WDCD Run #400: Zero Instruction Decay Across All 15 Models as GLM-4.6 Leads at 96.2 Points

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how reliably AI models hold user instructions across multi-turn dialogue, using 100% rule-based scoring and zero AI judges. In Run #400, conducted on 2026-10-28, all 15 evaluated models recorded 0% average commitment decay from Round 1 to Round 3 — the headline result of this cycle.

Leaderboard — Top 3:

• GLM-4.6 — 96.2 pts, -0% decay
• Gemini 3.1 Pro — 94.0 pts, -0% decay
• GPT-o3 — 92.3 pts, -0% decay

GLM-4.6 holds the top position on both raw score and decay resistance. With a 2.2-point margin over Gemini 3.1 Pro and a 3.9-point margin over GPT-o3, GLM-4.6's lead stems from stronger Round 1 instruction acknowledgment rather than differential degradation, since all three leaders — and in fact the entire 15-model field — posted identical 0% decay figures.

Decay pattern observations: Run #400 is notable for the complete absence of instruction decay across the cohort. R2, which inserts 2000–5000 word professional distractor documents between acknowledgment and the final constraint check, failed to dislodge any model's committed constraints at an aggregate level. The R3 final constraint integrity check confirmed that multi-turn commitment held across all 30 questions and all 5 scenario categories: data_boundary, resource_limit, business_rule, security, and engineering.

Because every model scored -0% decay, the "worst decay" and "best decay resistance" slots are both occupied by GLM-4.6 by virtue of its leaderboard position — a statistical artifact of a fully saturated decay metric rather than a differentiator between models. In this run, score separation is driven entirely by baseline instruction-following accuracy at R1, not by resistance to distractor-induced drift at R2 or R3.

Takeaway: Run #400 suggests the current top tier of models has, for this question set, effectively solved the surface-level multi-turn commitment problem measured by WDCD's three-round structure. Future runs may require harder distractor payloads or longer context chains to re-introduce measurable instruction decay and restore discrimination on the decay axis.

Methodology: https://www.winzheng.com/yz-index/methodology
Data API: https://www.winzheng.com/yz-index/api/v1/dcd