I build research systems that turn ambitious AI claims into reproducible, inspectable evidence.
Nullius in verba.Take nobody's word for it.
01
Featured research
Claim-relative inference
arXiv:2608.192692026 · v3
What Does an Evaluation License? A Commit-Bound Census of Claim-Relative Inference in Inspect Evals
Evaluation artifacts produce scores. They do not automatically bind the historical evidence or alternative semantics needed to replay the claim attached to that score.
DFrozen substrate
Versioned inputs, observations, traces, state, and support.
FGrounded family
Primary admissible readings and a separate review envelope.
qClaim query
Exact score, winner, complete order, or pairwise relation.
124mechanically eligible units
14deterministic analyses completed
110typed stops before inference
What the paper adds
Not another benchmark score—a contract for what may be inferred.
01
Every unit receives a terminal record. A STOP is not a zero score; it identifies the first missing object and the minimum artifact needed to unblock replay.
02
Primary and review families remain separate. A winner can stay identified even when the complete order does not.
03
A fresh run cannot reconstruct a historical evaluation event. Replay requires bound observations, mediator decisions, state, support joins, and semantic admission.
“The result is a reviewable claim-replay contract rather than a forced evaluator meaning or a single flip label.”Conclusion · arXiv v3
Ad fontesBack to the sources
02
Open-source contributions
Changes that landed
Selected contributions merged upstream. Updated .
01Query enginesMerged
apache / datafusion #25150
Skip hashing for fully ordered group keys
Added a columnar path for supported, fully ordered multi-column aggregations. Adjacent equal keys share a group without a hash table, reducing grouping work and retained state.
Implemented ROC AUC, Rand index, V-measure, and Kendall tau-b in NumPy, removing scikit-learn from the core dependencies while checking results against reference implementations.
Removed credential-taking actions from the natural-language planner until a separate credential channel exists. Schema, parser, and prompt checks enforce the same boundary.
1,883 Linux CI tests passed · schema and parser regressions+305 −32 · 11 files
Added planar rotation for batches of 2D vectors, restoring the VectorField curl example while preserving the existing rotation behavior for 3D vectors.
Four 2D/3D regressions · real VectorField reproduction+53 −2 · 2 files
Wearable application development on openvela, covering embedded systems, interaction design, and integration. Organized by the China Computer Federation (CCF); the 2026 competition received 586 valid entries and awarded 34 first prizes across all tracks.
Develop smart-home AI agents that connect natural-language requests with device tools and control workflows. This CCF competition track tests integration across language models, embedded software, and open-source systems.
围绕 openvela 构建智能家居 AI Agent,将自然语言请求接入设备工具与控制流程。作为 CCF 开源竞赛赛道,重点考察大模型、嵌入式软件与开源系统的综合实现能力。
A CCF-organized quantum programming challenge using Origin Quantum’s QPanda3 and VQNet tools. Tasks span circuit design and algorithm implementation, testing the ability to turn quantum theory into working code.
National Undergraduate Electrical Mathematical Modeling Competition
全国大学生电工数学建模竞赛
CSEE Cup · “Diangong Cup”
中国电机工程学会杯 · 电工杯
Second Prize二等奖
A national competition organized by the Chinese Society for Electrical Engineering. Teams of three have 72 hours to model an applied problem and submit a paper, testing mathematical reasoning, computation, and technical writing.