Removing or scrambling a trained norm gate’s read-level selectivity worsens loss, and reopening all reads the gate suppressed causes catastrophic loss, indicating that target choice matters beyond average attenuation and that the trained model depends on filtering.

arm: norm gate · tier: 124M · override: trained gate · dose_epsilon: 0

arm: norm gate · tier: 124M · override: trained gate · junk_type: structured · dose_epsilon: 0.2

arm: norm gate · tier: 124M · override: row mean · dose_epsilon: 0

arm: norm gate · tier: 124M · override: row mean · junk_type: structured · dose_epsilon: 0.2

arm: norm gate · tier: 124M · override: shuffled · dose_epsilon: 0

arm: norm gate · tier: 124M · override: shuffled · junk_type: structured · dose_epsilon: 0.2

arm: norm gate · tier: 124M · override: suppressed reads reopened · dose_epsilon: 0

arm: norm gate · tier: 124M · override: suppressed reads reopened · junk_type: structured · dose_epsilon: 0.2

arm: combo · tier: 124M · override: trained gate · dose_epsilon: 0

arm: combo · tier: 124M · override: trained gate · junk_type: structured · dose_epsilon: 0.2

arm: combo · tier: 124M · override: row mean · dose_epsilon: 0

arm: combo · tier: 124M · override: row mean · junk_type: structured · dose_epsilon: 0.2