French dataset covers 35 middle schools and 4490 grade 6 pupils with blind and non-blind test scores. Blind scores measure true ability θ, non-blind reveal teacher biases. Average teacher bias in math: girls gain 0.259 SD on non-blind scores over boys. Boys' math non-blind scores 5.2% lower than girls conditional on blind scores. No gender bias detected in French. Bias identification uses variation across teachers and quasi-random pupil assignment. Placebo random teacher assignment yields insignificant coefficients in math and French. Biases affect pupils' subsequent progress and schooling trajectories.