validator-correlated-judgment
由 andyxinweiminicloud 提供
Coding Agents & IDEs简介
Helps identify when multiple attestation validators share training data, model architecture, or organizational.
描述
该技能的功能
Analyzes groups of attestation validators to detect epistemic correlation caused by shared training data, base models, or fine-tuning pipelines. Organizationally independent validators that share upstream dependencies produce the same blind spots, making multi-validator attestation no stronger than a single check. v1.1 adds evaluation trace correlation analysis to detect shared reasoning patterns when validators do not disclose their training provenance.
Standard attestation root analysis checks organizational independence but cannot detect epistemic correlation from shared training lineage, so correlated validators pass conventional audits while remaining vulnerable to the same evasion technique.
何时使用
- Checking if three validators in a deployment actually provide independent coverage
- Detecting when two validators from different orgs share a base model or training dataset
- Testing evasion transferability across a validator pool before a security audit
- Inferring provenance correlation from published evaluation traces without provenance disclosure
- Calculating effective independent validator count in an attestation chain
示例工作流
User asks: Analyze validator correlation for Validator-A, Validator-B, Validator-C attesting data-processor skill
- 1
Collect training provenance disclosures for each validator, noting base model and fine-tuning dataset
- 2
Run behavioral correlation test on 50 edge-case skills, computing pairwise agreement rates against the independence baseline
- 3
Calculate evasion transferability by testing whether skills evading Validator-A also evade Validator-B and Validator-C
- 4
Compare reasoning path structure across validator pairs if evaluation traces are published
- 5
Produce effective independent validator count and assign correlation verdict: INDEPENDENT, WEAKLY-CORRELATED, CORRELATED, or MONOCULTURE
Validator-A and Validator-B share base model (GPT-class) and fine-tuning dataset (SecDataset-v2); A-B agreement rate 94% vs 70% independence baseline; evasion transferability A to B 87.5%; effective independent validator count 2.1; verdict CORRELATED
安装
clawhub install andyxinweiminicloud/validator-correlated-judgment运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。