Alomana Lab

We design reliable autonomous systems.

AI agents that run continuously, in production, against real data, and don't break in ways you can't reason about.

---

What we're working on

Autonomous agents fail in predictable ways: context bloat, structured output drift, brittle tool orchestration, silent degradation under load, and gaps between what models internally represent and what we can reliably control. We study these failure modes systematically and build the infrastructure to prevent them.

Our work spans:

• Evaluation environments — reproducible test harnesses for autonomous systems that capture production failure modes, not just benchmark accuracy
• Structured output reliability — why models with high reasoning accuracy still produce near-zero usable output, and how to fix it without fine-tuning
• Context management for long-horizon agents — token-aware compaction, pre-trimmed tool outputs, sidecar extraction patterns
• Agentic orchestration primitives — lightweight patterns for multi-step, multi-model workflows that are inspectable and recoverable
• Mechanistic interpretability for control — measuring when model-internal representations support intervention, and when detection and steering come apart

---

Research

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models arXiv 2606.24952 · 2026

Core finding: knowing where a behavior is represented in a model does not necessarily mean we can control it there. We measure the geometric gap between directions that detect a behavior and directions that cause an intervention. Output format collapses onto a shared axis, but hallucination does not: models can linearly detect fake entities with perfect separability while the corresponding refusal direction remains nearly orthogonal. The result is a reproducible detection-intervention gap across model families and scales, showing that model knowledge and model steering can dissociate even when the representation is easy to read.

When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models arXiv 2605.02363 · 2026

Core finding: small language models can achieve near-perfect reasoning accuracy while producing structurally invalid output on the vast majority of responses. Accuracy metrics don't capture what actually breaks in production. We introduce AloLab — an evaluation framework that measures output usability, not just correctness — and demonstrate strong output reliability without fine-tuning.

---

Design philosophy

Reliable before capable. A system that fails unpredictably is worse than a system with narrower scope. We start from failure modes and work backwards.

Environments matter as much as models. Most evaluation is contaminated by the gap between benchmark conditions and deployment conditions. We build evaluation environments that reproduce what actually happens when agents run in the wild — context pressure, malformed inputs, cascading tool errors.

Models deserve rigorous engineering. The frontier isn't the constraint for most real workloads. We care about what you can actually run, control, and audit.

Transparency about failure. Our most useful findings have come from systems that didn't work. We publish those too.

---

Background

Alomana Lab is the research and open source arm of Alomana, an enterprise AI company building autonomous agent systems for enterprises. Our tooling is extracted from production deployments, hardened, and released.

---

Get involved

If you're working on agent reliability, evaluation environments, structured generation in constrained models, or interpretability for controllable systems — we're likely solving adjacent problems.

Open an issue. Cite the papers. Reach out at lounge@alomana.com.