@Grauw If you want to take that approach then (VALUE+16)/32 is better... The problem is that you think that 255 should become 7 while actually you should think that 256 should become 8.
No, that way 8-bit values 0-47 translate to 3-bit value 0, you get way too much error there.
Simplifying the problem to grayscale, black is black and white is white. So the 8-bit (0-255) values are on a different scale than the 3-bit (0-7) values, they can’t be converted by simply dropping bits, you need to consider them on a linear scale as 0-1 where 0 is black and 1 is white.
The linear scale for 3-bit values is: 0.0, 0.143, 0.286, 0.429, 0.571, 0.714, 0.857, 1.0.
The linear scale for 8-bit values I’ll let you generate by yourself 
. But, for each value on the 8-bit scale, you need to pick the 3-bit scale’s value that is closest, in order to minimise the error.
There's probably a better solution : screen 8 with his 256 colors !
Why degrad the colors to fit in screen 5 or 7 ?
Palette colors are fewer in number, but have a little bit more accuracy than screen 8 RGB values (to be exact: 1 bit extra for the blue value). So depending on input, conversion to screen 5 or 7 may provide better result than conversion to screen 8. Well okay - probably not for most images... 
If output as YJK data is allowed, then you can get much better results. Conversion formulas for this are in the V9958 data book.
Btw. good info there, Grauw!
Uh... From the first post I could not see this coming such a big conversation
Try to do the conversion the other way around. What is the 24-bit RGB value of (7, 3, 0)?
Mmm perhaps you are right, your method needs less "human retouch" 
.
Uh... From the first post I could not see this coming such a big conversation
But it hasn't yet taken into acount the nonlinear DAC levels 
In the below code I got them in that l() function, I forgot from where I got them?
At the begining the differences between DAC levels are printed.
32, 40, 32, 40... mhm was that again about hardware savings or is it a trick suiting artists purposes?
some colors are in more harsh distance than 36, some are a more closer pack.
how does this fit a linear sheme of 9958 VUY mode - the 32,40,32,40 jumps would fit into a linear 5bit DAC, I assume things look identic on 9958.
the levels
function l(n) {
c = new Array();
c[0] = 0x06;
c[1] = 0x20;
c[2] = 0x48;
c[3] = 0x68;
c[4] = 0x90;
c[5] = 0xb0;
c[6] = 0xd8;
c[7] = 0xf7;
return c[n];
}
the diffs:
0-1: 26
1-2: 40
2-3: 32
3-4: 40
4-5: 32
5-6: 40
6-7: 31
the html file showing the colors demo
function hex(n) {
n1 = n / 16; n2 = n%16;
s = "0123456789abcdef";
return "" + s.charAt(n1) + s.charAt(n2)
}
function l(n) {
c = new Array();
c[0] = 0x06;
c[1] = 0x20;
c[2] = 0x48;
c[3] = 0x68;
c[4] = 0x90;
c[5] = 0xb0;
c[6] = 0xd8;
c[7] = 0xf7;
return c[n];
}
for (i = 0; i < 7; i++) { document.write(''+i+'-'+(i+1)+': '+(l(i+1)-l(i))+'<br>'); }
document.write('<table border="0" cellspacing="32" cellpadding="0"><tr><td>');
document.write('<table border="0" cellspacing="0" cellpadding="0"><tr>');
for (g = 0; g <= 7; g++) {
document.write("<tr>");
for (b = 0; b <= 7; b++) {
for (r = 0; r <= 7; r++) {
document.write('<td width="16" height="16">');
document.write(" ");
document.write("</td>");
}
}
document.write("</tr>");
}
document.write('</tr><table>');
document.write('</td></tr><table>');