Research company · Learning products

AI research for reliable systems.

We study how AI behaves, how to evaluate it, and what it costs to use.

Our research Meet SayPace

We share research preprints and are developing SayPace, our learning brand.

Company structureAI research and English-learning products, under one company.

01 / Research directions

Three questions guide our work.

Start with the question. Explore the methods and supporting papers.

Reliable evaluation

What does a success score measure?

Match the score to the task.

Explore this area

Model behavior & mechanisms

What changes a model’s behavior?

Use controlled comparisons to test an explanation.

Explore this area

Efficient AI systems

What does the deployed system cost?

Measure quality and resource use together.

Explore this area

02 / A research example

Method schematicEach box is a complete run; its dots are attempts. Keep runs separate when evaluating procedure success. No measured results are shown.

Studied search and self-correction designs

In the studied search and self-correction designs, pooling dependent attempts overstated the procedure’s run-level success rate.

Match the evaluation unit to the way the system operates.

View the paper

03 / Research library

Latest manuscripts

View the collection
Preprint · v5

Serving cost

Does matching accuracy also match serving cost?

Accuracy-Matched Is Not Cost-Matched: Pricing INT4 Reasoning Deployments by Served GPU Time per Solved Problem

Read manuscript Preprint · v2

Pooled statistics

What can a pooled success total estimate?

What a Pooled Success Total Supports for pass@k: Identification, Unbiased Estimation, and a Jackknife Below the Run Threshold

Read manuscript Preprint · v3

Evidence-controlled evaluation

Is the evidence fixed in a coherence test?

Information Effects Are Not Incoherence: Fixed-Evidence Coherence Tests for Tabular Foundation Models

Read manuscript

These manuscripts are preprints and have not been peer reviewed.

04 / Learning products

SayPace

Our learning brand for IELTS preparation and focused practice in everyday English skills.

Speaking is in planning and Numbers is in development. Neither app is publicly available yet.

Meet SayPace