Evenlight: I gave the AI the smallest job in the app, and measured why

Description and use

Evenlight measures sky colour and sun altitude, producing a card with a name chosen from a fixed vocabulary. After the model failed to pick from 35 words, the task was reduced to choosing from three options, achieving 17/18 accuracy.

Problem

The app claims to measure only what it sees from a photo taken by the user, but a language model on the phone performed poorly at selecting a name for the evening from a fixed vocabulary, leading to a design decision to give the model a smaller task.

Outcome

The app successfully measures sky colour and sun altitude without external data, and the redesigned language‑model workflow achieves high accuracy for name selection while preserving privacy.

Workflow

• 1. Capture a photo with the phone camera. 2. Sample eight horizontal colour bands from the image. 3. Compute sun altitude using stored coordinates and NOAA solar maths. 4. Match the most colourful band against a 35‑word vocabulary. 5. Generate a card with the colour bands, chosen name, date, and altitude. 6. Store the card locally; no network transmission.
• 7. Run the language model on‑device to select a name from a three‑choice list. 8. If the model fails, fall back to a template phrase.
• 9. Display a stack of recorded evenings, showing gaps where evenings were missed.

Evidence

The app claims exactly what it measures and nothing else.

Source

The shutter is gated on nothing but enabled = !busy. No hour check, no sun‑altitude check, no scene check.

Source

I ran it on the inputs the app would actually hand it, before writing any app code. Five designs, a 1B parameter open‑weight model on CPU, the same eighteen calls on the last two.

Source

"choose from exactly these words" (the real 35‑word vocabulary) – 0/6 on‑list.

Source

on‑list 18/18, single‑line 18/18, stable 1/6

Source