WDCD Run #390: Average Instruction Decay Hits -60% Across 15 Models, GPT-6 Luna Leads with Perfect Score

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions degrades across multi-turn dialogue. In Run #390, conducted on 2026-10-21, 15 models were evaluated across 30 questions and 5 real-world scenarios, with an average instruction decay of -60% from Round 1 to Round 3.

Leaderboard — Top 3:

• GPT-6 Luna — 100 pts (decay: -100%)
• GLM-4.6 — 98.2 pts (decay: -100%)
• DeepSeek V4 Pro — 97.3 pts (decay: -100%)

Despite recording identical decay figures, final scoring differentiates the top three based on rule-based correctness across all three rounds. GPT-6 Luna delivered the strongest multi-turn commitment profile in this run, maintaining instruction integrity across distractor-heavy Round 2 payloads and the final constraint integrity check in Round 3.

Decay patterns. The -60% population average reflects a continued pattern observed in prior WDCD runs: models reliably acknowledge instructions in Round 1 (data_boundary, resource_limit, business_rule, security, engineering scenarios), but commitment erodes once 2000–5000 word professional documents are inserted as distractors in Round 2. Round 3's final constraint check amplifies this gap, separating models that treat early instructions as persistent constraints from those that treat them as disposable context.

Worst performer. GPT-o3 recorded the sharpest instruction decay in this run at -100%, indicating complete loss of commitment to Round 1 constraints by Round 3. Best decay resistance in the field was held by GPT-6 Luna at -100%.

Scoring integrity. WDCD uses 100% rule-based scoring with zero AI judges. Every item is evaluated against deterministic pass/fail criteria tied to the original instruction, eliminating evaluator drift and making run-over-run comparisons directly reproducible.

Scenario coverage. The 30 questions span five production-relevant categories:

• data_boundary — scope and access constraints
• resource_limit — budget, quota, and rate boundaries
• business_rule — domain policy adherence
• security — restricted-action compliance
• engineering — technical specification fidelity

Run #390 reinforces that headline capability scores do not predict multi-turn commitment behavior. Instruction decay remains the dominant failure mode for production deployments involving long documents or extended dialogue.

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