diff --git a/utils/token-delta.py b/utils/token-delta.py
new file mode 100755
index 0000000..327fa92
--- /dev/null
+++ b/utils/token-delta.py
@@ -0,0 +1,251 @@
+#!/usr/bin/env python
+
+import os
+import re
+import subprocess
+import sys
+import tempfile
+
+###
+
+class DeltaAlgorithm(object):
+    def __init__(self):
+        self.cache = set()
+
+    def test(self, changes):
+        abstract
+
+    ###
+
+    def getTestResult(self, changes):
+        # There is no reason to cache successful tests because we will
+        # always reduce the changeset when we see one.
+
+        changeset = frozenset(changes)
+        if changeset in self.cache:
+            return False
+        elif not self.test(changes):
+            self.cache.add(changeset)
+            return False
+        else:
+            return True
+
+    def run(self, changes, force=False):
+        # Make sure the initial test passes, if not then (a) either
+        # the user doesn't expect monotonicity, and we may end up
+        # doing O(N^2) tests, or (b) the test is wrong. Avoid the
+        # O(N^2) case unless user requests it.
+        if not force:
+            if not self.getTestResult(changes):
+                raise ValueError,'Initial test passed to delta fails.'
+
+        # Check empty set first to quickly find poor test functions.
+        if self.getTestResult(set()):
+            return set()
+        else:
+            return self.delta(changes, self.split(changes))
+
+    def split(self, S):
+        """split(set) -> [sets]
+
+        Partition a set into one or two pieces.
+        """
+
+        # There are many ways to split, we could do a better job with more
+        # context information (but then the API becomes grosser).
+        L = list(S)
+        mid = len(L)//2
+        if mid==0:
+            return L,
+        else:
+            return L[:mid],L[mid:]
+    
+    def delta(self, c, sets):
+        # assert(reduce(set.union, sets, set()) == c)
+
+        # If there is nothing left we can remove, we are done.
+        if len(sets) <= 1:
+            return c
+        
+        # Look for a passing subset.
+        res = self.search(c, sets)
+        if res is not None:
+            return res
+
+        # Otherwise, partition sets if possible; if not we are done.
+        refined = sum(map(list, map(self.split, sets)), [])
+        if len(refined) == len(sets):
+            return c
+        
+        return self.delta(c, refined)
+
+    def search(self, c, sets):
+        for i,S in enumerate(sets):
+            # If test passes on this subset alone, recurse.
+            if self.getTestResult(S):
+                return self.delta(S, self.split(S))
+
+            # Otherwise if we have more than two sets, see if test
+            # pases without this subset.
+            if len(sets) > 2:
+                complement = sum(sets[:i] + sets[i+1:],[])
+                if self.getTestResult(complement):
+                    return self.delta(complement, sets[:i] + sets[i+1:])
+
+###
+
+class Token:
+    def __init__(self, type, data, flags, file, line, column):
+        self.type   = type
+        self.data   = data
+        self.flags  = flags
+        self.file   = file
+        self.line   = line
+        self.column = column
+        
+kTokenRE = re.compile(r"""([a-z_]+) '(.*)'\t(.*)\tLoc=<(.*):(.*):(.*)>""",
+                      re.DOTALL | re.MULTILINE)
+
+def getTokens(path):
+    p = subprocess.Popen(['clang','-dump-raw-tokens',path],
+                         stdin=subprocess.PIPE,
+                         stdout=subprocess.PIPE,
+                         stderr=subprocess.PIPE)
+    out,err = p.communicate()
+
+    tokens = []
+    collect = None
+    for ln in err.split('\n'):
+        # Silly programmers refuse to print in simple machine readable
+        # formats. Whatever.
+        if collect is None:
+            collect = ln
+        else:
+            collect = collect + '\n' + ln
+        if 'Loc=<' in ln and ln.endswith('>'):
+            ln,collect = collect,None
+            tokens.append(Token(*kTokenRE.match(ln).groups()))
+
+    return tokens
+
+###
+
+class TMBDDelta(DeltaAlgorithm):
+    def __init__(self, testProgram, tokenLists, log):
+        def patchName(name, suffix):
+            base,ext = os.path.splitext(name)
+            return base + '.' + suffix + ext
+        super(TMBDDelta, self).__init__()
+        self.testProgram = testProgram
+        self.tokenLists = tokenLists
+        self.tempFiles = [patchName(f,'tmp')
+                            for f,_ in self.tokenLists]
+        self.targetFiles = [patchName(f,'ok')
+                            for f,_ in self.tokenLists]
+        self.log = log
+        self.numTests = 0
+
+    def writeFiles(self, changes, fileNames):
+        assert len(fileNames) == len(self.tokenLists)
+        byFile = [[] for i in self.tokenLists]
+        for i,j in changes:
+            byFile[i].append(j)
+
+        for i,(file,tokens) in enumerate(self.tokenLists):
+            f = open(fileNames[i],'w')
+            for j in byFile[i]:
+                f.write(tokens[j])
+            f.close()
+
+        return byFile
+
+    def test(self, changes):
+        self.numTests += 1
+
+        byFile = self.writeFiles(changes, self.tempFiles)
+
+        if self.log:
+            print >>sys.stderr, 'TEST - ',
+            if self.log > 1:
+                for i,(file,_) in enumerate(self.tokenLists):
+                    indices = byFile[i]
+                    if i:
+                        sys.stderr.write('\n      ')
+                    sys.stderr.write('%s:%d tokens: [' % (file,len(byFile[i])))
+                    prev = None
+                    for j in byFile[i]:
+                        if prev is None or j != prev + 1:
+                            if prev:
+                                sys.stderr.write('%d][' % prev)
+                            sys.stderr.write(str(j))
+                            sys.stderr.write(':')
+                        prev = j
+                    if byFile[i]:
+                        sys.stderr.write(str(byFile[i][-1]))
+                    sys.stderr.write('] ')
+            else:
+                print >>sys.stderr, ', '.join(['%s:%d tokens' % (file, len(byFile[i]))
+                                               for i,(file,_) in enumerate(self.tokenLists)]),
+
+        p = subprocess.Popen([self.testProgram] + self.tempFiles)
+        res = p.wait() == 0
+
+        if res:
+            self.writeFiles(changes, self.targetFiles)
+
+        if self.log:
+            print >>sys.stderr, '=> %s' % res
+        else:
+            if res:
+                print '\nSUCCESS (%d tokens)' % len(changes)
+            else:                
+                sys.stderr.write('.')
+
+        return res
+
+    def run(self):
+        res = super(TMBDDelta, self).run([(i,j)
+                                          for i,(file,tokens) in enumerate(self.tokenLists)
+                                          for j in range(len(tokens))])
+        self.writeFiles(res, self.targetFiles)
+        if not self.log:
+            print >>sys.stderr
+        return res
+
+def tokenBasedMultiDelta(program, files, log):            
+    # Read in the lists of tokens.
+    tokenLists = [(file, [t.data for t in getTokens(file)])
+                  for file in files]
+
+    numTokens = sum([len(tokens) for _,tokens in tokenLists])
+    print "Delta on %s with %d tokens." % (', '.join(files), numTokens)
+    
+    tbmd = TMBDDelta(program, tokenLists, log)
+
+    res = tbmd.run()
+
+    print "Finished %s with %d tokens (in %d tests)." % (', '.join(tbmd.targetFiles),
+                                                         len(res),
+                                                         tbmd.numTests)
+        
+def main():
+    from optparse import OptionParser, OptionGroup
+    parser = OptionParser("%prog <test program> {files+}")
+    parser.add_option("", "--debug", dest="debugLevel",
+                     help="set debug level [default %default]",
+                     action="store", type=int, default=0)
+    (opts, args) = parser.parse_args()
+
+    if len(args) <= 1:
+        parser.error('Invalid number of arguments.')
+        
+    program,files = args[0],args[1:]
+
+    md = tokenBasedMultiDelta(program, files, log=opts.debugLevel)
+        
+if __name__ == '__main__':
+    try:
+        main()
+    except KeyboardInterrupt:
+        print >>sys.stderr,'Interrupted.'
+        os._exit(1) # Avoid freeing our giant cache.