KRContact
Technical reports

Measure the system, publish the method, state the limits

Patty reports do more than display results. They explain why the question matters, the conditions behind the measurement, how a decision changes, and what the evidence cannot establish.

Research agenda

Four research tracks for turning enterprise AI into an operating system

We study the complete deployable path rather than an isolated model score.

01

Korean models & evaluation

Evaluation designed for Korean instructions, long coding sessions, local documents, and repository conventions.

02

Inference & GPU systems

Measurements across distributed serving, scheduling, KV cache, and resource boundaries.

03

Agent execution

Finished-path analysis connecting observation, planning, approval, tool execution, and recovery.

04

Governance & evidence

Validation that identity, delegation, data boundaries, and execution receipts operate as real controls.

How to read

A report is a decision instrument, not a list of numbers

Inspect the measurement unit and exclusion boundary before the size of a result.

  1. 01

    Question

    The design or operating decision the report addresses

  2. 02

    Conditions

    Hardware, build, data, participants, and observation period

  3. 03

    Result

    What was observed and how it changes a decision

  4. 04

    Limits

    What cannot be generalized or has not yet been verified

Published research

Published reports

  1. R-01

    Finished-path measurements for an engine-native agent browser

    A measurement of the complete observe→classify→park→approve→dispatch path when an AI agent is embedded in the browser engine. Semantic snapshots remained around 14–16ms in the tested fixtures, and the governed engine cycle completed in under 25ms.

    Read report
  2. R-02

    Designing a benchmark corpus for coding agents

    Design notes for an end-to-end corpus stratified across eight workload classes, a 50-instance SWE-bench Verified subset, and a context-retention A/B harness, including anti-guess tasks and scorer validation.

    Read report
Editorial program

A cumulative technical record, not a one-time announcement

Reports evolve with releases, remeasurement, and environmental change. Modification dates and measurement boundaries remain explicit; earlier results are not silently generalized.

Measurement reports
Execution paths and system performance
Benchmark methodology
Task, scorer, and corpus design
Engineering deep dives
Architecture choices and failure modes
Technical essays
New operating problems and decision frames
NEXT STEP

Define the questions that matter in your environment

Instead of copying a public number, design an evaluation around your hardware, data, network, and approval requirements with Patty.

Discuss technical validation