Skip to content
brianletort.aiD. Brian Letort, Ph.D.

RESEARCH

The systems layer where enterprise AI becomes reliable.

Preprints and companion implementations focused on the architecture, governance, and operating mechanics that make AI systems inspectable, portable, and dependable in enterprise use.

Research sequence

Context Compilation advances as a systems stack.

Published work covers the first four layers. Evaluation and enterprise operation remain explicitly open research.

  1. 01
  2. 02

    Representation

    Context IR and Compiler Passes for Enterprise AI

    Published preprint

  3. 03
  4. 04
  5. 05

    Evaluation

    Benchmarking compilation quality and runtime behavior

    Open research

  6. 06

    Enterprise operation

    Operating, governing, and auditing compiled context

    Open research

Programs

Two programs, five published preprints.

Each program is framed by the question it is trying to answer, the outputs currently available, and the companion implementation where one exists.

Four published Zenodo preprints

Context Compilation

Central question
How should external context be selected, transformed, governed, optimized, and lowered into reliable execution?
Thesis
Enterprise AI needs a full context stack, not just retrieval and longer windows. This research program starts with a foundational paper that introduces Context Compilation Theory, then continues with a three-paper trilogy that splits the stack into representation and semantics, runtime and lifecycle, and optimization and efficiency. The MemoryOS project record connects the research to the reference implementation.
Outputs
1 precursor theory paper · 3 trilogy papers
Latest publication
April 12, 2026
View programMemoryOS ↗

One published Zenodo preprint

Cybernetic Software Delivery

Central question
How should software delivery be governed when autonomous agents become active producers of engineering artifacts?
Thesis
This paper argues that the traditional SDLC cannot govern a world where AI agents produce engineering artifacts. Cybernetic Software Delivery (CSD) reframes delivery as a control system — a closed-loop model with humans, agents, tools, evaluations, and feedback. The primary unit of delivery becomes the governed run: a bounded, observable, policy-aware execution of delegated work that is fully reconstructable from storage. The paper introduces a nine-stage lifecycle, a five-category artifact taxonomy, 13 new metrics extending DORA, and a four-tier governance framework.
Outputs
1 preprint · 1 companion repository
Latest publication
April 9, 2026
View programPraxo-AgentOS ↗

Featured contributions

The contribution is the stack, not a collection of isolated papers.

The precursor establishes the frame; the trilogy develops representation, runtime, and optimization; Cybernetic Software Delivery applies a control-system lens to agentic engineering.

Precursor · Theory

Toward a Theory of Context Compilation for Human-AI Systems

Every enterprise building with LLMs faces the same fragmentation problem: RAG pipelines, memory systems, prompt templates, and agent context are all solving pieces of the same puzzle without a unifying theory. Context Compilation Theory provides that missing layer — a principled way to think about how context flows from sources through transformation to execution. For enterprise AI leaders, this reframes context management from an engineering detail into an architectural concern that determines system quality, governance, and portability.

Read contribution

Trilogy · Paper 1 operationalizes Context Compilation Theory

Context IR and Compiler Passes for Enterprise AI

Most enterprise AI stacks still glue retrieval, summarization, prompts, and memory together without a stable internal representation. This paper gives teams the missing semantic layer: a way to reason about what context is, how it changes shape, where guarantees break, and how governance travels with the compiled artifact.

Read contribution

Independent program · Delivery governance

Cybernetic Software Delivery: A Governed Lifecycle for Agentic Engineering Work

As AI agents increasingly write code, generate tests, and produce documentation, engineering organizations need a new operating model. The traditional SDLC was designed for human developers working in sequential phases. CSD provides the governance, observability, and quality framework needed when agents are first-class producers. For engineering leaders, this is the missing operating manual for agentic software delivery — not just how to use AI tools, but how to govern the work they produce.

Read contribution

Browse the research

Filters
Program
Program
Has manuscript
Has manuscript

5 papers

Zenodo preprint · April 12, 2026

Paged Context Memory: Runtime Systems for Evidence Blocks, Locality, Speculation, Linking, and Policy Preservation

Paper 2 extends the trilogy from semantics into runtime behavior. It argues that compiled context should live like managed memory: evidence blocks enter working sets, locality governs movement, speculative execution prepares likely futures, context GC removes dead state, and provenance remains attached through lifecycle transitions.

Context Compilation · Published preprint

Zenodo preprint · April 12, 2026

Quantized Context: Utility-Preserving Compression and Mixed-Precision Context Assembly

Paper 3 turns optimization into a first-class concern. It defines a semantic precision ladder, a distortion model for compiled context, mixed-precision assembly strategies, and recovery-aware compression so systems can stay cheap until risk, policy, or task criticality demands higher fidelity.

Context Compilation · Published preprint

Zenodo preprint · April 9, 2026

Toward a Theory of Context Compilation for Human-AI Systems

This paper proposes that context in AI systems should not merely be retrieved or remembered — it should be compiled. Context Compilation Theory introduces a formal framework where context is selected, transformed, governed, optimized, and lowered into executable context packs for downstream models, agents, and interfaces. The paper defines Context IR as a portable intermediate representation, analogous to compiler IRs in traditional software, and presents CompileBench as a benchmark specification for measuring compilation quality rather than recall alone.

Context Compilation · Published preprint

Zenodo preprint · April 9, 2026

Cybernetic Software Delivery: A Governed Lifecycle for Agentic Engineering Work

This paper argues that the traditional SDLC cannot govern a world where AI agents produce engineering artifacts. Cybernetic Software Delivery (CSD) reframes delivery as a control system — a closed-loop model with humans, agents, tools, evaluations, and feedback. The primary unit of delivery becomes the governed run: a bounded, observable, policy-aware execution of delegated work that is fully reconstructable from storage. The paper introduces a nine-stage lifecycle, a five-category artifact taxonomy, 13 new metrics extending DORA, and a four-tier governance framework.

Cybernetic Software Delivery · Published preprint

Methods & citation

The publication record is explicit.

All five entries are published Zenodo preprints. A DOI is a persistent identifier, not evidence of peer review.

Zenodo hosts the deposited preprint PDFs and citation metadata. Manuscript versioning and companion code are maintained in the linked GitHub repositories, external to Zenodo.

Each record exposes the artifacts actually available: DOI/Zenodo, PDF, companion code, and, for the trilogy papers, a readable manuscript.

Research claims, limitations, and proposed next directions are stated in the individual preprints. No peer-review status is inferred here.

How to cite: open a paper’s DOI / cite link in the research browser.
Research | Brian Letort