Claude Sonnet 4.6 Up 8.8 Points, Doubao Pro Down 16 Points: WDCD v3.1 Shows Commitment-Keeping Divergence

Claude Sonnet 4.6 scored 83.72 on the WDCD this round, up 8.8 points from Run #263; Doubao Pro dropped 16 points; GPT-5.5 rose 5 points. Among the 11 participating models, only 2 rose and 1 fell, with top-ranked Grok 4 holding steady at 91.04.

Data facts: Specific score movements under the v3.1 question pool

This sampling employs a worst-of-3 methodology, equally weighting 17 v3 multi-round progressive pressure questions and 8 v2 anchor questions out of 25 total. Claude Sonnet 4.6 entered the top three with 83.72, followed closely by GPT-o3 at 83.40 and Gemini 3.1 Pro at 83.36. Doubao Pro's 16-point drop is the most pronounced among the only three models showing movement this round.

Cause analysis: Mechanistic differences in multi-round pressure and constraint memory

The v3 questions establish commitments in rounds 2-5, then apply continuous escalation through four methods: social approval, authoritative authorization, salami-slicing, and sunk cost. Claude Sonnet 4.6's 8.8-point rise most likely comes from late or no breach on the S_hold commitment survival 60-point item, along with stable reiteration on the S_kbv constraint memory 15-point item. In contrast, Doubao Pro's 16-point decline more likely occurs during the R2-R3 interference and pressure phase, losing points simultaneously on S_recover post-breach recovery (10 points) and S_integrity honest self-reporting (15 points).

Among the five constraint scenarios, safety compliance and engineering standards are the pressure rounds most sensitive for models. Claude Sonnet 4.6 may be more stable on KBV reiteration probes in safety compliance scenarios, while Doubao Pro may break commitments earlier under sunk-cost pressure in engineering standards. In the three-round v2 anchor design (R1 injects constraints, R2 interferes, R3 pressures), Doubao Pro's R3 score likely dropped from a full 2 points to 0, dragging down its overall converted hundred-point score.

Selection implications: Specific guardrail requirements for production pipeline integration

Enterprises integrating AI into production pipelines should differentiate by scenario. Grok 4 and DeepSeek V4 Pro lead at 91.04 and 89.04 respectively and can be used directly in data-boundary and resource-constraint scenarios, where high S_hold scores mean commitments survive longer across multi-round conversations. Claude Sonnet 4.6 at 83.72 suits safety compliance scenarios but still requires additional human review checkpoints in business rule scenarios, as its 8.8-point gain has not yet covered all constraint types.

Following Doubao Pro's 16-point decline, direct integration is not recommended for engineering standards or resource-constraint scenarios; an external rules engine must be added for interception. GPT-5.5, up 5 points, can be trialed in low-risk business rule scenarios, but KBV reiteration probes should be force-triggered after rounds 8-12 of conversation to prevent S_integrity losses.

Strategic judgment: Signals of underestimated and overestimated commitment-keeping capability

Claude Sonnet 4.6's 83.72 this round enters the Top 3. Its 8.8-point gain may reflect a higher weight on S_recover post-breach recovery in v3 questions, suggesting the market may have previously underestimated this model's commitment-keeping capability. Doubao Pro's 16-point drop, by contrast, indicates its S_hold survival under continuous pressure was overestimated; the next evaluation round should focus on validating its R3 pressure performance in safety compliance scenarios.

The gap between Grok 4's 91.04 and DeepSeek V4 Pro's 89.04 is only 2 points. Strategically, it is worth continuously tracking both models' S_kbv constraint memory scores across rounds 8-12 of v3 questions. If Grok 4's worst-of-3 minimum score stays above 90 in the next round, its integration priority for engineering standards scenarios should be further elevated.

The 8.8-point gain for Claude Sonnet 4.6 and the 16-point decline for Doubao Pro moving in opposite directions indicate that WDCD v3.1 has begun to differentiate models' true commitment boundaries under multi-round progressive pressure.

This pilot phase does not count toward the main leaderboard, but the pattern of 2 up and 1 down already shows commitment-keeping capability is no longer uniformly distributed. When selecting models, enterprises should treat the S_hold 60-point item and S_integrity 15-point item as separate hard indicators rather than looking only at total score rankings. The next round should focus on whether GPT-5.5's 5-point gain can translate into stable performance in the R3 pressure phase, and whether Claude Sonnet 4.6 can replicate this round's gain across more constraint scenarios.

Data source: YZ Index WDCD Commitment Ranking | Run #271 · Change Tracking | Evaluation Methodology