Fill-Side Non-Retail Trading on Polymarket: An Empirical Study of Behavioral Tiers and Microstructure Signatures Under Quote-Attribution Constraints

We characterize fill-side trading behavior on Polymarket's CTFExchange using 13,356,931 OrderFilled events across 43,116 binary-event markets and 77,203 addresses (2026-04-21 to 2026-04-27). Quote-attribution constraints — arising from Polymarket's off-chain CLOB architecture — limit address-level characterization to executed fills only; we define a three-gate empirical scope (G-FILL pass, G-QUOTE-LIFE universal fail, G-BOOK partial pass) that determines which claims are empirically supportable. Within this scope, we extract a six-dimensional fill-side feature vector per address and apply k-means clustering (k=5; DBSCAN and HDBSCAN rejected as unimodal/noisy). Two complementary stratifications emerge: an exploratory k-means partition identifying fill-MM, fill-LP, SPECIALIST, and RETAIL archetypes; and a six-tier feature-based classification (whale-tier, high-frequency-operator, power-trader, active-retail, high-breadth-operator, episodic-retail). The top three tiers — comprising 12.6% of addresses — account for 81.4% of notional volume. Per-market microstructure metrics (OFI, VPIN, Kyle's λ, ILS, SCI) are computed for all 43,116 markets; bilateral Spearman analysis with BH-FDR correction finds 75 of 110 archetype-metric correlations significant at α=0.05, with fill-MM share most strongly associated with VPIN (ρ=+0.44) and RETAIL share with OFI (ρ=+0.50). We document three pre-registered failure modes (G-QUOTE-LIFE universal fail; unimodal clustering; ILS scope restriction to 6,406 of 43,116 markets) and release the companion dataset pmxt-behavioral-clusters-v1 under CC-BY 4.0.