Difference between revisions of "Carver 2.0 Planning Page"

From Forensics Wiki
Jump to: navigation, search
(Filesystem Detection)
(Supported File Formats)
Line 104: Line 104:
** GIF
** GIF
** ZIP
** [[ZIP archive]] file format
** TAR (gz/bz2)
** TAR (gz/bz2)
Line 110: Line 110:
** Grapical Images
** Grapical Images
*** JPEG (the 3 different types with JFIF/EXIF support)
*** JPEG (the 3 different types with JFIF/EXIF support)
:: [[User:Joachim Metz|Joachim]] How different is JPEG 2000?
*** PNG
*** PNG
*** GIF
*** GIF
Line 115: Line 116:
*** TIFF
*** TIFF
** Office documents
** Office documents
*** OLE2 (Word/Excell content support)
*** Microsoft Office 97 - 2003 [[OLE Compound File]] format based with [[Word Document (DOC)]] and [[Excel Spreadsheet (XLS)]] file format support
*** PDF
*** [[PDF]]
*** Open Office/Office 2007 (ZIP+XML)
*** Open Office and Microsoft Office 2007 [[ZIP archive]] based file formats
:: Extension validation? AFAIK, MS Office 2007 [[DOCX]] format uses plain ZIP (or not?), and carved files will (or not?) have .zip extension instead of DOCX. Is there any way to fix this (may be using the file list in zip)? [[User:.FUF|.FUF]] 20:25, 31 October 2008 (UTC)
:: Extension validation? AFAIK, MS Office 2007 [[Word Document (DOCX)]] format uses plain ZIP (or not?), and carved files will (or not?) have .zip extension instead of DOCX. Is there any way to fix this (may be using the file list in zip)? [[User:.FUF|.FUF]] 20:25, 31 October 2008 (UTC)
:: [[User:Joachim Metz|Joachim]] Addition: Office 2007 also has a binary file format which is also a ZIP-ed data
:: [[User:Joachim Metz|Joachim]] Addition: Office 2007 also has a binary file format which is also a ZIP-ed data
==Archive Files==
==Archive Files==

Revision as of 10:50, 31 January 2009

This page is for planning Carver 2.0.

Please, do not delete text (ideas) here. Use something like this:

<s>bad idea</s>
:: good idea

This will look like:

bad idea

good idea




Joachim library based validators could require other licenses
Make the other libraries plug-able. If you them, you use them. Simsong 06:34, 3 November 2008 (UTC)



Shouldn't this just match what the underlying afflib & sleuthkit cover? RB
Yes, but you need to test and validate on each. Question: Do we want to support windows? Simsong 21:09, 30 October 2008 (UTC)
Joachim I think we would do wise to design with windows support from the start this will improve the platform independence from the start
Agreed; I would even settle at first for being able to run against Cygwin. Note that I don't even own or use a copy of Windows, but the vast majority of forensic investigators do. RB 14:01, 31 October 2008 (UTC)
Rob J Meijer Leaning heavily on the autotools might be the way to go. I do however feel that support requirements for windows would not be essential. Being able to run from a virtual machine with the main storage mounted over cifs should however be tested and if possible tuned extensively.
Joachim You'll need more than autotools to do native Windows support i.e. file access, UTF-16 support, wrap some basic system functions or have them available otherwise
Rob J Meijer That´s exactly my point, windows support as in being able to build and run on windows natively is much more trouble than its worth. Better make for a lean and mean autotools based build with little dependencies and no or little recursion, and better spent effort on a lean POLA design on POSIX based systems than on supporting building and running on non POSIX systems.

Name tooling

  • Joachim A name for the tooling I propose coldcut
How about 'butcher'?  ;) RB 14:20, 31 October 2008 (UTC)
Joachim cleaver ( scalpel on steroids ;-) )


Joachim Could we do a MoSCoW evaluation of these.

  • AFF and EWF file images supported from scratch. (Joachim I would like to have raw/split raw and device access as well)
If we base our image i/o on afflib, we get all three with one interface. RB Instead of letting the tools use afflib, better to write an afflib module for carvfs, and update the libewf module. The tool could than be oblivious of the file format. Rob J Meijer
Simsong 06:29, 3 November 2008 (UTC) The problem with using carvfs is that this adds another dependency. Do you really want to require that people install carvfs in order to run the carver? What about having the thing ported to Windows?
Rob J Meijer I would support adding one build dependency (libcarvpath) and removing two (libewf/libaff) by moving them to a layer more suited for them (carvfs) that would possibly allow some form of file handle (as cap) based POLA design. I am a proponent of making small things that do one and do one thing right, and to stack those to do what you need. In my view that would lead ideally to the following (simplified) chain:
  • recursive computer forensics framework (ocfa/pyflag)
    • The-(pola-based)-carving-tool
      • The-carving-lib working on open fd's.
        • libcarvpath
          • carvfs (Over cifs/nfs-v4 on platforms that don't support Fuse).
            • libewf
            • libaff
      • AppArmor (on supporting platforms)
      • suid (on supporting platforms)
      • iptables/ipfw (on supporting platforms)
As fow windows support, I would imagine making carvfs run over smb would come a long way, that is for as far as windows support is all that relevant.
There are two advantages to using libcarvpath and carvfs instead of libaff/libewf t this layer:
  • storage requirements for doing carving. Beyond what sleuthkit or alternatives provide I have seen many situations where carving was not done due to storage limitations.
  • File handles are like object capabilities. You can often do pretty simple POLA based implementations using file handles and something like AppArmor. POLA could IMHO be a strong weapon against the more nasty forms of anti forensics.
Next to this, I would consider making different tools for different stages instead of one semi recursive one, and looking at how to integrate these tools into existing frameworks (ocfa/pyflag).
Keep things simple but rigid and try to easily integrate things into existing frameworks as effectively as possible I would suggest.
Please note, I am not ptoposing the lib/tool should be useless without libcarvpath, only that usage without carvfs should limit the
supported image formats to raw images, and that libewf/libaff should be abstracted at the Fuse level or below and not at the tool level.
Joachim do you have an idea what the performance impact of this approach would be? It might be wise to do a proof of concept for this approach first.
Rob J Meijer It would I think depend greatly on behavior of the carving lib/tool. Small 512 byte reads are relatively very expensive, 128kb reads have negligible impact. Here are some numbers from ntfs-3g: http://article.gmane.org/gmane.comp.file-systems.fuse.devel/6397/match=ntfs+3g+performance+ext3 that might be relevant. More relevant than performance might be library footprint. For example, using OCFA, we often would want to keep e few hundred images that total to tens of TB of projected storage size fuse mounted. If libewf/libaff have a big combined memory footprint in such cases, this can be a major issue for this approach.
Joachim this layer should support multi threaded decompression of compressed image types, this speeds up IO
  • Joachim volume/partition aware layer (what about carving unpartioned space)
  • File system aware layer. This could be or make use of tsk-cp.
    • By default, files are not carved. (clarify: only identified? RB; I guess that it operates like Selective file dumper .FUF 07:00, 29 October 2008 (UTC)). Alternatively, the tool could use libcarvpath and output carvpaths or create a directory with symlinks to carvpaths that point into a carvfs mountpoint Rob J Meijer.
  • Plug-in architecture for identification/validation.
    • Joachim support for multiple types of validators
      • dedicated validator
      • validator based on file library (i.e. we could specify/implement a file structure API for these)
      • configuration based validator (Can handle config files,like Revit07, to enter different file formats used by the carver.)

Joachim Moderator: Could we limit the requirements for prototype version 1 of the tool to get a working version up and running ASAP? And keep discussing future options?

I think the following set will be large enough to handle: Input facilities

  • IO support (AFF, device, EWF, RAW and split RAW)
Abstraction of input format and multi threaded decompression (spin-off code out of afflib?)
  • Volume/Partitions support
at least for DOS based layout and GPT (spin-off code out of TSK/Photorec?)
  • File system support
VFAT/NTFS (spin-off code out of TSK/Photorec?)

Carving facilities

  • File format support using plug-able validator model (use dedicated validators Photorec/Scarve and/or wrap revit07 file format as validator?)
  • Content support using plug-able validator model (to handle text/mbox base64)
  • File system carving support (to handle file system fragments, could be linked to file system support layer?)
  • Basic fragment handling

Output facilities

  • audit/analysis/debug log
  • extraction of result files

Supported File Formats

  • Ship with validators for:

Joachim I think we should distinguish between file format validators and content validators

    • JPEG
    • PNG
    • GIF
    • MSOLE
    • ZIP archive file format
    • TAR (gz/bz2)

Joachim For a production carver we need at least the following formats

    • Grapical Images
      • JPEG (the 3 different types with JFIF/EXIF support)
Joachim How different is JPEG 2000?
Extension validation? AFAIK, MS Office 2007 Word Document (DOCX) format uses plain ZIP (or not?), and carved files will (or not?) have .zip extension instead of DOCX. Is there any way to fix this (may be using the file list in zip)? .FUF 20:25, 31 October 2008 (UTC)
Joachim Addition: Office 2007 also has a binary file format which is also a ZIP-ed data

Archive Files

    • Archive files
      • ZIP
      • 7z
      • gzip
      • bzip2
      • tar
      • RAR
    • E-mail files
      • PFF (PST/OST)
      • MBOX (text based format, base64 content support)
    • Audio/Video files
      • MPEG
      • MP2/MP3
      • AVI
      • ASF/WMV
      • QuickTime
      • MKV
    • Printer spool files
      • EMF (if I remember correctly)
    • Internet history files
      • index.dat
      • firefox (sqllite 3)
    • Other files
      • thumbs.db
      • pagefile?

Carving Strategies

Joachim Note to moderator could this section be merged with the carving algorithm section?

  • Simple fragment recovery carving using gap carving.
    • Joachim have hook in for more advanced fragment recovery?
  • Recovering of individual ZIP sections and JPEG icons that are not sector aligned.
    • Joachim I would propose a generic fragment detection and recovery
  • Autonomous operation (some mode of operation should be completely non-interactive, requiring no human intervention to complete RB)
    • Joachim as much as possible, but allow to be overwritten by user
  • Joachim When the tool output files the filenames should contain the offset in the input data (in hexadecimal?)
Mark I really like the fact carved files are named after the physical or logical sector in which the file is found (photorec)
Joachim This naming schema might cause duplicate name problem for extracting embedded files and extracting files from non sector aligned file systems.
  • Joachim Should the tool allow to export embedded files?
  • Joachim Should the tool allow to export fragments separately?
  • Mark I personally use photorec often for carving files in the whole volume (not only unallocated clusters), so I can store information about all potential interesting files in MySQL
Joachim interesting, Bas Kloet and me have been discussing to use information about allocated files in the recovery process, i.e. recovered fragments could be part of allocated files. Do we want to be able to extract them? Or could we rebuild the file from the fragments and the allocated files.
  • Mark It would also be nice if the files can be hashed immediately (MD5) so looking for them in other tools (for example Encase) is a snap

Performance Requirements

  • Tested on 500GB-sized images. Should be able to carve a 500GB image in roughly 50% longer than it takes to read the image.
    • Perhaps allocate a percentage budget per-validator (i.e. each validator adds N% to the carving time) RB
    • Joachim have multiple carving phases for precision/speed trade off?
  • Parallelizable
    • Joachim tunable for different architectures
  • Configuration:
    • Capability to parse some existing carvers' configuration files, either on-the-fly or as a one-way converter.
    • Disengage internal configuration structure from configuration files, create parsers that present the expected structure
    • Joachim The validator should deal with the file structure the carving algorithm should not know anything about the file structure (as in revit07 design)
    • Either extend Scalpel/Foremost syntaxes for extended features or use a tertiary syntax (Joachim I would prefer a derivative of the revit07 configuration syntax which already has encountered some problems of dealing with defining file structure in a configuration file)


  • Can output audit.txt file.
  • Joachim Can output database with offset analysis values i.e. for visualization tooling
  • Joachim Can output debug log for debugging the algorithm/validation
  • Easy integration into ascription software.
Joachim I'm no native speaker what do you mean with "ascription software"?
I think this was another non-native requesting easy scriptability. RB 14:20, 31 October 2008 (UTC)
Joachim that makes sense ;-)
Incorrect. Ascription software is software that determines who the owner of a file is. Simsong 06:36, 3 November 2008 (UTC)


  • Use as much TSK if possible. Don't carry your own FS implementation the way photorec does.
Joachim using TSK as much as possible would not allow to add your own file system support (i.e. mobile phones, memory structures, cap files) I would propose wrapping TSK and using it as much as possible but allow to integrate own FS implementations.
  • Extracting/carving data from Thumbs.db? I've used foremost for it with some success. Vinetto has some critical bugs :( .FUF 19:18, 28 October 2008 (UTC)

Recursive Carving

Joachim do we want to support (let's call it) 'recursive in file carving' (for now) this is different from embedded files because there is a file system structure in the file and not just another file structure

  • Is it just me, or do a lot of the above (and below) ideas somewhat skirt around the fact that many of us want recursive carving? Can we bend back to that instead of discussing object particulars? I think this can be distilled down to three requirements:
    • Simple recursion: once an object is identified, have the ability to re-carve it for internal structures
    • Directed recursion: the carver should be able to be directed at arbitrary blobs and told to carve it as a specified type. This allows programmatically more simple methods of dealing with unidentifiably compressed or encrypted data. Or filesystem fragments.
    • Export: the ability to export an object (recognized or not) for later or external "recursion". Should go without saying for a carver, but...
--RB 18:45, 2 November 2008 (UTC)
Simsong 06:30, 3 November 2008 (UTC) pyflag already does recursive carving. Are we just going to reimplement pyflag as a single executable?

Library Dependencies

Rob J Meijer :

  • Use libcarvpath whenever possible and by default to avoid high storage requirements.
Joachim For easy deployment I would not opt for making an integral part of the tool solely dependant on a single external library or the library must be integrated in the package
Rob J Meijer Integrating libraries (libtsk,libaff.libewf,libcarvpath etc) is bad practice, autotools are your friend IMO.
Joachim I'm not talking about integrating (shared) libraries. I'm talking about that an integral part of a tool should be part of it's package. Why can't the tool package contain shared or static libraries for local use? A far worse thing to do is to have a large set of dependencies and making the tool difficult to install for most users. The tool package should contain the most necessary code. afflib/libewf support could be detected by the autotools a neat separation of functionality.
From a packager's standpoint, Joachim's other libraries do a really good job of this, carrying around what they need but using a system-global version if available. RB
  • libtsk
  • libaff ? : possibly the discussion in the requirements section should move to his section.
  • libewf ? : possibly the discussion in the requirements section should move to his section.
  • posix ? : Can we depend especially on the availability of UNIX domain sockets and the possibility to use msg_accrights for passing opn file handles as ocaps?

Filesystem Detection

  • Dont stop with filesystem detection after the first match. Often if a partition is reused with a new FS and is not all that full yet, much of the old FS can still be valid. I have seen this with ext2/fat. The fact that you have identified a valid FS on a partition doesn't mean there isn't an(almost) valid second FS that would yield additional files. Identifying doubly allocated space might in some cases also be relevant.
Joachim What your saying is that dealing with file system fragments should be part of the carving algorithm
  • Allow use where filesystem based carving is done by other tool, and the tool is used as second stage on (sets of) unallocated block (pseudo) files and/or non FS partition (pseudo) files.
Joachim I would not opt for this. The tool would be dependent on other tools and their data format, which makes the tool difficult to maintain. I would opt to integrate the functionality of having multiple recovery phases (stages) and allow the tooling to run the phases after one and other or separately.
Rob J Meijer More generically, I feel a way should exist to communicate the 'left overs' a previous (non open, for example LE-only) tool left.
Joachim I guess if the tool is designed to handle multiple phases it should store its data somewhere. So it should be possible to convert results of such non open tooling to the format required. However I would opt to design the recovery functionality of these non-open tools into open tools. And not to limit ourselves making translators due to the design of these non-open tools.
  • Ability to be used as a library instead of a tool. Ability to access metadata true library, and thus the ability to set metadata from the carving modules. This would be extremely usefull for integrating the project into a computer forensics framework .
Joachim I guess most of the code could be integrated into libraries, but I would not opt integrating tool functionality into a library
  • Mark I think it would be very handy to have a CSV, TSV, XML or other delimited output (log)file with information about carved files. This output file can then be stored in a database or Excel sheet (report function)

Anti forensics and system integrity concerns

  • It might be very interesting to look at the possibilities of using a multi process style of module support and combine it with a least authority design. On platforms that support AppArmor (or similar) and uid based firewall rules, this could make for the first true POLA (principle of least authority) based forensic tool ever. POLA based forensics tools should make for a strong integrity guard against many anti forensics. Alternatively we could look at integrating a capability secure language (E?) for implementation of at least validation modules. I don't expect this idea to make it, but mentioning it I hope might spark off less strong alternatives that at least partially address the integrity + anti-forensics problem. If we can in some way introduce POLA to a wider forensics public, other tools might also pick up on it what would be great.
Joachim Could you give an example of how you see this in action?
Rob J Meijer I see two layers where using POLA could be applied. The best one would require one of the folowing as prerequisites:
  • The libaff/libewf layer is moved to a fuse implementation (for example carvfs).
  • Libewf/Libaff are updated to accept opened filhandles instead of demanding to open their own files.
If one of these is fulfilled, than the tool running as some user can just have the simple task of opening the image files, starting up the 'real' tool and handing over the appropriate file handles. If the real tool runs with a restrictive AppArmor profile, and is started suid to a tool specific user that also has its own iptables uid based filter, than the real tool will run with least authority.
A second alternative, if neither of the first prerequisite could not be bet, would be to run the modules as confined processes and have a non confined process run as proxy for the first.
A third probably far fetched alternative would be to embed an object capability language in the tool and make the module interface thus that modules are to be written in this ocap language.
A 4th alternative might include minorfs or plash, but I havn't geven those sufficient thinking hours yet.

Format syntax specification

  • Carving data structures. For example, extract all TCP headers from image by defining TCP header structure and some fields (e.g. source port > 1024, dest port = 80). This will extract all data matching the pattern and write a file with other fields. Another example is carving INFO2 structures and URL activity records from index.dat .FUF 20:51, 28 October 2008 (UTC)
    • This has the opportunity to be extended to the concept of "point at blob FOO and interpret it as BAR"

.FUF added: The main idea is to allow users to define structures, for example (in pascal-like form):

Field1: Byte = 123;
SomeTextLength: DWORD;
SomeText: string[SomeTextLength];
Field4: Char = 'r';

This will produce something like this:

Field1 = 123
SomeTextLength = 5
SomeText = 'abcd1'
Field4 = 'r'

(In text or raw forms.)


Opinion: Simple pattern identification like that may not suffice, I think Simson's original intent was not only to identify but to allow for validation routines (plugins, as the original wording was). As such, the format syntax would need to implement a large chunk of some programming language in order to be sufficiently flexible. RB

Joachim In my option your example is too limited. Making the revit configuration I learned you'll need a near programming language to specify some file formats. A simple descriptive language is too limiting. I would also go for 2 bytes with endianess instead of using terminology like WORD and small integer, it's much more clear. The configuration also needs to deal with aspects like cardinality, required and optional structures.

This is simply data structures carving, see ideas above. Somebody (I cannot track so many changes per day) separated the original text. There is no need to count and join different structures. .FUF 19:53, 31 October 2008 (UTC)
Joachim This was probably me is the text back in it's original form?
I started it by moving your Revit07 comment to the validator/plugin section in this edit, since I was still at that point thinking operational configuration for that section, not parser configurations. RB
Joachim I renamed the title to format syntax, clarity is important ;-)

Please take a look at the revit07 format syntax specification (configuration). It's not there yet but goes a far way. Some things currently missing:

  • bitwise alignment
  • handling encapsulated streams (MPEG/capture files)
  • handling content based formats (MBOX)

Caving algorithm


  • should we allow for multiple carving phases (runs/stages)?
I opt yes (separation of concern)
  • should we allow for multiple carving algorithms?
I opt yes, this allows testing of different approaches
  • Should the algorithm try to do as much in 1 run over the input data? To reduce IO?
I opt that the tool should allow for multiple and single run over the input data to minimize the IO or the CPU as bottleneck
  • Interaction between algorithm and validators
    • does the algorithm passes data blocks to the validators?
    • does a validator need to maintain a state?
    • does a validator need to revert a state?
    • How do we deal with embedded files and content validation? Do the validators call another validator?
  • do we use the assumption that a data block can be used by a single file (with the exception of embedded/encapsulated files)?
  • Revit07 allows for multiple concurrent result files states to deal with fragmentation. One has the attribute of being active (the preferred) and the other passive. Do we want/need something similar? The algorithm adds block of input data (offsets) to these result files states.
    • if so what info would these result files states require (type, list of input data blocks)
  • how do we deal with file system remainders?
    • Can we abstract them and compare them against available file system information?
  • Do we carve file systems in files?
I opt that at least the validator uses this information

Caving scenarios


  • normal file (file structure, loose text based structure (more a content structure?))
  • fragmented file (the file entirely exist)
  • a file fragment (the file does not entirely exist)
  • intertwined file
  • encapsulated file (MPEG/network capture)
  • embedded file (JPEG thumbnail)
  • obfuscation ('encrypted' PFF) this also entails encryption and/or compression
  • file system in file

File System Awareness

Background: Why be File System Aware?

Advantages of being FS aware:

  • You can pick up sector allocation sizes
Joachim do you mean file system block sizes?
  • Some file systems may store things off sector boundaries. (ReiserFS with tail packing)
  • Increasingly file systems have compression (NTFS compression)
  • Carve just the sectors that are not in allocated files.

Tasks that would be required


As noted above, TSK should be utilized as much as possible, particularly the filesystem-aware portion. If we want to identify filesystems outside of its supported set, it would be more worth our time to work on implementing them there than in the carver itself. RB
I guess this tool operates like Selective file dumper and can recover files in both ways (or not?). Recovering files by using carving can recover files in situations where sleuthkit does nothing (e.g. file on NTFS was deleted using ntfs-3g, or filesystem was destroyed or just unknown). And we should build the list of filesystems supported by carver, not by TSK. .FUF 07:08, 29 October 2008 (UTC)
This tool is still in the early planning stages (requirements discovery), hence few operational details (like precise modes of operation) have been fleshed out - those will and should come later. The justification for strictly using TSK for the filesystem-sensitive approach is simple: TSK has good filesystem APIs, and it would be foolish to create yet another standalone, incompatible implementation of filesystem(foo) when time would be better spent improving those in TSK, aiding other methods of analysis as well. This is the same reason individuals that have implemented several other carvers are participating: de-duplication of effort. RB
Joachim A design problem might be that TSK currently is a single library operating on multiple layers (storage media IO, volume/partition analysis and file system analysis). I'm not aware how easily the parts can be used separately. But I estimate that for the carver we want to use these 3 layers differently than TSK currently does.

Joachim I would like to have the carver (recovery tool) also do recovery using file allocation data or remainders of file allocation data.

Joachim I would go as far to ask you all to look beyond the carver as a tool and look from the perspective of the carver as part of the forensic investigation process. In my eyes certain information needed/acquired by the carver could be also very useful investigative information i.e. what part of a hard disk contains empty sectors.

Supportive tooling


  • validator (definitions) tester (detest in revit07)
  • tool to make configuration based definitions
  • post carving validation
  • the carver needs to provide support for fuse mount of carved files (carvfs)



  • automated testing
  • test data

Validator Construction


  • Write validators in C/C++
    • Joachim you mean dedicated validators
  • Have a scripting language for writing them (python? Perl?) our own?
    • Joachim use easy to embed programming languages i.e. Phyton or Lua
  • Use existing programs (libjpeg?) as plug-in validators?
    • Joachim define a file structure api for this

Existing Code that we have

Joachim Please add any missing links


  • DFRWS2006/2007 carving challenge results
  • DFRWS2008 paper on carving


Possible file structure validator libraries

Input support

Volume/Partition support

File system support

  • TSK
  • photorec FS code
  • implementations of FS in Linux/BSD

Content support

Zero storage support


Implementation Timeline

  1. gather the available resources/ideas/wishes/needs etc. (I guess we're in this phase)
  2. start discussing a high level design (in terms of algorithm, facilities, information needed)
    1. input formats facility
    2. partition/volume facility
    3. file system facility
    4. file format facility
    5. content facility
    6. how to deal with fragment detection (do the validators allow for fragment detection?)
    7. how to deal with recombination of fragments
    8. do we want multiple carving phases in light of speed/precision tradeoffs
  3. start detailing parts of the design
    1. Discuss options for a grammar driven validator?
    2. Hard-coded plug-ins?
    3. Which existing code can we use?
  4. start building/assembling parts of the tooling for a prototype
    1. Implement simple file carving with validation.
    2. Implement gap carving
  5. Initial Release
  6. Implement the threaded carving that .FUF is describing above.

Joachim Shouldn't multi threaded carving (MTC) not be part of the 1st version? The MT approach makes for different design decisions

It is virtually impossible to turn a non-MT application into an MT application .Simsong 06:37, 3 November 2008 (UTC)