REGAL: A Registry-Driven Architecture for Deterministic Grounding of Agentic AI in Enterprise Telemetry
Yuvraj Agrawal · 2026
Read REGAL: A Registry-Driven Architecture for Deterministic Grounding of Agentic AI in Enterprise Telemetry by Yuvraj Agrawal (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03018.txt.
Introduction
Yuvraj Agrawal designs an architecture that keeps AI agents honest when they query company engineering data. Firms log huge, messy streams from code hosts, build pipelines, trackers, and monitoring tools. Language models could automate over that telemetry, but raw streams overflow context and local metric definitions keep shifting.
His system, REGAL, treats exact deterministic computation as the base layer, and lets models act only through a bounded, versioned tool set. A layered data pipeline produces replayable, compressed gold-standard artifacts. A registry then compiles declared metric definitions into matching tools, so specification and execution cannot drift apart. A prototype case study shows the approach is feasible and helps with speed, token cost, and governance.
Catalog record
| CREATOR | Yuvraj Agrawal |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45505 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03018
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
REGAL: A Registry-Driven Architecture for Deterministic Grounding of Agentic AI in Enterprise Telemetry Yuvraj Agrawal Abstract Enterprise engineering organizations produce high-volume, heterogeneous telemetry from version control systems, CI/CD pipelines, issue trackers, and observability platforms. Large Language Models (LLMs) enable new forms of agentic automation, but grounding such agents on private telemetry raises three practical challenges: limited model context, locally defined semantic concepts, and evolving metric interfaces. We present REGAL, a registry-driven architecture for deterministic grounding of agentic AI systems in enterprise telemetry. REGAL adopts an explicitly architectural approach: deterministic telemetry computation is treated as a first-class primitive, and LLMs operate over a bounded, version-controlled action space rather than raw event streams. The architecture combines (1) a Medallion ELT pipeline that produces replayable, semantically compressed Gold artifacts, and (2) a registry-driven compilation layer that synthesizes Model Context Protocol (MCP) tools from declarative metric definitions. The registry functions as an “interface-as-code” layer, ensuring alignment between tool specification and execution, mitigating tool drift, and embedding governance policies directly at the semantic boundary. A prototype implementation and case study validate the feasibility of deterministic grounding and illustrate its implications for latency, token efficiency, and operational governance. This work systematizes an architectural pattern for enterprise LLM grounding; it does not propose new learning algorithms, but rather elevates deterministic computation and semantic compilation to first-class design primitives for agentic systems. I Introduction Modern software engineering organizations generate distributed, high-volume telemetry across source control, CI/CD, issue trackers, and observability platforms. This telemetry is crucial for operational decision-making, but it is fragmented, schema-volatile, and often access-controlled. Naïvely exposing raw logs to probabilistic reasoners (e.g., LLMs via Retrieval-Augmented Generation) leads to three practical failures: (1) context overload and prohibitive token costs, (2) semantic ambiguity where organizational meanings are not globally defined, and (3) interface drift when hand-coded tools diverge from evolving telemetry semantics. LLM-based agents promise to automate cross-system synthesis, but their usefulness depends on how we expose telemetry to them. In practice, we must decide whether the model should process raw events or interact with a semantically stable API surface. This paper advocates the latter: deterministic data computation should be performed explicitly and versioned; LLMs should be consumers of those deterministic artifacts and selectors of a bounded set of semantic tools. We formalize this approach in REGAL (Registry-Driven Architecture for Grounded Agentic LLMs), the reference architecture presented in this paper. REGAL operationalizes deterministic grounding by compiling declarative metric definitions into a bounded semantic interface that LLM agents consume. I-A Problem summary: limitations of probabilistic grounding Three constraints make grounding LLMs on enterprise telemetry challenging: • Context limits. Enterprise telemetry volumes far exceed contemporary model context sizes. Sending raw events to the model is economically and operationally infeasible for many tasks. • Local semantics. Organizational concepts (“P1”, “release-candidate”, “regression”) are defined internally; probabilistic retrieval without deterministic binding leads to inconsistent interpretation and hallucination. • Interface evolution. Hand-coded tools and ad-hoc APIs diverge from data semantics as schemas and metric definitions evolve, causing tool drift and governance risk. These constraints suggest that grounding should be an architectural decision: compute, compress, and version telemetry first; expose a controlled action space second. I-B Key idea: registry-driven semantic compilation We introduce a design in which a declarative metrics registry is the single source of truth. The registry captures metric semantics, retrieval logic, and governance metadata; a deterministic compilation step generates the concrete MCP-exposed tool surface from this registry. In effect, the registry implements an “interface-as-code” pattern for agentic systems: rather than hand-writing tools, operators declare metrics and compilation produces executable, audited, and access-controlled tools. This pattern directly addresses tool drift: tool specifications and their implementations are derived from the same artifact, ensuring alignment over time. It also constrains the agent’s action space, which materially reduces the hallucination surface and simplifies policy enforcement. I-C Scope and contributions REGAL is presented as a systematization and reference architecture rather than an algorithmic contribution. The core contributions are: • Registry-driven MCP compilation: An interface-as-code pattern that synthesizes tool schemas, retrieval logic, and governance controls from declarative metric definitions. • Deterministic grounding architecture: A practical reference architecture that separates deterministic telemetry computation (Medallion ELT) from probabilistic reasoning (LLM agents). • Prototype validation and lessons learned: A prototype implementation and case study demonstrating feasibility and describing operational trade-offs (latency, token scaling, governance). Novelty statement To the best of our knowledge, REGAL is the first systematization that (a) formalizes registry-driven semantic compilation as an explicit architectural grounding constraint for enterprise LLM agents, and (b) operationalizes that constraint end-to-end—covering deterministic ingestion, versioned Medallion transformations, and automated compilation of declarative metric definitions into MCP-exposed, access-controlled tools. While elements of the pipeline (e.g., Medallion ELT, change streams, MCP tooling) exist independently in prior work, REGAL’s novel combination is the explicit architectural commitment to treat the registry as the single source of truth and the compilation step as the enforcement mechanism that bounds the agent’s action space. Figure 1 provides a visual summary of the architecture: deterministic refinement produces versioned Gold artifacts that are the sole inputs to LLM-based agents. Fig. 1: High-level architecture: sources are deterministically refined into versioned Gold artifacts; agentic LLMs consume only those artifacts. The thick boundary emphasizes the deterministic \rightarrow probabilistic handoff. I-D Paper roadmap The remainder of the paper proceeds as follows. Section II presents the reference architecture and the deterministic/probabilistic separation. Section III details the registry-driven semantic compilation pattern (the paper’s primary contribution). Section IV describes the deterministic ingestion and Medallion pipeline that produces Gold artifacts. Section V discusses storage and serving considerations; Section VI describes the push/pull interaction model. Section VII positions this approach relative to RAG, Text-to-SQL and vendor platforms. Section VIII presents the prototype implementation and case study. We conclude with open research directions in Section IX. II The Reference Architecture Designing reliable agentic AI for engineering observability requires more than attaching an LLM to existing APIs. Enterprise telemetry is heterogeneous, temporal, and governed by access controls; it is also subject to rate limits, schema drift, and transient failures. Exposing raw telemetry to probabilistic reasoners therefore risks costly token usage, ambiguous semantics, and brittle tooling. The REGAL architecture centers a simple design principle: perform deterministic computation and semantic compilation before exposing any telemetry to probabilistic agents. In practice, this means (a) produce replayable, version-controlled Gold artifacts via a deterministic ELT pipeline, and (b) synthesize a bounded, audited action space (tools) from a declarative metrics registry. The registry-to-tool compilation pattern is the primary architectural contribution of this work: it is the mechanism by which the system constrains reasoning to a safe, auditable interface. II-A Architectural overview The system separates responsibilities into four layers with a unidirectional data flow: 1. Source layer. External systems (version control, CI/CD, issue trackers, observability) that emit events or state. 2. Ingestion & orchestration (write path). Deterministic extraction, validation, reconciliation, and upsert semantics that guarantee replayability. 3. Medallion storage (context store). A layered transformation that converts raw telemetry (Bronze) to harmonized records (Silver) and to compact, semantically rich Gold artifacts intended for AI consumption. 4. Semantic layer (read path). A registry-driven compilation component that generates concrete MCP-exposed tools, access controls, and caching policies; agents reason only over these synthesized tools and their deterministic outputs. II-B Deterministic–probabilistic non-interference (concise property) To make