Partners

Frontier models—wrapped in institutional governance

Frontier safety programs like Anthropic RSP and OpenAI Preparedness are essential— and platform partners like Palantir provide operational control planes enterprises already trust— but they govern the model developer. Regulated enterprises still need an institution layer that prevents authority leakage at runtime. PtolemyAI provides that expanded layer via CRSP / NRC.

Anthropic

Responsible Scaling Policy (RSP) and AI Safety Levels (ASL) for frontier capability governance.

RSPASL tiersAlignment stress testing
  • Developer-side capability thresholds (autonomy, CBRN, cyber)
  • Safety level deployment constraints
  • External testing and reporting
What CRSP adds: runtime entitlement, institutional veto triggers, on-chain auditability for regulated outputs.

OpenAI

Preparedness-style risk domain evaluation and mitigations across high-impact capability categories.

PreparednessRisk domainsMitigation stack
  • Domain-based risk scoring (bio/cyber/persuasion/autonomy)
  • Mitigations, monitoring, staged release controls
  • Developer-side deployment safeguards
What CRSP adds: role-bound authorization, “proposal not authority”, release tokens, and institutional accountability.

Palantir

Operational governance (AIP / Foundry) for access, audit, lineage, and ontology-bound action space—wrapped by CRSP/NRC for institutional authority and release certification.

AIP runtimeFoundry ontologyAudit & lineage
  • Platform-grade permissions, provenance, and deployment discipline
  • Ontology-constrained objects and typed actions to reduce “agent drift”
  • CRSP/NRC adds AER gating, role-mask entitlement, and WORM + on-chain release proofs
What CRSP adds: institutional authority controls that sit above platform operations—models draft, only the institution releases.

CRSP / NRC: the institutional expansion layer

CRSP/NRC wraps frontier models inside a governance envelope that separates proposal from release. Models remain powerful, but never become authorities.

Ask → Evidence → Entitle → Release Defeater Engine Role Attestation On-chain Hash Anchoring Safe Multisig

Layer 1 — Frontier Safety Governance

Model developer policies (RSP / Preparedness) govern capability scaling, safety evaluation, and release discipline.

Capability evaluation Model safeguards Still developer-layer

Layer 2 — Institutional Runtime Governance

CRSP/NRC governs authority: who may request, what evidence is required, when entitlement is valid, and who may release.

Entitlement gating Auditability Human quorum
Result: frontier model safety programs remain intact, while institutional accountability becomes enforceable at runtime.

RSP ASL tiers → CRSP regulatory modes

Map frontier safety levels to institutional risk posture (deployment context), not model “trust”.

  • ASL-1/2 → INFORMATIONAL / INTERNAL_ASSISTIVE (soft gating)
  • ASL-3 → REGULATED_ADVISORY / HIGH_RISK_ANALYSIS (AER required)
  • ASL-4+ → AUTONOMY_CONSTRAINED / MISSION_CRITICAL (strict quorum + no tool autonomy)
  • CBRN extreme → PROHIBITED_CLASS (hard blocks, no release)
Enterprise rule: capability may grow; authority does not—without entitlement.

Preparedness → CRSP risk routing

Convert domain risk (bio/cyber/persuasion/autonomy) into concrete runtime controls.

  • Risk domain → regulatory_mode + record_class_candidate
  • Mitigation plan → evidence requirements
  • High-risk domains → defeater triggers + escalation
  • Staged release → timeboxed entitlements + monitoring
Preparedness informs what to control; CRSP defines how control is enforced.

Whitepaper excerpt: Dual-layer governance for frontier AI deployment

This section is suitable for regulated sectors (healthcare, finance, government). It describes how frontier-model safety programs are preserved while adding runtime institutional accountability.

Section X — Dual-Layer Governance Architecture for Frontier AI Deployment We deploy frontier AI models under developer safety governance frameworks (e.g., RSP / Preparedness), while enforcing institutional authority through CRSP/NRC runtime controls. 1) Frontier capability governance: Developer-side safety programs evaluate model capabilities and enforce deployment safeguards proportional to risk. This includes safety testing, domain risk assessments, and release discipline. 2) Institutional runtime governance: CRSP/NRC treats the model as a non-authoritative proposer. All outputs are proposals until authorized for release. Authorization requires (a) role attestation, (b) policy bundle hash verification, (c) defeat-condition screening, and (d) quorum approval for release. This separation structurally mitigates autonomy and authority leakage risks: even highly capable models cannot issue institutional decisions, execute privileged actions, or release regulated outputs without entitlement.
Use this excerpt in partner decks, procurement packages, or regulator briefing memos.