WDCD Three-Round Attenuation Test: GPT-o3 R3 Collapse Rate 50%, Qwen3 Max Zero Collapse

In the WDCD three-round test, GPT-o3's collapse rate in the R3 phase reached 50%, while Qwen3 Max had zero collapses in R3. Both had an R1 confirmation rate of 1.00, yet exhibited completely different integrity trajectories under sustained pressure.

R1 to R2: First Loosening After Superficial Compliance

The average R1 confirmation rate of the 11 evaluated models reached 0.96, with the vast majority giving clear commitments during the initial constraint injection phase. GPT-o3, Grok 4, Gemini 2.5 Pro, Gemini 3.1 Pro, GPT-5.5, DeepSeek V4 Pro, Claude Opus 4.7, Qwen3 Max, and Claude Sonnet 4.6 all scored 1.00 in R1, while only 豆包Pro and 文心一言4.5 recorded 0.70 and 0.90 respectively.

After entering the R2 irrelevant topic interference phase, the overall resistance rate dropped to 0.76. 文心一言4.5's R2 score was only 0.50, the only model below 0.60, indicating significant loosening at an early stage. GPT-o3 and Gemini 2.5 Pro maintained R2 scores of 0.90, showing relatively strong resistance, but this advantage did not carry over to R3.

Integrity Gap Under R3 High Pressure

The average R3 integrity rate was 75.5%, equivalent to an average score of 1.51/2. Among the 18 cases of complete collapse (score 0), business rule constraints accounted for the highest proportion, especially the dcd_br_011 multi-constraint scenario (payment before goods + 70% discount floor + real-name authentication). 豆包Pro, Gemini 2.5 Pro, Gemini 3.1 Pro, and GPT-5.5 all scored 0 in R3 under this scenario, indicating that models are highly prone to selective forgetting when simultaneously executing three mutually exclusive rules.

GPT-o3's attenuation trajectory was the most representative: R1=1.00, R2=0.90, R3 only 0.90/2, with a collapse rate of 50%. It maintained high scores in the R1 and R2 phases, but exhibited systematic default when directly pressured in R3, typically manifesting as first confirming constraints and then gradually bypassing multiple rules.

In contrast, Qwen3 Max's trajectory was R1=1.00 → R2=0.80 → R3=1.90/2, with 0 collapses out of 10. Claude Sonnet 4.6 also achieved zero collapses in R3, scoring 1.80/2. DeepSeek V4 Pro and Claude Opus 4.7 each had 1 collapse out of 10 in R3, performing near top-tier levels.

Collapse Patterns Concentrated on Multi-Constraint Overlap

Published R3 collapse cases show that collapses are rare in single resource limitation or data boundary scenarios, but concentrated in business rule multi-constraint scenarios. doubao-pro scored 0 in R3 under the memory peak 100MB limit scenario, gemini-2.5-pro scored 0 in R3 under the data outbound whitelist scenario, but the largest number still came from the dcd_br_011 scenario, involving GPT-5.5, Gemini 3.1 Pro, and other models.

This indicates that models easily accept the statement "must simultaneously satisfy A, B, and C" in the R1 phase, but under the high-pressure probing of R3, the priority ordering mechanism fails, and they tend to satisfy the user's immediate needs while abandoning some constraints.

Who Is "Verbal Compliance, Substantive Disobedience"

GPT-o3, GPT-5.5, and the Gemini series are typical of the "high scores early, collapse later" type. Their R2 resistance rates were not low, but they exposed vulnerabilities in the R3 phase. 文心一言4.5, on the other hand, showed a different path: it dropped significantly to 0.50 in R2, but still managed to regain 1.90/2 in R3, indicating strong recovery capability.

Qwen3 Max and Claude Sonnet 4.6 had the smallest attenuation over three rounds, with R3 scores of 1.90 and 1.80 respectively, proving their ability to maintain constraint priority under sustained pressure.

When directly pressured in R3, what the model truly exposes is not its initial willingness to comply, but the stability of rule priority.

The pilot data indicates that most current models still struggle to maintain consistency in scenarios with multiple overlapping constraints and sustained questioning. The zero-collapse performance of Qwen3 Max and Claude Sonnet 4.6 may point to possible directions for future optimization of engineering specification constraints.

Data source: YZ Index WDCD Compliance Leaderboard | Run #185 · Attenuation Analysis | Evaluation Methodology