I build tools for understanding what AI agents did, why they did it, and whether their behavior can be trusted.

My work combines statistical inference, mathematical modeling, provenance, and software engineering to make autonomous systems more observable and auditable.

Rather than treating AI evaluation as a collection of pass/fail tests, I'm interested in systems where uncertainty is quantified, behavior is measurable, evidence is preserved, and claims about reliability can be independently examined.

My work focuses on three related problems:

AI Validation Developing empirical methods for evaluating autonomous-agent behavior and detecting unexpected, unstable, or unsupported behavior.

Statistical & Geometric Analysis Investigating how optimization, parameterization, Fisher information, sampling, and other mathematical structures affect conclusions drawn about learned systems.

Agent Provenance & Observability Building infrastructure that records agent actions and state transitions so behavior can be reconstructed, analyzed, and audited.

My projects approach reliable AI from two complementary directions: the mathematical foundations of valid inference and the infrastructure required to preserve evidence about autonomous systems.

A framework for reasoning about whether transformations preserve the evidence required for downstream inference.

The project explores the relationship between representation, information preservation, and the validity of conclusions drawn from transformed data.

Its central concern is not simply whether two representations appear similar, but whether the transformation preserves the information necessary to support the inference being made.

Research themes:

• Inferential validity
• Evidence preservation
• Statistical information
• Representation and transformation
• Mathematical AI validation

Explore Inferential Fidelity Framework →

An evidence and observability layer for autonomous AI agents.

AgentLedger records agent operations as structured, auditable events so developers can reconstruct what an agent did and analyze its behavior programmatically.

The project is evolving toward a safety-oriented infrastructure component for agent systems, with features including:

• APE provenance events and operational-boundary validation
• Durable run checkpoints with leases and fencing tokens
• External-effect lifecycle with explicit ambiguous (UNKNOWN) outcomes and reconciliation
• Request-bound idempotency for consequential operations
• Tamper-evident hash chains with optional Ed25519 evidence signatures
• OpenTelemetry and MCP integration surfaces
• Bayesian-ready evidence export for downstream inference
• Transactional controls, spending limits, automated tests, and CI

with ledger.transaction():
    ledger.debit("research_agent", 10)
    ledger.credit("analysis_agent", 10)

Agent retries can be made idempotent:

ledger.debit(
    "research_agent",
    10,
    idempotency_key="task_2026_001"
)

The objective is straightforward:

Give autonomous systems an evidence trail that can be inspected, tested, and independently verified.

Explore AgentLedger →

I also maintain work on lightweight decoder-only pretraining architecture adapters and training controls for TransformerLens, including provenance hooks that can emit reproducibility evidence without coupling TransformerLens to a specific audit backend.

Explore TransformerLens-Training-Controls →

Using posterior updating and uncertainty quantification to evaluate evidence about model and agent behavior.

Studying whether observed properties of learned systems are intrinsic to the learned function or artifacts of parameterization, optimization, sampling, or measurement.

This includes investigating Fisher information spectra, optimizer-dependent behavior, identifiability, and the distinction between properties of a learned function and properties of its coordinate representation.

Exploring mathematical representations of allowable agent states and transitions as a foundation for enforceable behavioral constraints.

Creating verifiable chains of evidence connecting agent actions, state transitions, policies, observations, and resulting claims.

My work follows a common progression:

Mathematical foundations ↓ What information is required to justify an inference?

Measurement & empirical validation ↓ Can that information be reliably observed and quantified?

Evidence infrastructure ↓ Can the relevant actions, states, and observations be preserved?

Autonomous systems ↓ Can agent behavior be reconstructed, tested, and audited?

The goal is to connect mathematical questions about what can legitimately be inferred with engineering systems that preserve the evidence needed to answer those questions.

Reliable autonomous systems require more than successful outputs.

They require infrastructure capable of answering:

1. What happened?
2. What evidence produced the decision?
3. What changed between runs?
4. Was the action inside its permitted operating boundary?
5. What uncertainty accompanies the result?
6. Can the result be reproduced or independently audited?

A system should not merely produce an answer. It should preserve enough evidence for that answer—and the process that produced it—to be examined.

I'm interested in collaborations around agent infrastructure, AI validation, probabilistic modeling, interpretability, statistical inference, and reliable autonomous systems.

GitHub · X · LinkedIn