TRACE AI Labs

Trustworthy Routing, Agents, Compliance, and Efficiency

TRACE AI Labs is the joint research of Mika Okamoto and Ansel Erol on making AI agents trustworthy, auditable, and efficient. We care about leaving a clear trace — the audit trails that make agent behavior explainable and verifiable.

What we work on

Agent compliance & enterprise AI — whether agents actually follow the rules they're given, and why they don't.
LLM & agent routing — explainable, cost-aware routing that picks the right model for each task and shows its reasoning.
Efficient AI — frontier-level results under real cost and latency budgets.

…and adjacent directions in agentic evaluation and explainability as they come up.

Papers

Agent compliance and enterprise AI

PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?

Mika Okamoto, Ansel Kaplan Erol

Preprint, 2026 · under review

PACT (Pressure-Applied Compliance Testing) is a novel benchmark testing whether enterprise AI assistants keep following company rules when violating them is the convenient choice: a deadline looms, a manager says to make an exception, or the user pushes back. Its 3,364 items each pair a rule with a shortcut that violates it, set in realistic multi-turn workplace conversations across 48 scenarios, 12 regulated domains, and nine psychology-grounded pressures. Across 22 models, ordinary pressure raises violation rates by 65% and no model clears the bar for unsupervised use.

Paper Preprint DOI Interactive results Code Dataset

BibTeX

@misc{okamoto2026pact,
  title  = {PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?},
  author = {Okamoto, Mika and Erol, Ansel Kaplan},
  year   = {2026},
  note   = {Under review},
}

Agent Memory Is a Surface for Endogenous Authorization Laundering

Tommaso Cerruti, Mika Okamoto, Ansel Kaplan Erol

Preprint, 2026 · EleutherAI Summer of Open AI Research (SOAR), mentored by Trace AI Labs

We study a failure mode where an agent's own persistent memory accidentally creates authority that the underlying history never granted, without any external attacker. Across procurement, cybersecurity, and finance, incremental memory updates create false authority for up to 50.2% of unauthorized requests; once present, executors act on it in 98.6% of trials. We also introduce EAL-Bench, an open-source benchmark for studying these failures and potential mitigations.

EAL-Bench system overview: a memory-writer agent compresses an organizational history of grants and amendments into persistent authorization memory, an executor acts on that memory plus a later request, and a hidden canonical ledger with a deterministic oracle scores each tool action as compliant or non-compliant.

Paper Code

BibTeX

@misc{cerruti2026agentmemory,
  title  = {Agent Memory Is a Surface for Endogenous
            Authorization Laundering},
  author = {Cerruti, Tommaso and Okamoto, Mika and Erol, Ansel Kaplan},
  year   = {2026},
  eprint = {2609.01836},
  archivePrefix = {arXiv},
}

Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance

Mika Okamoto, Ansel Kaplan Erol, Kutluhan Erol

AIES 2026 · also at the COLM 2026 Workshop on Agent Behavior

We diagnose why AI agents violate legal constraints in enterprise settings by mapping systematic compliance failure modes across 12 instruction-tuned procurement chatbots. Safety-fine-tuned models remain broadly compliant, while task-optimized and agentic models trade off compliance under weak regulatory signals, non-command phrasing, and organizational pressures.

Experimental setup: a corporate LLM chatbot receives a system context with corporate policy, rule framing, financial pressures, and social norms, then chooses between an ISO-compliant and a non-compliant vendor.

Paper Interactive results Code & data

BibTeX

@inproceedings{okamoto2026legalincentives,
  title     = {Why Do AI Agents Break Rules? How Framing, Context,
               and Social Signals Shape Compliance},
  author    = {Okamoto, Mika and Erol, Ansel Kaplan and Erol, Kutluhan},
  booktitle = {AAAI/ACM Conference on AI, Ethics, and Society (AIES)},
  year      = {2026},
  eprint    = {2608.12323},
  archivePrefix = {arXiv},
}

Routing and explainability

Explainable Model Routing for Agentic Workflows

Mika Okamoto, Ansel Erol, Mark Riedl

Workshop on Human-Centered Explainable AI (HCXAI) at CHI 2026

Agentic AI systems increasingly route subtasks across multiple specialized models, but these routing decisions are opaque to end users. We propose a framework for explaining model routing decisions in natural language, supporting meaningful human oversight of multi-step pipelines.

Paper

BibTeX

@inproceedings{okamoto2026explainable,
  title     = {Explainable Model Routing for Agentic Workflows},
  author    = {Okamoto, Mika and Erol, Ansel and Riedl, Mark},
  booktitle = {Workshop on Human-Centered Explainable AI (HCXAI) at CHI},
  year      = {2026},
  eprint    = {2604.03527},
  archivePrefix = {arXiv},
}

Trust by Design: Skill Profiles for Transparent, Cost-Aware LLM Routing

Mika Okamoto, Ansel Erol, Glenn Matlin

MLSys Young Professionals Symposium (YPS) 2025

We introduce BELLA, a routing system that constructs interpretable skill profiles for LLMs to match incoming queries to the most capable and cost-efficient model. Routing decisions stay transparent and grounded in empirically measured model competencies.

Two-stage pipeline: profiling runs tasks through a benchmarker and a critic-LLM skill profiler to produce cost/performance metrics and skill matrices, which train a model selector that returns a best model plus a selection rationale for each new application.

Paper

BibTeX

@inproceedings{okamoto2025bella,
  title     = {Trust by Design: Skill Profiles for Transparent,
               Cost-Aware LLM Routing},
  author    = {Okamoto, Mika and Erol, Ansel and Matlin, Glenn},
  booktitle = {MLSys Young Professionals Symposium (YPS)},
  year      = {2025},
  eprint    = {2602.02386},
  archivePrefix = {arXiv},
}

People

Ansel Erol Mika Okamoto