TWO SAFETY APPROACHES : On the Stochastic Envelope and the Deterministic Gate — A Comparative Case Study of Anthropic's Responsible Scaling Policy and ST-T1024 A1–A17
There are two ways to make an AI system safe. One is to contain a stochastic process —measure it, mitigate it probabilistically, and govern it through institutions of trust. The other is to gate a deterministic process — bound it by construction, verify it mechanically, and govern it through signed declarations. These are not two grades of the same thing. They are two different theories of what safetyis. This paper presents both, on the evidence each has produced, as two cases. Case One isAnthropic's Responsible Scaling Policy, the Claude Constitution, and the interpretabilityprogramme that underwrites them. Case Two is ST-T1024 A1–A16, extended by the newlypublished A17, which moves the safety question from the individual agent to the collective. The paper does not argue that one case refutes the other. It argues that the two casesanswer different questions, and that the difference is visible in what each can and cannot certify. Case One can certify a process — that a lab measured, reported, and mitigated. Case Two can certify a system — that a declared class was respected or refused. The first is governance by disclosure. The second is governance by construction. And A17 — the most recent ST-T1024 extension — is the case that makes the difference sharpest, because it asks the question the stochastic envelope cannot answer: can a collection of individually safe agents become collectively unsafe?
Authors
- M. TAALABI
- Team LLM-DIPLOMAT
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24429464
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00