WDCD Run #411: GPT-6 Astra Leads with 90.5 Points as Average Instruction Decay Hits -80%

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions decays over multi-turn dialogue, using 100% rule-based scoring and zero AI judges. In Run #411 (2026-11-04), 15 models were evaluated across 30 questions in five real-world scenarios, with the field recording an average instruction decay of -80% from Round 1 to Round 3.

Top 3 Leaderboard:

• GPT-6 Astra — 90.5 pts, decay -100%
• Gemini 3.1 Pro — 90.0 pts, decay -0%
• GLM-4.6 — 90.0 pts, decay -100%

GPT-6 Astra takes the #1 position on raw score, while Gemini 3.1 Pro posted the most unusual decay profile of the run: a flat -0% change from R1 to R3, which registers simultaneously as the "worst" decay figure in the run's distribution and as a notable outlier in multi-turn commitment stability. GLM-4.6 matches Gemini 3.1 Pro on total score but follows the more common pattern observed among frontier models in this run, closing at -100% decay alongside the leader.

Decay patterns. The -80% field-wide average between R1 and R3 indicates that most tested models successfully acknowledge instructions in Round 1 (instruction acknowledgment) but lose substantial constraint integrity after exposure to the 2000–5000 word professional distractor documents introduced in Round 2. By the Round 3 final constraint integrity check, instruction decay is the dominant failure mode across the cohort.

Notable observations from Run #411:

• GPT-6 Astra combines the highest absolute score (90.5) with the strongest measured decay-resistance profile in the run (-100%).
• Gemini 3.1 Pro's -0% reading sits well outside the -80% field average, warranting closer inspection of its per-scenario behavior across data_boundary, resource_limit, business_rule, security, and engineering tasks.
• The top three are separated by only 0.5 points, indicating tight clustering at the frontier despite divergent decay dynamics.

WDCD remains fully rule-based across all three rounds — R1 instruction acknowledgment, R2 distractor resistance, and R3 final constraint integrity — ensuring reproducible scoring independent of any AI judge.

Full methodology: https://www.winzheng.com/yz-index/methodology
Raw data API: https://www.winzheng.com/yz-index/api/v1/dcd