← back to index

Settlement-risk radar

A 4B model I fine-tuned reads the resolution rules on Polymarket markets and flags how likely the market is to end in a dispute. It shows the exact clauses it's worried about and similar past markets. I analyzed 250k past markets, ID'd the markets that had poor resolutions for traders and then fine tuned the model on Prime Intellect compute. 220x cheaper inference than Kimi K3 with the improved performance. I wrote in more detail how I built it here.

Paste a market URL and hit Analyze — usually a few seconds.

v5 vs Kimi K3

Head-to-head on the same held-out markets. The 4B fine-tune ties a frontier model on ranking, beats it on usable decisions, and costs ~220× less to run.

Ranking quality · ROC-AUC

Ranking (AUC)
0.975
ties K3 at 0.973
Decision quality (F1)
0.45
vs 0.15 for K3 · 3×
Cost / 1k markets
~$0.07
vs ~$16 for K3 · 220×

On brand-new markets neither model trained on, both drop (ours 0.70 AUC vs K3 0.66) — dispute risk decays fast. The fine-tune still wins, just from a harder baseline.

Familiar numbers: v3 on the frozen natural test. Fresh: v5 vs K3 head-to-head on June 2026 markets.