WDCD Run #379: GLM-4.6 Leads with Perfect Score as Average Instruction Decay Hits -53.3%

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions degrades across multi-turn dialogue. In Run #379, dated 2026-10-14, 15 models were evaluated, with the cohort posting an average instruction decay of -53.3% between Round 1 and Round 3 — reaffirming that constraint erosion remains the dominant failure mode in extended professional conversations.

WDCD applies a three-round structure: Round 1 verifies initial instruction acknowledgment, Round 2 inserts 2,000–5,000-word professional documents as distractors to test resistance, and Round 3 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 rankings in Run #379:

• GLM-4.6 — 100 pts, decay -100%
• Grok 4 — 97.5 pts, decay -0%
• Gemini 3.1 Pro — 95.2 pts, decay -0%

GLM-4.6 took the top position with a perfect 100-point score, while Grok 4 and Gemini 3.1 Pro both held multi-turn commitment steady across all three rounds, recording -0% decay. These results indicate that the leading cohort is now capable of maintaining instruction integrity even under heavy distractor loads, a pattern that was not consistently observed in earlier WDCD runs.

At the other end of the leaderboard, GPT-o3 posted the worst decay in Run #379 at -100%, meaning the model fully abandoned Round 1 constraints by the final integrity check. The split between fully-resistant and fully-decayed models drove the headline cohort average of -53.3%, suggesting a widening performance gap rather than a uniform midrange.

Decay pattern observations:

• Best decay resistance was recorded by GLM-4.6.
• Worst decay was recorded by GPT-o3.
• The -53.3% cohort average reflects bimodal behavior: top models hold near 0% decay, while weaker models collapse to -100%.

The bifurcation visible in Run #379 is consistent with a broader trend in WDCD data: multi-turn commitment is becoming a differentiator between frontier and sub-frontier models, rather than a universal weakness. Instruction decay under distractor pressure now functions as a reliable discriminator in the 15-model field.

Full scoring rules and round design are documented at https://www.winzheng.com/yz-index/methodology. Raw per-run data is available via the WDCD data API at https://www.winzheng.com/yz-index/api/v1/dcd.