right, and where it is not. Current run: 85% agreement across 600 scored location-nights, with 16 opposite calls.

Overall agreement: 85%
Locations: 102
Countries: 58
Last run: 2026-07-30

By observing goal

Each goal uses a different weighting profile, so each is validated separately. Tolerance setting: Mostly clear.

ISS passes99%
101/102 nights · 0 opposite calls · The easiest case: pass geometry is deterministic and only cloud can spoil it.
Planets97%
99/102 nights · 1 opposite call · Bright targets tolerate moonlight and thin cloud, so the two methods rarely diverge.
Moon92%
94/102 nights · 2 opposite calls · Disagreements cluster on nights with broken cloud, where 'worth it' is genuinely a judgement call.
Astrophotography78%
70/90 nights · 2 opposite calls · Most disagreements are WAIT vs GO on mediocre nights early in a week that improves later.
Milky Way75%
77/102 nights · 4 opposite calls · The engine is stricter about a risen Moon than the baseline, which costs agreement around first and last quarter.
Meteor showers66%
67/102 nights · 7 opposite calls · The weakest domain, and we know why: the engine does not yet score radiant altitude or shower activity (ZHR). That is the next fix, not a threshold tweak.

Method

How this is measured

One honest sentence first: this is agreement with a baseline, not a measurement against observed sky.

For every location the harness fetches a live seven-night hourly forecast, computes hours of astronomical darkness and Moon geometry with the same ephemeris the site uses, and runs the production scoring code — not a copy of it. The resulting recommendation (go, wait or skip) is compared with a baseline that sees only raw numbers: A rule-based expert baseline computed from raw cloud cover, Moon illumination weighted by how long the Moon is above the horizon, and hours of astronomical darkness — with no access to the engine's weighting.

Disagreements are classified. Go versus wait is a soft disagreement — two defensible readings of a mediocre night. Go versus skip is a hard disagreement, the only class that would actually send someone outside for nothing, and it is reported separately for every goal above.

What this does not prove: that the sky was in fact clear. Verifying against observed conditions requires post-hoc satellite cloud data, and that is the next step for this page rather than something we claim today.

Open items

What we are fixing next

Known gaps, in priority order
Gap	Effect	Status
Radiant altitude and ZHR for meteor showers	Meteor agreement stuck at 66%	Next engine change
Post-hoc verification against satellite cloud observations	Turns agreement into true accuracy	Planned
Atmospheric transparency and aerosol load	Milky Way nights can score well and still look milky	No public data source yet
Local terrain and coastal fog	Below weather-model grid resolution	Mitigated by the drive suggestion

Night Engine accuracy

By observing goal

How this is measured

What we are fixing next

Keep reading