\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={SpatialClaw: Oral Knowledge Questions}}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate,1]{leftmargin=*,itemsep=1.1em,topsep=0.7em}
\setlist[enumerate,2]{label=(\alph*),leftmargin=*,itemsep=0.7em,topsep=0.7em}
\begin{document}
\raggedright
{\LARGE\bfseries SpatialClaw\par}
\vspace{0.3em}
{\Large\bfseries Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words.
\begin{enumerate}
\item \textbf{What is an action interface?}

Explain what an action interface determines about how a VLM invokes
tools and receives their results.

\medskip
How does SpatialClaw combine the flexibility of writing Python with
the ability to inspect results and revise its approach?

\newpage
\item \textbf{How do segmentation and reconstruction work together?}

Suppose the question is ``How far is the chair from the table?''
\begin{enumerate}
\item What information does a segmentation mask provide, and what
information does a 3D reconstruction provide?
\item How can the mask select the object's points from a dense 3D point map?
\item Why must the mask and point map correspond to the same frame,
with measurements expressed in compatible coordinates?
\end{enumerate}

\newpage
\item \textbf{Why must the agent inspect intermediate results?}

A tool returns a mask, and the next Python cell successfully computes
a distance from it.
\begin{enumerate}
\item Why does this not establish that the answer is correct?
\item What can displaying the mask with \textit{show()} reveal that
printing only the distance cannot?
\item If the mask is wrong, how can the agent revise its analysis while
reusing valid results already stored in the kernel?
\end{enumerate}

\newpage
\item \textbf{What does ``training-free'' mean for SpatialClaw?}

SpatialClaw improves a VLM's spatial reasoning without updating its
model weights.
\begin{enumerate}
\item Where does the additional capability come from?
\item Why does ``training-free'' not imply low inference cost?
\item How does SpatialClaw's use of explicit code and intermediate arrays
differ from learning continuous visual thought tokens, as in CoVT?
\end{enumerate}

\newpage
\item \textbf{Why can a tool-using agent still give a wrong answer?}

Suppose SpatialClaw returns an incorrect distance even though every
tool call and Python cell completed successfully.
\begin{enumerate}
\item Give one possible \textbf{perception error} and one possible
\textbf{reasoning or measurement error} that could explain the result.
\item What checks could help distinguish these two causes?
\item Why does access to accurate numerical computation not guarantee
accurate spatial reasoning?
\end{enumerate}
\end{enumerate}
\end{document}
