Skip to content

SpellBench v2 score on the Pauper pool

2026-11-02 → 11-03

Where does the FDN-trained M_CTLX search bot stand on SpellBench's Pauper board (pauper-bo1, eight decks, 112 distinct cards) against gorge's rule-based bots, and does a Pauper-trained net change that?

Verdict

M_CTLX at 400 sims came first at 1446, but the top five entries are not separated, and it lost its direct match with gorge-bot. At 100 sims the net adds little over the same search with no net (p 0.30): most of its strength on this pool comes from search. This motivated the Pauper retrain.

Setup

SpellBench arena run, 2026-11-02: 1,792 games, all rated, validate: OK; SpellBench 404c4c0, mtg-kernel 92ec2393, stats seed 20261102. Elo is anchored at uniform = 1000; intervals are the arena's CI95, a paired bootstrap over seat-swapped pairs. SpellBench v2 agents describes how gorge's bots sit a SpellBench seat.

Results

  • az search
  • rule-based or baseline
  • bar length is Elo above the uniform anchor, 1000 to 1500
Bot Elo CI95
az-mctlx-400 (M_CTLX, 400 sims) 1446.0 1392–1511
az-mctlx-100 1416.5 1364–1478
sb-search-lite-atk 1408.9 1352–1474
gorge-bot 1392.1 1339–1457
sb-tactical 1392.1 1338–1455
az-nonet-100 1368.0 1312–1430
heuristic 1201.4 1146–1263
uniform 1000 —

az-mctlx-400 went 295-153 over 448 games with a 3.53% fallback rate. Head to head (64 games each):

Match Result Score p
mctlx-100 vs nonet-100 37-27 0.578 (0.469–0.688) 0.30
mctlx-400 vs nonet-100 45-19 0.703 0.002
mctlx-400 vs gorge-bot 30-34 0.469 —

Only 6 of the pool's 112 distinct cards appear in the FDN deck pool.

The Pauper-trained net's boards (sb-g6, sb-fresh) are on the Pauper retrain page: pg6 led sb-g6 at 1572.8 [1489.8, 1673.8], then was level with gorge-bot (1418.5 against 1414.2) on the fresh, latency-capped board.