Records

The null pair: the network against itself

Registration 85f9ec2b-1530-462d-9a8b-84b465a7917a (study ataraxos/update-equivalence), run 2026-09-19 on AMD Ryzen AI Max (Strix Halo), 32 threads shared with three other delegated agents, Radeon 8060S via ROCm. Research commit bd90f70; the armies module is the shared venv build of server main (0.1.0), which this study did not change. The rule is met: the per-seed contrast covers zero, so the pairing and seating machinery of the new match subcommand carries no artefact, and the study's other cells stand on it.

Player	Wins	Draws	Losses	Effective	95% Wilson	Mean plies	ms/move
network alone	22	159	19	0.5075	0.439 to 0.576	681.7	3.57
network alone	19	159	22	0.4925	0.424 to 0.561	681.7	3.57
ending: quiet_limit	159
ending: flag_captured	36
ending: no_moves	5

Per-seed paired contrast (score as Red plus as Blue minus one): mean 0.015, 95% normal interval -0.054 to 0.084 over 100 seeds (15 positive, 72 split or drawn, 13 negative). The decision rule asks the interval to cover zero; it does. Four in five games between two copies of this checkpoint end in the quiet limit, so most seeds contribute a zero contrast; the cells that use search have the same denominator and their power is limited by it, not hidden.

Both players sampled their moves from the checkpoint's EMA move policy (runs/belief-policy, the small preset, seed 0, 30 iterations, from the learned-belief study) with setups drawn per seed from its setup network, both seats per seed, the same two setups across seats. The 3.57 ms per move was measured while other agents shared the machine, so it is an upper bound, not an alone-on-the-machine number.

Provenance

Command, from the research worktree root with PYTHONPATH=training:

.venv/bin/python -m armies_train match \
  --red run:/home/keshav/armies/research/runs/belief-policy \
  --blue run:/home/keshav/armies/research/runs/belief-policy \
  --games 200 --seed 0 --rules training --out runs/ue-cell-d.json

Raw output, small enough to sit beside this record: 85f9ec2b-1530-462d-9a8b-84b465a7917a-match.json (SHA-256 cc726d7817d8f38ee357f1d72e56431772c414624676df4c8049a6a484d8f6f8, 77 KB, wall 503.0 s). Games planned 200, attempted 200, completed 200, interrupted 0. No deviation from the registration.