A guide to the site

Find a thread.
Follow it somewhere.

Start with a result, a project, or a question. The connections matter, but each page should also make sense on its own.

Start here

The overview, my background, and the notebook index.

HomeResearch, projects, and a little about me
中文首页研究、项目与个人介绍
CVPublications and experience · printable
Field notesWriting organized by question

Research

The experiments, my contribution, and the limits of each result.

The Recall DebtAccepted poster
Before You ScaleAccepted
Diagnosing Agent CapabilitiesAccepted poster
Factor(U,T)Accepted
A Benchmark Audit of Site ConfoundsAccepted full paper · poster
Detecting Hidden Chain-of-Thought in Large Language ModelsAccepted
Factor(U,T) · 中文解读只看计划，够吗？

Projects

What I built and how the work connects to the research.

Eval Evidenceevaluation infrastructure
Circuit Compassbiomedical
Brewery Control Systemsindustrial systems
POP2 Protein Expressionwet-lab system
Try Eval EvidenceA reproducible synthetic walkthrough

Notebook

Personal essays, working ideas, and questions I am still testing.

Optimization without an optimizerText revised 10 Sep 2026
Inside the loop.Text revised 10 Sep 2026
A zero needs adjudication.Text revised 10 Sep 2026
Where does the model end?Text revised 10 Sep 2026
The frontier is funded.Text revised 10 Sep 2026
Systems are strange loops.Text revised 10 Sep 2026
Red teaming is accountability.Text revised 10 Sep 2026
The loop is an evidence system.Text revised 10 Sep 2026
When synthesis gets cheap, truth gets expensive.Text revised 10 Sep 2026
Control has grammar.Text revised 10 Sep 2026
Education is an evaluation problem.Text revised 11 Sep 2026
The tiers are not one model.Text revised 11 Sep 2026
Veracity is infrastructure.Text revised 10 Sep 2026
The shape of what's absent.Text revised 10 Sep 2026
Handoffs are memory.Text revised 10 Sep 2026
Epistemic grounding is the missing layer.Text revised 10 Sep 2026
The ontology is the interface.Text revised 10 Sep 2026
Orientation precedes grounding.Text revised 10 Sep 2026
Do not collapse the spirals.Text revised 10 Sep 2026
One untrusted agent is enough.Text revised 10 Sep 2026

Dates describe different things: a notebook’s revision date tracks its text; a paper’s acceptance or publication status describes the research record; “Last checked” records the evidence review. A visual refresh does not reset those dates.