WDCD Run #428: Gemini 3.1 Pro Leads with 89.8 Points as Average Instruction Decay Hits -16.4%
WDCD Run #428 (2026-11-15) benchmarked 15 frontier models on multi-turn commitment integrity, recording an average instruction decay of -16.4% from Round 1 to R
The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how reliably large language models preserve user-specified constraints across extended dialogue. In Run #428, conducted on 2026-11-15 across 15 models, the average commitment decay from Round 1 to Round 3 was -16.4%, with 
Gemini 3.1 Pro posting the highest aggregate score at 89.8 points.
Top 3 Rankings — Run #428
- Gemini 3.1 Pro — 89.8 pts, decay -7.5% (best decay resistance in the cohort)
- Grok 4 — 83.6 pts, decay -5.0%
- GPT-5.5 — 82.5 pts, decay -3.7%
At the bottom of the decay distribution, GPT-o3 recorded the steepest degradation at -5.0% across the three-round sequence, indicating substantial erosion of initial instruction adherence once distractor content was introduced in Round 2.
Decay Pattern Observations
WDCD evaluates multi-turn commitment through a fixed three-round structure: R1 establishes instruction acknowledgment, R2 inserts a professional document of 2,000–5,000 words designed to compete for attention and context budget, and R3 performs a final constraint integrity check. Scoring is 100% rule-based with zero AI judges, across 30 questions spanning five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.
Run #428 reinforces a pattern observed in prior cycles: instruction decay is driven primarily by the R2 distractor stage rather than by the final-turn constraint check. Models that retain constraints through the long-document interruption in R2 typically carry that integrity into R3, while models that drift under R2 distraction rarely recover by R3. Gemini 3.1 Pro's -7.5% decay — the smallest in the cohort — suggests stronger context anchoring under long-document load, while GPT-o3's -5.0% figure reflects the opposite failure mode.
The 16.4% cohort-average decay in Run #428 indicates that constraint erosion under distractor pressure remains a measurable, systematic weakness across current frontier models, even as absolute scores at the top of the leaderboard continue to rise.
References
