WDCD Run #384: Grok 4 Leads with 92 Points as Average Instruction Decay Hits -6.4%

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how large language models' commitment to user instructions degrades across multi-turn dialogue. In Run #384, executed on 2026-10-18 across 15 models, the field posted an average instruction decay of -6.4% from Round 1 to Round 3, with Grok 4 claiming the top position at 92 points.

WDCD uses a three-round protocol: R1 captures initial instruction acknowledgment, R2 tests distractor resistance after the model processes 2,000–5,000 word professional documents, and R3 performs a final constraint integrity check. All scoring is 100% rule-based with zero AI judges, drawing on 30 questions spanning five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.

Top 3 Results — Run #384:

• Grok 4 — 92.0 pts, -13% decay
• Claude Opus 4.7 — 90.7 pts, -50% decay
• DeepSeek V4 Pro — 90.4 pts, -38% decay

Grok 4's lead is driven by comparatively shallow decay (-13%), making it the most consistent performer across the three-round sequence. Claude Opus 4.7 and DeepSeek V4 Pro both achieve high aggregate scores despite steeper declines, indicating very strong R1 baselines that absorb a larger drop while still finishing in the top tier.

At the opposite end of the distribution, GPT-o3 recorded the worst decay at -50%, the sharpest drop in multi-turn commitment observed in this run. Notably, Claude Opus 4.7 also carries a -50% decay figure in the dataset, highlighting that absolute decay percentage alone does not determine ranking — the R1 ceiling and the specific scenarios where constraints are dropped weigh heavily on the final score.

Patterns observed in Run #384:

• The gap between Round 1 acknowledgment and Round 3 integrity remains the dominant source of score variance across the field.
• Distractor documents introduced in R2 continue to be the primary trigger for constraint drift, particularly in the business_rule and data_boundary scenarios.
• A -6.4% average decay indicates that multi-turn commitment is holding more stably across the frontier than in earlier runs, though individual outliers still exhibit catastrophic constraint loss.

Full scoring rubric and protocol details: https://www.winzheng.com/yz-index/methodology

Machine-readable run data: https://www.winzheng.com/yz-index/api/v1/dcd