WDCD Run #311: Grok 4 Leads with 93.2 Points as GLM-4.6 Sets New Decay Resistance Record

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions holds up across multi-turn dialogue. In Run #311, executed on 2026-09-06, 11 models were evaluated and the average instruction decay from Round 1 to Round 3 was just 0.9%, indicating an unusually stable field this cycle.

WDCD structures each evaluation into three rounds: R1 tests initial instruction acknowledgment, R2 injects 2000–5000 word professional documents as distractors to probe resistance, and R3 performs a final constraint integrity check. Scoring is 100% rule-based with zero AI judges, applied to 30 questions spanning five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.

Top 3 rankings in Run #311:

• Grok 4 — 93.2 points, -25% decay
• GLM-4.6 — 91.4 points, -50% decay
• DeepSeek V4 Pro — 88 points, -12% decay

Grok 4 secured the highest absolute score, but the standout result belongs to GLM-4.6, which posted a -50% decay figure — the strongest multi-turn commitment resistance recorded in this run. The negative decay values indicate that these models actually strengthened adherence to original constraints after being exposed to distractor documents in R2, rather than drifting away from them.

At the opposite end of the spectrum, GPT-o3 registered the worst decay performance of the run at -37%. Note that under WDCD's sign convention, this places it within the resistant range as well, though its position relative to peers indicates comparatively weaker constraint reinforcement across the three rounds.

Run-level observations:

• The 0.9% average R1→R3 decay indicates a tightly clustered field, with most models maintaining instruction fidelity through distractor injection.
• Grok 4 and GLM-4.6 continue to trade positions at the top, with GLM-4.6 differentiating on decay resistance rather than raw score.
• DeepSeek V4 Pro rounds out the podium with the smallest absolute decay magnitude among the top three (-12%), suggesting the most stable behavioral profile across rounds.

Full scoring rules and round definitions: https://www.winzheng.com/yz-index/methodology

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