Changing the sequence of variables in a random forest model changes the classification accuracy I find that the classification accuracy of the random forest model changes when the sequence of the input variables change.
E.g. 
set.seed(143)
artCheck.forest<-randomForest(GoodArt.check~Brush+Min.Guarantee.Cost+Top.3.artists+Brush.Size+Brush.Finesse+Art.Nationality+Art.Type, data=noYes,imp=T,type='classification')

The confusion matrix is
    NO YES class.error
NO  84  12   0.1250000
YES 15  47   0.2419355

When the "Brush" variable is moved to a different position, say
set.seed(143)
artCheck.forest<-randomForest(GoodArt.check~Min.Guarantee.Cost+Top.3.artists+Brush+Brush.Size+Brush.Finesse+Art.Nationality+Art.Type, data=noYes,imp=T,type='classification')

The confusion matrix becomes
    NO YES class.error
NO  83  13  0.13541667
YES  5  57  0.08064516

The sequence of the variable names on the dataframe is as follows :
1   Art Auction House
2   IsGood Purchase
3   Critic Ratings
4   Buyer No
5   Zip Code
6   Art Purchase Date
7   Year of art piece
8   Acq Cost
9   Art Category
10  Size
11  Length
12  Width
13  SizeRatio
14  Border of art piece
15  Art Type
16  Prominent Color
17  CurrentAuctionAveragePrice
18  Premium
19  Brush
20  Brush Size
21  Brush Finesse
22  Art Nationality
23  Top 3 artists 
24  CollectorsAverageprice
25  Profit
26  GoodArt check
27  AuctionHouseGuarantee
28  Vnst
29  Is It Online Sale
30  Min Guarantee Cost

 A: Proportion classified correctly is not only an improper scoring rule (one that is optimized by choosing the wrong features and giving them the wrong weights) but it is extremely information-losing and has high variance.  Make sure you can get probabilities out of the random forest, and consider using the Brier score or other proper scoring rules in judging predictions.
For more information see Section 17.4 and Chapter 18 of Biostatistics for Biomedical Research from http://biostat.mc.vanderbilt.edu/ClinStat as well as http://www.citeulike.org/user/harrelfe/tag/proper-scoring-rule .
A: I think the reason you are having different results is because, even if you set the seed, you obtain trees on different variables changing their sequence.
Moreover, as Frank said, classification accuracy has high variance. How did you compute the confusion matrix? Try to get a lower variance confusion matrix based on leave-on-out given that you have just a few records. Also, you might want to increase ntree to get less variance given that the default is 500.
