You want to compete in specific Marathon competition, but it's rules (or part of the rules) looks unfamiliar and distinct from usual Marathon matches.
From time to time TopCoder features Marathon matches with specific rules that aimed to solve some real-world problem or test/promote new hardware or software platform.
Usually such matches require specific skills (knowledge of specific hardware, language, library or method) and have sponsors and monetary prizes.
(But some normal Marathon matches can have sponsors and prizes too).
There were several specific Marathon matches series in the history of TopCoder.
Here is brief description of all past (before October 2010) specific Marathon matches that should give you some understanding what specific Marathon match you may encounter in the future.
Intel Multi-Threading Competition series (12 matches, 2006)
There were twelve matches with well-suited for parallel programming problems in this series.
Submissions were tested on a dedicated Intel multi-processor server.
Unlike normal marathon matches, multi-threaded programs were allowed and Intel C++ compiler was used instead GCC.
AMD Multicore Threadfest (four matches, 2008)
There were four matches with well-suited for parallel programming problems (simulation and image-processing related).
Unlike normal marathon matches, multi-threaded programs were allowed and the only language to use was C++.
Predictive Marathon Competition 1 (Jun 2008)
The task was to predict the outcomes of TopCoder Component competitions based on large set of real historical data for these competitions.
FundingPrediction (Feb 2009)
Large amount of real data about loans for period of two years was provided as training data set.
The task was to create model for predicting outcome of loan funding based on the training data set.
Linden Lab OpenJPEG (Feb 2009)
The goal of this competition was to speed up decoding of the "JPEG 2000" format in open source OpenJPEG library to use in the "Second Life" virtual world.
Only C/C++ were allowed.
Experimental Event 1 (Apr 2009)
In this match contestants were working on sequence alignment problem - very common real-world bioinformatics problem.
The match was part of the scientific experiment conducted by Harvard Business School about solving problems individually and in groups and thus had very unusual format - all contestants were divided into 3 groups:
A - traditional Marathon match format (no code or idea sharing allowed)
B - traditional + sharing format (first week - individual competition, second - information sharing mode)
C - sharing code and ideas was allowed during the entire length of the competition
CUDA series (3 matches, Sep-Nov 2009)
Another one parallel programming competition series, but this time on GPU.
Contestants were allowed and encouraged to use nVidia CUDA API to run code on a graphics processing unit.
Submissions were tested on a dedicated server with powerful CUDA-enabled GPU.
AgentMatching (Aug 2009)
This contest was similar to the FundingPrediction contest.
The task was to create model for predicting outcome of real estate deals.
MessageDispatcher (Sep 2009)
Again data-based contest.
The task was to simulate a message dispatching system in a most efficient way.
NASA-TopCoder (Nov 2009)
Another one data-based contest.
The task was to design best space medkit model based on provided real-world data corpus.
The match was another one part of the scientific experiment conducted by Harvard Business School - all contestants were divided into 2 groups:
Group A - Traditional Marathon Match - individual work - i.e. no information sharing at all.
Group B - Teams - work in 5-member teams individuals, code and ideas sharing were allowed, score of a team was equal to the highest individual score in the team.