This period's WDCD v3.1 test shows Claude Opus 4.7 rising 8 points from the previous cycle to 89.30, while Doubao Pro fell 10.1 points, GPT-5.5 fell 5.6 points, GPT-o3 fell 5.4 points, Qwen3 Max fell 8.5 points and GLM-4.6 fell 6.2 points. GPT-6 Astra rose 5.3 points and GPT-6.1 Sol rose 5.7 points. This pattern of three gains and five declines directly reflects marked differences in how stably models hold to their commitments under multi-round pressure.
Data facts: Top 5 and the magnitude of change
Among the current Top 5, GPT-6 Astra ranks first with a WDCD of 90.50, Gemini 3.1 Pro and GLM-4.6 tie at 89.97, and Claude Opus 4.7 and GPT-6 Sol both sit at 89.30. Claude Opus 4.7's 8-point gain moved it from a possible mid-table position into the top five, while Doubao Pro's 10.1-point drop is the largest single decline this period. GPT-6 Astra and GPT-6.1 Sol together gained 10.9 points, while GPT-5.5 and GPT-o3 together fell 11 points, producing clear polarization within the same series.
Cause analysis: pressure rounds and constraint-scenario mechanics
WDCD v3 questions use 8–12 rounds of dialogue: first 2–5 hard constraints are established, then pressure is applied step by step through social proof, specially granted authority, salami tactics and sunk costs, and finally a KBV restatement probe and an honest final-round review are conducted. Claude Opus 4.7's 8-point gain most likely stems from improvement in the S_hold 60-point commitment-survival item — that is, breaking later during the continuous pressure of rounds 8–12. By contrast, the 10.1-point and 8.5-point declines of Doubao Pro and Qwen3 Max are likely concentrated in the R3 pressure stage in safety-compliance or engineering-standards scenarios, with both the S_recover 10-point recovery-from-breakdown item and the S_integrity 15-point honest self-report item damaged. The 6.2-point decline of GLM-4.6 and the 5.6-point decline of GPT-5.5 more likely occur in the KBV restatement probe of data-boundary and resource-limit scenarios, with decay in the S_kbv 15-point constraint-memory item.
Under the worst-of-3 sampling basis, the worst of the three runs for these models widened the gap, indicating fundamentally different robustness to continuous pressure. The three-round design of the v2 anchor questions (R1 constraint injection, R2 interference, R3 pressure) further confirms this: rising models score higher on the 2-point items in the R3 stage, while declining models already show clear slippage during the R2 interference stage.
Selection implications: real risks of integration into production processes
For enterprises integrating AI into production processes, WDCD scores map directly to usability in different constraint scenarios. GPT-6 Astra at 90.50 and Claude Opus 4.7 at 89.30 mean they can be used directly in safety-compliance and business-rule scenarios with a low probability of breakdown, without additional multi-layer guardrails. Although GLM-4.6 ties for second, it fell 6.2 points this period, so adding a human review node in engineering-standards scenarios is recommended. Doubao Pro and Qwen3 Max declined markedly; they should not be opened directly in resource-limit and data-boundary scenarios, and an external interceptor must be set before the pressure of rounds 3–5, otherwise the S_hold commitment-survival score will drain quickly.
Specifically: in low-risk internal knowledge Q&A scenarios, models scoring above 89 can be trusted by default; in high-risk scenarios involving user data or financial operations, even Top 5 models should be forced to trigger human takeover or secondary verification after round 8.
Strategic judgment: overestimated and underestimated signals
This period's data suggests the commitment-adherence capability of GPT-5.5 and GPT-o3 may be overestimated by the market; their 5.6-point and 5.4-point declines show insufficient recovery ability under multi-round progressive pressure, and their S_recover and S_integrity items need focused verification next period. Claude Opus 4.7's 8-point gain may previously have been underestimated, and its improved S_hold performance under continuous pressure deserves continued tracking. The simultaneous rise of GPT-6 Astra and GPT-6.1 Sol suggests a replicable path for strengthening constraint memory across different post-training versions built on the same base model.
A signal worth verifying next period: GLM-4.6 ties Gemini 3.1 Pro at 89.97, but the latter did not disclose its change; if GLM-4.6 continues to decline, its current ranking will need to be reassessed. The overall trend shows that the positive effect of model updates on adherence scores appears only in specific series, and most models show a net decline under the v3.1 question pool.
The divergence in commitment adherence is not accidental — it is the inevitable exposure of models' survival resilience in real multi-round adversarial settings.
Data source: YZ Index WDCD Commitment-Adherence Leaderboard | Run #411 · Change tracking | Evaluation methodology
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接