./research

Questions we're working on.

Running notes from the lab. These are open problems, not finished papers; we publish the questions first and the results when they survive contact with production.

• aug 2026

  Detecting Sarcasm in AI: Building Emotional Subtext Understanding

  Sarcasm detection is an emotional intelligence problem, not a text classification problem. It requires understanding intent, register, cultural context, and often the history of the conversation. Here is how to build an AI agent that handles sarcasm and emotional subtext reliably.

  sarcasm detectionemotional subtextagentsactive
• aug 2026

  Human Emulation in AI Systems: Beyond Language Models to Behavioral Reasoning

  True human emulation in AI is not fine-tuning a language model on more human text. It requires behavioral reasoning, contextual consistency, and the kind of state that carries meaning across a conversation. Here is where the gap is, and what it takes to close it.

  human emulationbehavioral reasoningagentsactive
• aug 2026

  Tools for Building Emotionally Intelligent AI Agents in 2026

  The tooling landscape for emotionally intelligent AI agents falls into three categories: emotion recognition APIs, reasoning and orchestration frameworks, and evaluation platforms. Here is a practical map of the leading options, what each does well, and where each falls short.

  emotional intelligencetoolsagentsactive
• aug 2026

  Building Emotional Reasoning into AI Agents: A Step-by-Step Framework

  Emotional reasoning in AI agents is the capacity to model a user's emotional state, maintain it across turns, and select responses that are appropriate to that state - not just to the informational content of the request. Here is a practical framework for integrating it into production systems.

  emotional reasoningagentshuman layeractive
• aug 2026

  What a code index actually saves an agent

  A code index is supposed to stop an agent re-searching files it has already read, and the published figure for one of the best of them is 10x fewer tokens. We measured it on 24 symbols across two codebases. At default settings it saves 2.8x to 4.0x. Asked for the complete answer it costs more than grep on 22 of the 24. The saving is real, and it is selection rather than compression.

  efficiencyagentsmeasurementactive
• jul 2026

  A verifiable audit layer for AI

  A tamper-evident audit layer for model inference and agent execution: canonical serialization, domain-separated hashing, DSSE signatures, a per-run Merkle DAG, and a cross-run Merkle Mountain Range that proves no run was deleted. Every primitive was chosen from a head-to-head benchmark against the real alternatives, with the raw data committed. This note is the design and the experiments behind it.

  auditcryptographybenchmarksshipped
• jul 2026

  Reducing the token cost of long-running agents

  Agents burn most of their tokens re-reading context they've already seen: the same instructions, the same company facts, the same history, on every call. We're studying how far caching, context compression and precompiled knowledge can cut the marginal cost of an agent step without degrading its decisions.

  efficiencyagentsactive