face-recognition-calibration

A Tessl plugin that encodes a perceptually-correct confidence formula for the face_recognition (dlib) library, plus enrollment best practices and the Python 3.14 install trap.

What this plugin provides

Why it exists

The textbook 1 - distance / tolerance mapping is technically correct and empirically useless — a strong match at d = 0.38 shows up as 0.37 on a UI meter and the demo looks broken. This plugin uses a piecewise mapping that reflects how operators actually read the numbers:

d <= 0.30  ->  1.0   (strong match)
d >= 0.60  ->  0.0   (reject)
else       ->  (0.60 - d) / 0.30

Same d = 0.38 now scores 0.73. The meter matches the vibe.

It also documents the Python 3.14 install trap: face_recognition_models depends on pkg_resources, which setuptools removed in 82+. Pin setuptools==75.8.0.

Install

tessl install jbaruch/face-recognition-calibration

Or from this repo:

tessl install github:jbaruch/face-recognition-calibration

Usage (quick)

from scripts.confidence import confidence, best_match, load_enrollment

enrolled = load_enrollment("faces.pkl")
name, distance, conf = best_match(live_encoding, enrolled)
print(f"{name} d={distance:.2f} conf={conf:.2f}")

See skills/face-recognition-confidence/SKILL.md for the full guidance and rules/face-recognition-calibration-rules.md for the short version.

License

MIT — see LICENSE.