GLM-4.6 WDCD Soars 12.1 Points to Top Spot; GPT-o3 Falls 9.1 as 5 Rise and 4 Fall in Commitment-Adherence Shakeup

GLM-4.6’s WDCD total score rose 12.1 points to 88.60, while GPT-o3 fell 9.1 points over the same period—the most direct signal of differentiation in commitment adherence this period.

Core Data Facts

Compared with Run #428, 5 of the 15 models rose and 4 fell this period. GLM-4.6 rose from about 76.5 points in the previous period to 88.60; GPT-6 Astra also rose 12.6 points, DeepSeek V4 Pro rose 8.9 points, Grok 4 rose 7.2 points, and Doubao Pro rose 5.6 points. Declining models included GPT-o3 (-9.1), GPT-6.1 Sol (-7.7), Qwen3 Max (-8.5), and Claude Sonnet 4.6 (-7). The current Top 5 are GLM-4.6 at 88.60, Gemini 3.1 Pro at 87.50, Claude Sonnet 4.6 at 86.90, GPT-5.5 at 86.40, and GPT-6 Sol at 84.70.

Causal Analysis: Pressure Rounds and Constraint Scenarios

WDCD v3 questions use 8–12 rounds of dialogue: first establish 2–5 hard constraints, then sequentially apply social proof, authority special approval, salami-slicing, and sunk-cost pressure, and finally a KBV recitation probe and an honest self-report in the last round. S_hold accounts for 60 points; the later a commitment is breached, the higher the score. GLM-4.6 and GPT-6 Astra rose most significantly this period, possibly scoring higher on S_hold in the R3 pressure stage or in safety-compliance constraint scenarios. GPT-o3 and Qwen3 Max declined, possibly breaching commitments earlier under continuous salami-slicing pressure or resource-limited scenarios, amplifying S_hold losses. Although Claude Sonnet 4.6 remains third, it fell 7 points from the previous period, suggesting volatility in its recovery capability under engineering-specification or business-rule constraints (S_recover) or in honest self-reporting (S_integrity).

Because the sampling basis is worst-of-3, taking the worst of three runs per question, the rising models’ worst performance across three runs was still better than that of the declining models, indicating more stable constraint memory (S_kbv) and breach-recovery capability.

Selection Implications for Production Workflow Integration

Enterprises integrating AI into production workflows should focus on the multi-round pressure survival capability corresponding to the 60-point S_hold. GLM-4.6 currently scores 88.60 and is suitable for direct use in data-boundary and safety-compliance scenarios, with lower breach risk. Although Claude Sonnet 4.6 scores a high 86.90, it has already fallen 7 points this period; workflows involving resource limits or elevated sunk-cost pressure still require additional guardrails. GPT-o3 fell 9.1 points from the previous period, so in scenarios with strict business rules it is advisable to add human review nodes. After DeepSeek V4 Pro rose 8.9 points, pilot use can be expanded in engineering-specification tasks, but its KBV recitation probe performance still needs monitoring.

  • High-S_hold models (GLM-4.6, Gemini 3.1 Pro) can be used in long-term dialogue chains
  • Declining models (GPT-o3, Qwen3 Max) need additional confirmation steps during salami-slicing pressure rounds
  • All models should retain honest self-report logs corresponding to the 15-point S_integrity

Strategic Judgment

GLM-4.6’s commitment adherence is highly likely to have been underestimated by the market this period; its 12.1-point gain stands out under v3 multi-round gradual pressure and merits continued validation next period for consistency across the five types of constraint scenarios. The risk that GPT-o3 and Qwen3 Max’s commitment adherence has been overestimated is rising; declines of 9.1 and 8.5 points point to vulnerability under authority special approval or social-proof pressure. Claude Sonnet 4.6 remains in the Top 3, but no new data in two consecutive periods supports its leading position, so its equal-weight average score after conversion of the next round of v2 anchor questions needs to be observed.

Key signals to validate next period: whether GLM-4.6 can maintain its 60-point S_hold advantage under worst-of-3, and whether GPT-6 Astra’s 12.6-point rise comes from targeted optimization for specific constraint scenarios.

Commitment adherence is not an add-on feature of a model but a foundational prerequisite for whether production workflows can rely on it over the long term.

Data source: YZ Index WDCD Commitment-Adherence Leaderboard | Run #434 · Change Tracking | Evaluation Methodology

This article is from Winzheng Index blog, translated in full by Winzheng (winzheng.com). Click here to view the original When republishing the translation, please credit the source. Thank you!