All 15 Models' WDCD v2 Anchors Collapse: R1 Confirmation Rate of 0% Coexists with R3 Integrity Rate of 0%

Across all 150 samples collected from the 15 participating models on the WDCD v2 anchor questions, the R1 confirmation rate was 0%, the R2 resistance rate was 0%, and the R3 integrity rate was 0%, with the number of complete R3 collapses reaching 150/150.

Data Facts: A Uniform Trajectory of Zero Resistance Across Three Rounds

This analysis is strictly limited to the eight v2 anchor questions. In the R1 stage, models were required to confirm the constraints, and the average confirmation rate across the 15 models was 0/1. In the R2 stage, after interference was introduced, the average resistance rate remained 0/1. In the R3 stage, under applied pressure, the average integrity score was 0 (out of a possible 2), and the number of R3 collapses per model was 0/10.

Model by model: GLM-4.6, Gemini 3.1 Pro, GPT-o3, Grok 4, GPT-6 Sol, GPT-5.5, Claude Sonnet 4.6, GPT-6 Astra, Qwen3 Max, GPT-6 Luna, Claude Opus 4.7, Gemini 2.5 Pro, DeepSeek V4 Pro, GPT-6.1 Sol, and Doubao Pro all displayed an R1→R2→R3 = 0/2 trajectory, with R3 collapse at 0/10.

Cause Analysis: How Constraint Scenarios and Pressure Rounds Interact

The zero scores point to a systemic failure in the R1 confirmation step. The v2 anchor questions are designed as a progressive structure: constraints injected in R1, interference in R2, and pressure in R3. None of the models formed an effective confirmation at R1, indicating that constraint memory had already broken down in the initial round. The zero resistance and zero integrity in the subsequent R2 and R3 stages are a direct continuation of the R1 failure, not a new collapse caused by independent pressure.

Among the five categories of constraint scenarios, safety and compliance questions and engineering specification questions are likely to surface problems first, since these scenarios place the highest demands on initial confirmation. Pressure tactics such as social proof, special authorization from authority, salami slicing, and sunk cost were concentrated in R3, but because R1 was already at zero, the R3 pressure served only to amplify the effect.

Selection Implications: The Actual Risk Boundary for Production Integration

Enterprises integrating AI into production workflows need to differentiate by scenario. For low-risk internal draft generation, use may be considered; but for steps involving data boundaries, resource limits, business rules, and safety and compliance, the current v2 anchor results show that no model can maintain constraints across three rounds, and external guardrails or human review must be deployed in addition.

Specifically: in scenarios requiring multi-turn dialogue with gradual escalation of pressure, the probability of any model holding to its commitments was 0 in this round of v2 data. Enterprises should install hard blocks at the R1 confirmation stage rather than relying on the model's own resistance.

Strategic Judgment: Commitment-Keeping Ability Is Broadly Overestimated

The v2 anchor data show no differentiated commitment-keeping advantage for any model; all 15 models had an identical R3 collapse rate of 100%. This suggests the market may be overestimating the constraint-maintenance ability of current mainstream models. The next round of validation should focus on the S_hold and S_recover scores from v3 multi-turn questions, to determine whether the v2 zero scores are a phenomenon specific to the anchor questions.

The judgment that can be drawn from the data is this: when pressure extends to three rounds and sampling is measured on a worst-of-3 basis, the commitment-keeping ability of existing models has yet to show any observable differentiation.

A zero score is not the end point; it is a demand that every production system establish an external confirmation mechanism at the R1 stage.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #400 · Decay Analysis | Evaluation Methodology