-The cost of a flow is the sum of its reviewer overload costs,
-assignment costs, and reviewer boring / very boring load costs,
-minus paper bonuses. Any one of those components can be traded off against
-the others. We attempted to assign reasonable weights to each component,
-but it is difficult to know without testing the algorithm on real data.
-In any event, all the parameters are easy to tune to realize the priorities
-of a particular application.
-
-\begin{figure*}
-\begin{center}
-\centerline{\includexfig{flow}}
-\caption{Flow Construction.}
-\label{flow-fig}
-\end{center}
-\end{figure*}
-
-
-\section{Experimental Results}
-Waiting for data from NSF.
-
-Synthetic Data.
-
-\section{Conclusions}
-
-The source code for the current version of the proposal matcher may be
-browsed or downloaded at:
-\[\hbox{\url{TODO}}\]
+In the example in Figure~\ref{flow-fig},
+paper 1 is interesting to reviewer 1 and boring to reviewers 2 and 3.
+Reviewer 2 is expert on paper 1, with reviewers 1 and 3 merely knowledgeable.
+(Reviewer edges for paper 2 are not shown.)
+This illustrates how, in principle,
+the desirability and expertise relations might differ.
+Each is taken into account at a different stage of the construction.
+
+The cost of a flow (assignment) is the sum of its reviewer overload costs,
+per-review costs, and reviewer boring / very boring load costs,
+minus paper bonuses. Any one of these components can be traded off against
+the others.
+
+\section{Evaluation}
+
+We have implemented the matching algorithm, based on the construction above, in
+Haskell.
+The construction is intended to illustrate
+ways to model real concerns that we find reasonable a priori.
+We do not have enough experience with real-world
+instances to be confident that each part of the construction serves its intended
+purpose or that the parameter values we have chosen are suitable.
+Fortunately, the parameter values are easy to change in the source code of our
+implementation, and even substantive changes to the graph structure are not too
+hard to make. At a minimum, we believe that min-cost max-flow provides a
+reasonable framework for global optimization of an assignment.
+
+We have worked with Michael Hicks, program chair of POPL 2012, to apply our
+method to the paper assignment problem for that conference.
+The set of POPL 2012 reviewers consisted of the
+program committee (PC) and an external review committee (ERC),
+where the ERC served two purposes:
+\begin{itemize}
+\item To provide up to one knowledgeable (or expert) review per paper if
+prior knowledge of the topic was hard to find on the PC.
+\item To provide all reviews of papers with an author on the PC (``PC papers''),
+which were considered to
+pose an implicit conflict of interest to all PC members.
+\end{itemize}
+The special policy for ERC reviews of non-PC papers was realized via a more
+complex paper-side gadget, not described here.
+Based on an evaluation tool he wrote as well as manual inspection,
+Dr.~Hicks was satisfied that the decisions made by the matching tool
+closely reflected the best he could have done manually.
+
+We are looking for additional opportunities to apply the matching tool.
+Anyone interested is invited to contact us so we can help adapt it
+to the scenario and document the experience gained.
+
+%Waiting for data from NSF.
+
+%Synthetic Data.
+
+\section{Getting the Tool}
+
+A distribution containing the source code for the matching tool as well as this
+document may be browsed or downloaded at (NOT YET):
+\[\hbox{\url{https://mattmccutchen.net/match/}}\]
+There are currently two branches:
+\begin{itemize}
+\item \code{master} has the tool as originally designed for NSF, with no
+distinction between desirability and expertise.
+\item \code{popl2012} is the basis of the version used for POPL 2012. The main
+differences are that it has separate desirability and expertise, support for
+``fixing'' previously chosen reviewer-paper pairs (buggy, however),
+and the special ERC gadget.
+\end{itemize}
+We regret that we do not currently have a single well-maintained version of the
+tool to recommend.