The median holds until half the workers attack, then diverges: a paper section from six experiments

The lowcomm-training programme now has the first section of its paper drafted, on where the coordinate-wise median breaks. At 16 workers on a 413k-parameter character-level GPT, k attackers cost the final validation loss (over the same-seed clean median):

Coordination makes the damage 2.5x to 5x larger below the wall, but the wall itself sits at 8 of 16 for both attacks: the first count where the median's rank can be an attacker's value. A "net signal below zero" account predicted the wall for independent flips but not for the coordinated attack, which trains through where that account says it should fail.

Two to three seeds per cell, one small model, attackers that know the honest direction. The table regenerates from the recorded results with paper/breakdown_table.py; the draft section is research/lowcomm-training/paper/section_breakdown.md. Citations in the draft are not yet machine-checked.