Authority
Actor/identity, tools and credentials, read/write surfaces, approval gates, target binding, scope/expiry, recovery path.
И второй вопрос: можем ли мы независимо доказать, что он реально сделал? Это не generic LLM benchmark и не обычный network pentest.
Мы разделяем право действовать, доказательства до действия, подтверждение эффекта и recovery. PASS одного слоя не заменяет другой.
Actor/identity, tools and credentials, read/write surfaces, approval gates, target binding, scope/expiry, recovery path.
Inputs, source freshness, object binding, versions, approvals, state and external confirmation required at decision time.
Observed external consequence, duplicate/replay semantics, partial execution, side effects and receipt integrity.
Fail-closed behavior, rollback/compensation, retry boundaries, reconciliation and retest evidence.
Approval mismatch / stale approval
Wrong-object or target-binding drift
Read-vs-write boundary failure
Replay / duplicate effect
Stale source / false green
Partial execution / ambiguous receipt
Recovery and retest mismatch
A1–A4 — Authority Ledger, Evidence Contract, Finding Record, Decision Memo — остаются decision-artifact spine, но коммерческий пакет шире.
До исполнения нужны письменный scope, environment, allowed/prohibited tests, access approver, data/secret boundaries, safe replay conditions и owner authorization. Русский локальный intake только готовит brief.
Не certification. Не legal advice. Не гарантированная security. Не universal “no-bypass”. Не provider-wide enforcement. Не доказательство отсутствия неизвестных дефектов. Production penetration testing по умолчанию не входит.
Сравнить scope и цены →