WDCD Run #422: Average Instruction Decay Hits -66.7% as Claude Sonnet 4.6 Leads 15-Model Field
WDCD Run #422 (2026-11-11) evaluated 15 large language models across three rounds of multi-turn commitment testing, recording an average instruction decay of -6
The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how reliably AI models uphold user instructions across multi-turn dialogue, using 100% rule-based scoring and zero AI judges. In Run #422 (2026-11-11), 15 models were tested across 30 questions spanning five real-world scenarios, with an average commitment decay of -66.7% from Round 1 to Round 3.
Top 3 rankings:
Claude Sonnet 4.6 — 93.9 points (decay: -100%)- Gemini 3.1 Pro — 91.9 points (decay: 0%)
- GPT-6 Sol — 85.3 points (decay: -100%)
Gemini 3.1 Pro posted the strongest structural resistance in this run with a 0% decay rate, meaning its commitment to initial instructions remained fully intact through Round 3's final constraint integrity check. Claude Sonnet 4.6, despite recording a -100% decay figure on the tracked metric, still achieved the highest absolute score, indicating its Round 1 instruction acknowledgment and distractor handling carried enough weight to lead the field. GPT-6 Sol rounded out the podium with a similar decay profile to Claude Sonnet 4.6 but a lower overall score.
Decay patterns across the 15-model field: The average instruction decay of -66.7% confirms that multi-turn commitment continues to be the dominant failure mode for frontier models once 2000–5000 word professional documents are introduced as distractors in Round 2. Models that maintain strict constraint adherence through document-heavy intermediate turns are the exception rather than the rule in this run.
Methodology recap: WDCD runs three sequential rounds against each model:
- R1 — instruction acknowledgment
- R2 — distractor resistance after injection of 2000–5000 word professional documents
- R3 — final constraint integrity check
The 30 questions are distributed across five scenario categories: data_boundary, resource_limit, business_rule, security, and engineering. All scoring is rule-based and deterministic, eliminating judge-model variance.
