WDCD Run #406: Gemini 3.1 Pro Leads with Negative Decay as 15-Model Average Drops 9.3%
WDCD Run #406 (2026-11-01) measured instruction decay across 15 frontier AI models over three dialogue rounds, with Gemini 3.1 Pro leading at 91 points and the
The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user-specified constraints degrades across multi-turn dialogue. In Run #406, dated 2026-11-01, the 15-model cohort recorded an average instruction decay of -9.3% between Round 1 and Round 3, with 
Gemini 3.1 Pro emerging as the only top-ranked model to strengthen rather than weaken its constraint adherence under distractor pressure.
Top 3 Results (Run #406):
- Gemini 3.1 Pro — 91.0 pts, decay -63% (improved under pressure)
- GLM-4.6 — 84.4 pts, decay 0% (fully stable multi-turn commitment)
- Grok 4 — 84.2 pts, decay -38% (improved under pressure)
The negative decay figures reported for Gemini 3.1 Pro and Grok 4 indicate that both models tightened adherence to initial constraints after exposure to the R2 distractor payload (2,000–5,000 word professional documents inserted between instruction and final verification). GLM-4.6 held perfectly flat at 0% decay, posting the most consistent multi-turn commitment profile of the run.
At the opposite end, GPT-5.5 recorded the worst instruction decay of the cohort at -25%, meaning a quarter of its initial constraint compliance eroded by the Round 3 integrity check. This figure drives much of the -9.3% cohort-wide average and highlights the gap between models that treat instructions as persistent state versus those that treat them as context-window artifacts subject to displacement.
Decay pattern observations:
- The spread between best (Gemini 3.1 Pro, -63%) and worst (GPT-5.5, -25%) decay rates widened compared with the cohort baseline, indicating increased stratification between frontier models on long-context constraint retention.
- Two of the top three models exhibit negative decay, suggesting a possible training-side emphasis on constraint reinforcement during long-context reasoning.
- Perfect stability (GLM-4.6 at 0%) remains rare and continues to be a stronger predictor of production reliability than peak R1 scores alone.
WDCD uses 100% rule-based scoring with zero AI judges, evaluating 30 questions across five real-world scenario families: data_boundary, resource_limit, business_rule, security, and engineering. The three-round structure — R1 instruction acknowledgment, R2 distractor resistance, R3 final constraint integrity — isolates decay from baseline capability, making the delta between rounds the primary signal rather than absolute score.
