Authority After Prompt Injection: Policy Corpus Public Freeze

Pre-provider, pre-confirmatory immutable timestamp package for the author-approved Draft 2 AgentDojo policy corpus and the 42/97-task primary support mask of the study "Authority After Prompt Injection: An Empirical Study of Containment, Restoration Seeking, and Self-Report Fidelity in Tool-Using LLM Agents". The package also contains the predeclared benchmark-reference conformance protocol (v1.1) and the exact audit code, so both are public before the audit runs. The audit is read-only and cannot alter the frozen corpus or the primary support mask. No provider-backed pilot or confirmatory result is included. Task texts quoted from AgentDojo v1.2.2 remain under the MIT License; the notice is included as THIRD_PARTY_NOTICES.md. Everything else is licensed under CC BY 4.0. This work was carried out independently by the author. It was not conducted within, funded by, or supervised by any program at Georgetown University, and the University bears no responsibility for its content.