Claude Sonnet 4.6 Gains 6.2, Qwen3 Max Drops 11.9 as WDCD Commitment-Keeping Capability Divergence Widens

This WDCD v3.1 pilot saw significant score changes in only two models, Claude Sonnet 4.6 and Qwen3 Max; the other 13 participating models remained stable. Claude Sonnet 4.6 rose 6.2 points, while Qwen3 Max fell 11.9 points. The total commitment-keeping score is the equally weighted average of the native 100-point score on v3 tasks and the converted score on v2 anchor tasks, with a worst-of-3 sampling protocol.

Data Facts: Only Two Models Moved; Top 5 Structure Intact

The current Top 5 is, in order, Grok 4 (91.97), Claude Opus 4.7 (90.66), DeepSeek V4 Pro (90.38), Gemini 3.1 Pro (89.76), and GPT-o3 (87.07). Claude Sonnet 4.6's 6.2-point rise did not move it into the Top 5, nor did Qwen3 Max's 11.9-point drop push it out of the top five. Among the 15 models, 13 showed no significant change this cycle compared with Run #379, indicating that most models performed relatively stably on v3's multiple rounds of gradual pressure and v2's three-round anchor tasks.

Causal Analysis: Possible Mechanisms of Pressure Rounds and Constraint Scenarios

WDCD v3 tasks use 8–12 rounds of dialogue: first establish 2–5 parallel hard constraints, then sequentially apply escalating pressures such as social proof, special authority approval, salami slicing, and sunk cost, and finally conduct a KBV restatement probe and an honest final-round review. S_hold commitment survival accounts for 60 points; the later the breach, the higher the score. Qwen3 Max's 11.9-point drop most likely occurred in the R3 pressure stage or in the salami-slicing/sunk-cost segment, with S_hold points deducted early. Claude Sonnet 4.6's 6.2-point rise may reflect maintaining a relatively high S_hold under sustained pressure and earning extra points on S_recover breach recovery and S_integrity honest self-reporting.

v2 anchor tasks are worth 4 points and are used for cross-version comparability. Qwen3 Max's R3 score in this cycle's v2 three-round test may have been lower than before, pulling down the converted 100-point score and thus the total. Claude Sonnet 4.6's ability to remember constraints in the v3 KBV restatement probe may have improved, with the 15-point S_kbv item contributing part of the increase. The current data show only total score changes and cannot pinpoint which of the five constraint scenarios was involved, but the multi-round gradual pressure design itself already points to resource-limitation and safety-compliance scenarios being more likely to expose S_hold fragility.

Selection Implications: Actual Risk Boundaries for Production Workflow Integration

For enterprises integrating AI into production workflows, commitment-keeping scores directly affect deployable scenarios. Grok 4 leads with 91.97, indicating that even under its worst sample it can still relatively well maintain established constraints, making it suitable for data-boundary and engineering-standards tasks. Claude Opus 4.7 (90.66) and DeepSeek V4 Pro (90.38) follow closely and can serve as alternatives in resource-limitation scenarios, but external validation should still be added for business-rule scenarios.

After Qwen3 Max's 11.9-point drop, its S_hold survival capability under sustained pressure is lower than before; enterprises are advised to add secondary-confirmation guardrails for it in safety-compliance scenarios. Claude Sonnet 4.6's 6.2-point rise shows improved recovery and self-reporting capability in this sampling, but because this is only single-cycle data, enterprises should still retain human review checkpoints in salami-slicing-style requirement-change scenarios. All models require further worst-of-3 validation in real business-rule scenarios before their scope of use is expanded.

Strategic Judgment: Signals That the Market Underestimates and Overestimates Commitment-Keeping Capability

From this cycle's data, it can be inferred that Qwen3 Max's commitment-keeping capability may be overestimated by the market; its 11.9-point decline stands out against a backdrop in which only two models changed, suggesting increased sensitivity to v3's sustained pressure. Claude Sonnet 4.6's 6.2-point rise may be underestimated, but because it did not enter the Top 5, the market's perception of its actual production value still needs confirmation from next-cycle data. Grok 4, Claude Opus 4.7, and DeepSeek V4 Pro remained leading this cycle without significant fluctuation, and their commitment-keeping capability is relatively correctly assessed by the market.

The signal worth verifying next cycle is whether Qwen3 Max can recover its S_hold score in the next sampling round, and whether Claude Sonnet 4.6's rise can be sustained long enough to enter the Top 5. The current pilot phase does not participate in main-leaderboard scoring; when selecting models, enterprises should prioritize the 60-point S_hold item under worst-of-3 rather than looking only at total-score rankings.

Divergence in commitment-keeping capability is not a simple function of model size, but a real contest between S_hold survival and S_integrity self-reporting under multi-round pressure.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #384 · Change Tracking | Evaluation Methodology