Here is my description of Minolta RAW file format. It's based on work of Paul K. Hansen and Jake. All this information is unconfirmed by Minolta and is just my guess based on experiments and hex editor so use it carefully. If you have any comments please let me know by e-mail.

Jake's description is here (or my mirror here)
Paul K. Hansen's work is here (or as a text file)
Thanks to Patrick Porlan for help

Version 1.02 published on 25.10.2002. Changes in: MRM, WBG, values > 3965.
Version 1.03 published on 30.7.2004. Changes: some info in A1,A2; new information in PRD block deciphered by Cliff Biffle.
Version 1.04 published on 1.6.2005. Changes: some info for A200; fixed typos in WBG block.

Basic structure of .mrw file is here:

Every block in MRM (Minolta Raw M???) has this structure:

The length of whole block is then "Block Length" + 8.

0. MRM Block (Minolta Raw M???)

This superblock contains all the rest blocks (except image data)

1. PRD Block (Picture Raw Dimensions)

All numbers are stored in big endian format (Macintosh/Motorola style). Offset is counted from start of Block Data. This means that you have to add 8 to get offset from start of block. Size is in bytes.

2. TTW Block (Tiff Tags W??)

This is a classic TIFF header as described in TIFF format specification ver. 6.0

Description of EXIF tags can be found JEITA CP-3451 Standard
I do not know where to get a description of PIM tags. If you know it please let me know by e-mail.
Minolta MakerNote tag (part of EXIF) specification is not publicly available. I have decoded some parts of it and I will publish its structure some day on my website.

3. WBG Block (White Balance Gains)

Usually those coefficient are the same as those in MakerNote in EXIF. This is when G=G'=1.
But there are some shots where G and G's is not equal to one. In this case the MakerNote has different white balance coefficients this time with G=1.
When user chooses in DIVU to use white balance as measured by camera the RAW image data are multiplied by these white balance coefficients.
The meaning of those coefficients depends on the Bayer pattern used by the camera. Usually it is RGGB, but A200 uses GBRG which does change the meaning of the WB coefficients above.

4. RIF Block (Requested Image Format)

Those values do not reflect camera setting when taking pictures. You can change them with DIVU when you choose different setting for RAW file and save back (works only with DIVU 1, DIVU 2 doesn't save values back to RAW file.).
There are much more values but I was able to decrypt only those in the table. My problem is that I have only a few raw files on my harddisk. If you have a large RAW collection and you are willing to help with the rest please let me know.

5. PAD Block (Padding)

Here are just zeroes used to create a gap so that Raw Image Data start at 512 bytes boundary.

Raw Image Data

The rest of the file are RAW image data read from CCD.
They are stored sequentially in lines.
Length of the line is defined in PRD block.
The number of lines is in PRD block also.
Odd lines are RGRGRG..., even lines are GBGBGB...
Values are 12bit numbers stored in two bytes in big endian byte order.
D5, D7xx: Two pixel values 0x123, 0x456 are stored as 01 23 04 56.
A1, A2: Image data are packed. Two pixel values 0x123, 0x456 are stored as 12 34 56. This method reduces the file size by cca 25%.

It may seem that the largest pixel value is is 0xF7D = 3965
but this in not true there are some pictures with full range of values up to 0xFFF = 4095.
According to latest information by Mark Horton the values >3965 appears only in
pictures taken with aperture f2.8. Because of a large aperture there is a vignetting
in the corners of the image. So the camera multiplies corner values to reduce
this effect. This could then result in values larger than 3965. Here is a picture which
roughly shows the affected areas:

This is a new vignetting compensation mask used in A1 and A2:

There are eight pixels extra in horizontal and vertical directions.
They are lost when interpolating Bayer pattern.
Nice description of Bayer interpolation algorithms by Ting Chen can be found here (here is my pdf mirror).
Very simple one is here and C source of conversion algorithm is here

Nice article about CCD sensors is here.
A site devoted to CCD technology is here.