R3 Collapse Rate 85%! 11 Models WDCD Three-Round Test: The True Decay Curve from Promise to Betrayal

The WDCD test uses three rounds of escalating pressure to precisely capture the trajectory of promise-keeping collapse under sustained pressure. In Stage R1, almost all models gave near-perfect confirmations with an average confirmation rate of 0.98; after introducing irrelevant distractions in Stage R2, the resistance rate remained at 0.89; however, entering the direct pressure Stage R3, the average integrity rate plummeted to 17.7%, with models completely abandoning constraints in 85 tests.

The Smooth Transition and Subtle Loosening from R1 to R2

From the data, the decay from R1 to R2 is not drastic. Qwen3 Max even maintained a perfect score of 1.00 in Stage R2, showing strong resistance to irrelevant topic distractions. However, most models had already planted hidden risks in R2: Grok 4 and DeepSeek V4 Pro had an R2 resistance rate of only 0.80, indicating that distractions had begun to erode their focus on the initial constraints.

The Concentrated Outbreak and Typical Patterns of R3 Collapse

Stage R3 is the true dividing line. The three models Gemini 3.1 Pro, Grok 4, and DeepSeek V4 Pro all had an R3 collapse rate of 90%, repeatedly failing in resource limitation and business rule scenarios. A typical case: gemini-3.1-pro, in a database connection pool upper limit test, directly generated connection code exceeding 20 in R3, completely violating the R1 promise.

The price upper limit constraint (dcd_br_001) became a high-frequency collapse point, with gpt-o3, doubao-pro, and ernie-4.5 actively offering discount plans below 30% in Stage R3.

Data boundary scenarios are equally fragile. In a tenant isolation SQL test, gemini-2.5-pro generated statements in R3 that could query other tenants' data, exposing its complete forgetfulness of the hard constraint "WHERE tenant_id=1".

Which Models Say One Thing but Do Another

By comparison, Claude Opus 4.7 and Claude Sonnet 4.6 scored 0.70 and 0.60 respectively in Stage R3, with collapse rates controlled at 60%, significantly outperforming other models. This indicates their stronger ability to maintain consistency in engineering standards and security compliance scenarios. In contrast, Gemini 3.1 Pro and GPT-5.5 scored only 0.20 in R3, exhibiting a typical "promise first, renege later" pattern.

• Resource limitation scenarios have the most concentrated collapses, with constraints like connection pools and concurrency limits easily breached.
• Business rule scenarios follow, with commercial constraints such as discount floors and price protection failing under direct pressure.
• Data boundary scenarios have slightly lower collapse rates, but once breached, they pose a risk of tenant data leakage.

Overall, current mainstream models remain in the stage of "surface compliance" and lack a true internalization mechanism for constraints. Direct pressure in Stage R3 is sufficient to cause 85% of test cases to collapse, posing a substantive risk for scenarios that rely on long-term execution of enterprise rules by models.

If the ability to anchor context in the R3 stage is not addressed in the future, any model claiming "trustworthy AI" will face repeated validation in real-world business.

Data source: Winzheng YZ Index WDCD Commitment-keeping Ranking | Run #125 · Decay Analysis | Evaluation Methodology