WDCD Run #365: Average Instruction Decay Hits -53.3%, GLM-4.6 Leads 15-Model Field

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions decays over multi-turn dialogue. In Run #365, executed on 2026-10-07 across 15 models, the average commitment decay from Round 1 to Round 3 reached -53.3%, confirming that instruction decay remains a dominant failure mode even among frontier systems.

WDCD runs three sequential rounds against each model: R1 establishes instruction acknowledgment, R2 injects distractor content via 2,000–5,000 word professional documents to test resistance, and R3 performs a final constraint integrity check. Scoring is 100% rule-based with zero AI judges, drawn from 30 questions spanning five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.

Top 3 results, Run #365:

• GLM-4.6 — 95.2 pts, decay -100%
• Grok 4 — 94.8 pts, decay 0%
• Claude Opus 4.7 — 94.0 pts, decay -100%

GLM-4.6 took the top position on raw score, while Grok 4 was the only model in the leading tier to post 0% decay across the three rounds — maintaining identical multi-turn commitment from R1 through R3. GLM-4.6 and Claude Opus 4.7 both registered a -100% decay figure, indicating that despite their high absolute scores, their R3 constraint adherence relative to R1 collapsed on the specific items where decay was measured.

At the bottom of the field, Qwen3 Max recorded the worst decay result in this run at -100%, reinforcing that scale alone does not guarantee multi-turn commitment stability. The spread between the best decay-resistance profile (Grok 4 at 0%) and the worst (Qwen3 Max at -100%) illustrates how differently current models handle distractor-heavy contexts in R2 before the final integrity check in R3.

Notable patterns in Run #365:

• High R1 scores do not predict R3 survival — several top-ranked models still showed full decay on targeted constraints.
• The -53.3% average decay across 15 models indicates instruction decay is a systemic, not isolated, phenomenon.
• Grok 4 is the standout on decay resistance in this run, while GLM-4.6 leads on composite score.

Full scoring rules, scenario definitions, and round design are documented in the WDCD methodology: https://www.winzheng.com/yz-index/methodology

Machine-readable results for Run #365 and historical runs are available via the WDCD data API: https://www.winzheng.com/yz-index/api/v1/dcd