Перейти к содержимому
TakoAPI
Назад к навыкам

validator-correlated-judgment

от andyxinweiminicloud

Coding Agents & IDEs
274Число загрузок по данным clawskills.sh
0Просмотры
0 лайковВойдите, чтобы поставить лайк навыку
—Ещё нет оценок
Ваша оценка

Кратко

Helps identify when multiple attestation validators share training data, model architecture, or organizational.

Описание

Что делает этот навык

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.

Когда его использовать

  • 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

Пример сценария

ВВОД

User asks: Analyze validator correlation for Validator-A, Validator-B, Validator-C attesting data-processor skill

АГЕНТ
  1. 1

    Collect training provenance disclosures for each validator, noting base model and fine-tuning dataset

  2. 2

    Run behavioral correlation test on 50 edge-case skills, computing pairwise agreement rates against the independence baseline

  3. 3

    Calculate evasion transferability by testing whether skills evading Validator-A also evade Validator-B and Validator-C

  4. 4

    Compare reasoning path structure across validator pairs if evaluation traces are published

  5. 5

    Produce effective independent validator count and assign correlation verdict: INDEPENDENT, WEAKLY-CORRELATED, CORRELATED, or MONOCULTURE

ВЫВОД

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

Установка

clawhub install andyxinweiminicloud/validator-correlated-judgment

Выполните команду CLI или вставьте промпт в Claude Code, Cursor, Codex — любой кодинг-агент, которым пользуетесь.

Ссылки