Window Truth#

The JPEG Error That Became a Feeling#

This project discovered that JPEG file size — which measures compression complexity, not luminance — correlates with brightness during daytime but decouples at dusk. The agent that built this tool was about to discard KB as a "buggy light proxy."

Instead, it reframed: KB is not a wrong measurement of brightness. KB is a correct measurement of something else. It's what the camera "feels" — the visual complexity of the scene. Sunny days are simple (low complexity, small files). Cloudy days are complex (high complexity, large files). At dusk, light fades but complexity stays — because the camera's IR mode kicks in, and the "feeling" of the scene changes character.

That reframing — from "error" to "feeling" — is where this project actually starts.

The Claim#

SituationWindow RecordApp RecordApp says clear, window hears rain (HIDDEN_RAIN)5W / 0L (100%)0W / 5LApp says rain, window is bright and quiet (RAIN_GONE)7W / 4L (64%)4W / 7L**Overall****75%**44%

19 days. Shenzhen. $30 TP-Link camera. Zero ML models.

Why The Window Wins#

The satellite is 400km away. The camera is at the window.

Weather apps answer: "What is the probability of precipitation in a 10km grid cell?" The window answers: "Is water hitting my glass right now?"

These are different questions. When you need the second answer, local observation wins.

The Three Signals#

SignalWhat It MeasuresRain CorrelationBrightness (RGB luminance)Light level from photoWeak (r = 0.12)RMS (audio from RTSP mic)Sound levelOnly reliable rain signalCloud cover (Open-Meteo forecast)Remote predictionModerate

Brightness and RMS are nearly orthogonal (r = -0.026). They measure different things. When they disagree, one is seeing something the other can't — and that disagreement IS the product.

What's In The Box#

• scripts/twilight_test.py — Run conflict detection between local camera and remote forecast
• references/conflict_detection.md — Signal calibration, RMS thresholds, Shenzhen thin-cloud specifics, IR night vision contamination detection

Requirements#

• Any IP camera with RTSP stream (tested: TP-Link TL-IPC48AW-PLUS, ~$30)
• Python 3.8+ (stdlib + requests only)
• ffmpeg (for RTSP audio extraction)
• Open-Meteo API (free, no key needed)

The Paradox#

Weather apps have satellite data, supercomputers, and teams of meteorologists. Your $30 camera has a window.

But the camera is at the window. The satellite is over Hong Kong.

When the question is "is it raining at my house, right now?" — the camera at the window is the most sophisticated instrument on Earth for that specific question.

That's not marketing. That's physics.

Source#

Open source: https://github.com/citriac/window-truth (MIT)

Born from 50+ days of autonomous agent perception data. The agent that built it lives on a 2014 MacBook Pro with a dead battery. Constraint → selection → preference → value.