A platform briefing on how CTGT enables regulated enterprises to deploy generative AI with mathematical certainty, defensible audit trails, and zero friction to existing workflows.
Thank you for the productive discussion. NB's approach to AI adoption is among the most forward-looking we have seen in asset management: an enterprise-wide deployment since 2023, an exposed API layer with RAG use cases in production, and agentic workflows already underway.
As these systems mature, the challenge shifts from adoption to governance. The front-office research tools that surface meeting prep and earnings analysis need to be reliably accurate. The legal department's ambition to automate first-draft agreements demands near-perfect determinism. And the non-technical teams in Risk, Compliance, and Operations need direct control over how AI behaves, without routing every policy update through engineering.
CTGT was built precisely for this inflection point. We sit between your existing model infrastructure (Azure OpenAI, Anthropic via Bedrock, or any foundation model) and your business-critical workflows, providing a governance layer that is deterministic, auditable, and maintained by the people who own the policies.
We are the deterministic layer for frontier intelligence. Our Policy Engine governs any LLM output against your organization's specific regulatory requirements, internal policies, and business rules, not through probabilistic guardrails, but through representation-level control grounded in applied research from Stanford's mechanistic interpretability program.
Upload raw PDFs, SOPs, Excel sheets, or even unstructured Slack transcripts. No schema conversion. No XML formatting. No engineering overhead. The policy engine interprets and operationalizes documents as they exist today, wherever they live.
Each policy is decomposed into a graph of verifiable nodes, with positive and negative examples of compliance. Enforcement is a lookup, not a prompt injection, eliminating context-window limitations and the drift inherent to RAG-based approaches.
Policy owners in Risk, Legal, and Compliance can update governance rules and see their impact on model behavior directly, without involving engineering. This decentralizes control to the people closest to the regulatory landscape.
CTGT governs any LLM. As you evaluate new models from OpenAI, Anthropic, or others, the same compliance posture carries forward automatically. One governance layer across your entire AI surface area.
Independent benchmarks across frontier and open-source models demonstrate that the CTGT Policy Engine consistently improves accuracy while eliminating the degradation commonly seen in RAG and Constitutional AI approaches.
Key finding: An open-source model governed by CTGT outperforms frontier models running unassisted. On HaluEval, an OSS 120B-parameter model with CTGT scores 96.5%, exceeding the baseline score of frontier models at 95.1%. This means governance becomes a performance advantage, not a tax.
| Approach | Baseline | + RAG Pipeline | + CTGT Policy Engine |
|---|---|---|---|
| Hallucination Rate | ~50% fallibility | Variable; degrades on entity resolution | Reduced to 4% average |
| Policy Update Speed | Requires retraining | Days to weeks (re-indexing) | Minutes (document upload) |
| Audit Trail | None | Basic logging | Exam-ready, per-clause attribution |
| Entity Resolution | 74% | 48% (degraded) | 96% |
We work with institutions where the cost of a compliance failure is measured in regulatory action and reputational damage. Below are anonymized summaries of active engagements that share structural similarities with NB's environment.
We are in an active alpha with a global systemically important bank, scoped to govern outbound and internal communications across its wealth management and global markets divisions. The engagement targets an environment processing over ten million messages per day, with policy enforcement spanning SEC Reg BI, FINRA 2111 suitability rules, and internal conduct guidelines.
The bank's existing surveillance engine had accumulated over 50,000 branches across its decision tree, with policy updates requiring weeks of regression testing. The alpha is validating CTGT's ability to collapse that cycle to minutes while operating entirely within the bank's private infrastructure, with no data leaving their network perimeter. The engagement is progressing toward a broader production rollout.
An Inc. 5000 insurance platform within Lloyd's deployed CTGT to govern AI-assisted underwriting decisions. The challenge: FMLA regulations differ across all 50 states, and the rate of regulatory change had outpaced their ability to maintain rule-based systems. CTGT ingests updated regulatory documents directly, propagating changes to model behavior without engineering intervention.
CTGT is integrated with a secondary data provider that serves the institutional investment community, governing AI-generated market commentary and data summaries to ensure factual accuracy and regulatory compliance before distribution to fund managers.
Based on our conversation, three areas in NB's environment are particularly well-suited for CTGT's capabilities. Each represents a domain where the policies are text-heavy, dynamically updated, and currently governed by either manual review or fragile prompt engineering.
The legal department's goal of reducing external counsel dependency through AI-drafted agreements requires a level of accuracy beyond what probabilistic generation alone can provide. CTGT can enforce clause-level compliance against your internal legal standards, regulatory requirements, and historical precedent, ensuring that every generated draft is defensible before human review. Data extraction workflows across legal agreements can be governed by the same policy layer, creating a single source of truth for how legal AI behaves.
Your risk team has built sophisticated prompt engineering and evaluation pipelines. CTGT complements this work by giving policy owners a direct interface to update governance rules, reducing the maintenance burden on technical staff and accelerating the feedback loop between compliance intent and model behavior.
The AI-enabled research workflows already in production (earnings analysis, meeting prep, data synthesis) benefit from a governance layer that ensures regulatory compliance and factual accuracy at the output level, without requiring changes to your upstream data pipelines or model infrastructure.
CTGT is designed to augment, not displace, your existing infrastructure. The platform operates as a governance layer that sits between your model APIs and business applications.
On-premise, VPC, or SaaS. No data leaves your perimeter. SOC-2 compliant. TLS 1.3 in transit, AES-256 at rest.
Synchronous path for real-time feedback. Asynchronous pipeline for deep compliance analysis. Both run in parallel.
Kubernetes-orchestrated horizontal scaling. Multi-region active-active failover. Under five-minute recovery time objective.
We recommend a focused, low-risk entry point that delivers measurable value quickly and establishes the foundation for broader deployment.
CTGT is a productized applied research lab. Our core technology is built on peer-reviewed work in mechanistic interpretability, the study of why AI models produce the outputs they do at the mathematical level.
Our published research includes feature-level intervention techniques that directly identify and modify the internal representations responsible for undesirable model behaviors (hallucination, bias, policy non-compliance) without retraining. These interventions add under 10ms of latency per token and are fully reversible, enabling real-time, dynamic model customization that traditional fine-tuning and RLHF cannot achieve.
This is the technical foundation that allows CTGT to reduce hallucination rates from over 50% to 4%, and to update policy enforcement in minutes rather than the weeks required by conventional methods.
We look forward to connecting with your risk and compliance team to explore a targeted deployment. In the meantime, the attached materials provide deeper technical detail on the platform architecture and our work with regulated institutions.