diff --git a/TODO.org b/TODO.org
index 1716ec28..8da01239 100644
--- a/TODO.org
+++ b/TODO.org
@@ -1,11 +1,14 @@
 * untypy
 ** add tests with type annotations
 *** for records
-**** construction
 **** write access
+**** test for correct location
+** understand why commit 150c7a5b209ddb264fd104feae42eba4905c2be7 is necessary to fix the replTester test
 * other
+** mode where all functions or methods without type signatures are reported as errors
 ** upload dist to pypi
 ** update README
+*** python version
 *** pip
 *** records
 *** import
diff --git a/python/.ignore b/python/.ignore
new file mode 100644
index 00000000..eed76414
--- /dev/null
+++ b/python/.ignore
@@ -0,0 +1,2 @@
+site-lib/wypp
+site-lib/untypy
diff --git a/python/deps/untypy b/python/deps/untypy
index 5e58090c..16ce8157 160000
--- a/python/deps/untypy
+++ b/python/deps/untypy
@@ -1 +1 @@
-Subproject commit 5e58090c59b5a382a596e67718e64bc00bce3916
+Subproject commit 16ce81576c36bcd15b0db3d3147416921823f450
diff --git a/python/file-tests/testArgs.out b/python/file-tests/testArgs.out
deleted file mode 100644
index 6840595f..00000000
--- a/python/file-tests/testArgs.out
+++ /dev/null
@@ -1 +0,0 @@
-['file-tests/testArgs.py', 'ARG_1', 'ARG_2']
diff --git a/python/file-tests/testTraceback.err b/python/file-tests/testTraceback.err
deleted file mode 100644
index e20739f5..00000000
--- a/python/file-tests/testTraceback.err
+++ /dev/null
@@ -1,7 +0,0 @@
-
-Traceback (most recent call last):
-  File "file-tests/testTraceback.py", line 9, in <module>
-    foo(lst)
-  File "file-tests/testTraceback.py", line 7, in foo
-    print(lst[10])
-IndexError: list index out of range
diff --git a/python/file-tests/testTraceback2.err b/python/file-tests/testTraceback2.err
deleted file mode 100644
index 55342d7e..00000000
--- a/python/file-tests/testTraceback2.err
+++ /dev/null
@@ -1,5 +0,0 @@
-
-  File "file-tests/testTraceback2.py", line 3
-    lst = [1,2,3
-               ^
-SyntaxError: unexpected EOF while parsing
diff --git a/python/file-tests/testTypes1.err b/python/file-tests/testTypes1.err
deleted file mode 100644
index 86fffb7c..00000000
--- a/python/file-tests/testTypes1.err
+++ /dev/null
@@ -1,17 +0,0 @@
-
-given: '1'
-expected: int
-          ^^^
-
-inside of inc(x: int) -> int
-                 ^^^        
-declared at: 
-file-tests/testTypes1.py:1
-  1 | def inc(x: int) -> int:
-  2 |     return x + 1
-
-
-caused by: 
-file-tests/testTypes1.py:4
-  4 | inc("1")
-
diff --git a/python/file-tests/testTypes2.err b/python/file-tests/testTypes2.err
deleted file mode 100644
index d3bc73e6..00000000
--- a/python/file-tests/testTypes2.err
+++ /dev/null
@@ -1,17 +0,0 @@
-
-given: '1'
-expected: int
-          ^^^
-
-inside of inc(x: int) -> int
-                 ^^^        
-declared at: 
-file-tests/testTypes2.py:1
-  1 | def inc(x: int) -> int:
-  2 |     return x
-
-
-caused by: 
-file-tests/testTypes2.py:4
-  4 | inc("1")
-
diff --git a/python/integration-tests/shell.py b/python/integration-tests/shell.py
index 7e573fc6..0d53673d 100644
--- a/python/integration-tests/shell.py
+++ b/python/integration-tests/shell.py
@@ -322,10 +322,12 @@ def hook(self):
 
     def exit(self, code=0):
         if code is None:
-            code = 0
+            myCode = 0
         elif type(code) != int:
-            code = 1
-        self.exitCode = code
+            myCode = 1
+        else:
+            myCode = code
+        self.exitCode = myCode
         self._origExit(code)
 
     def exc_handler(self, exc_type, exc, *args):
diff --git a/python/integration-tests/testIntegration.py b/python/integration-tests/testIntegration.py
index 51cf445a..d59cf9d4 100644
--- a/python/integration-tests/testIntegration.py
+++ b/python/integration-tests/testIntegration.py
@@ -32,67 +32,65 @@ def stripTrailingWs(s):
 
 LOG_FILE = shell.mkTempFile(prefix="wypp-tests", suffix=".log", deleteAtExit='ifSuccess')
 print(f'Output of integration tests goes to {LOG_FILE}')
-LOG_REDIR = f'> {LOG_FILE} 2>&1'
+LOG_REDIR = f'>> {LOG_FILE} 2>&1'
 
 class TypeTests(unittest.TestCase):
     def test_enumOk(self):
-        out = runInteractive('file-tests/typeEnums.py', 'colorToNumber("red")')
+        out = runInteractive('test-data/typeEnums.py', 'colorToNumber("red")')
         self.assertEqual(['0'], out)
 
     def test_enumTypeError(self):
-        out = runInteractive('file-tests/typeEnums.py', 'colorToNumber(1)')[0]
-        self.assertIn("expected: Literal['red', 'yellow', 'green']", out)
+        out = runInteractive('test-data/typeEnums.py', 'colorToNumber(1)')[0]
+        self.assertIn("expected: value of type Literal['red', 'yellow', 'green']", out)
 
     def test_recordOk(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out1 = runInteractive(rec, 'Person("stefan", 42)')
         self.assertEqual(["Person(name='stefan', age=42)"], out1)
         out2 = runInteractive(rec, 'incAge(Person("stefan", 42))')
         self.assertEqual(["Person(name='stefan', age=43)"], out2)
 
-    @unittest.skip
     def test_recordFail1(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out = runInteractive(rec, 'Person("stefan", 42.3)')[0]
-        self.assertIn('expected: int', out)
+        self.assertIn('expected: value of type int', out)
 
     def test_recordFail2(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out = runInteractive(rec, 'mutableIncAge(Person("stefan", 42))')[0]
-        self.assertIn('expected: MutablePerson', out)
+        self.assertIn('expected: value of type MutablePerson', out)
 
     def test_recordMutableOk(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out1 = runInteractive(rec, 'MutablePerson("stefan", 42)')
         self.assertEqual(["MutablePerson(name='stefan', age=42)"], out1)
         out2 = runInteractive(rec, 'p = MutablePerson("stefan", 42)\nmutableIncAge(p)\np')
         self.assertEqual(['', '', "MutablePerson(name='stefan', age=43)"], out2)
 
-    @unittest.skip
     def test_mutableRecordFail1(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out = runInteractive(rec, 'MutablePerson("stefan", 42.3)')[0]
-        self.assertIn('expected: int', out)
+        self.assertIn('expected: value of type int', out)
 
     def test_mutableRecordFail2(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out = runInteractive(rec, 'incAge(MutablePerson("stefan", 42))')[0]
-        self.assertIn('expected: Person', out)
+        self.assertIn('expected: value of type Person', out)
 
     @unittest.skip
     def test_mutableRecordFail3(self):
-        rec = 'file-tests/typeRecords.py'
+        rec = 'test-data/typeRecords.py'
         out = runInteractive(rec, 'p = MutablePerson("stefan", 42)\np.age = 42.4')
-        self.assertIn('expected: int', out)
+        self.assertIn('expected: value of type int', out)
 
     def test_union(self):
-        out = runInteractive('file-tests/typeUnion.py', """formatAnimal(myCat)
+        out = runInteractive('test-data/typeUnion.py', """formatAnimal(myCat)
 formatAnimal(myParrot)
 formatAnimal(None)
         """)
         self.assertEqual("'Cat Pumpernickel'", out[0])
         self.assertEqual("\"Parrot Mike says: Let's go to the punkrock show\"", out[1])
-        self.assertIn('given: None\nexpected: Union[Cat, Parrot]', out[2])
+        self.assertIn('given:    None\nexpected: value of type Union[Cat, Parrot]', out[2])
 
 class StudentSubmissionTests(unittest.TestCase):
     def check(self, file, testFile, ecode, tycheck=True):
@@ -100,72 +98,83 @@ def check(self, file, testFile, ecode, tycheck=True):
         if not tycheck:
             flags.append('--no-typechecking')
         cmd = f"python3 src/runYourProgram.py {' '.join(flags)} --test-file {testFile} {file} {LOG_REDIR}"
-        print(cmd)
         res = shell.run(cmd, onError='ignore')
         self.assertEqual(ecode, res.exitcode)
 
     def test_goodSubmission(self):
-        self.check("file-tests/student-submission.py", "file-tests/student-submission-tests.py", 0)
-        self.check("file-tests/student-submission.py", "file-tests/student-submission-tests.py", 0,
+        self.check("test-data/student-submission.py", "test-data/student-submission-tests.py", 0)
+        self.check("test-data/student-submission.py", "test-data/student-submission-tests.py", 0,
                    tycheck=False)
 
     def test_badSubmission(self):
-        self.check("file-tests/student-submission-bad.py",
-                   "file-tests/student-submission-tests.py", 1)
-        self.check("file-tests/student-submission-bad.py",
-                   "file-tests/student-submission-tests.py", 1, tycheck=False)
+        self.check("test-data/student-submission-bad.py",
+                   "test-data/student-submission-tests.py", 1)
+        self.check("test-data/student-submission-bad.py",
+                   "test-data/student-submission-tests.py", 1, tycheck=False)
 
     def test_submissionWithTypeErrors(self):
-        self.check("file-tests/student-submission-tyerror.py",
-                   "file-tests/student-submission-tests.py", 1)
-        self.check("file-tests/student-submission-tyerror.py",
-                   "file-tests/student-submission-tests.py", 0, tycheck=False)
-        self.check("file-tests/student-submission.py",
-                   "file-tests/student-submission-tests-tyerror.py", 1)
-        self.check("file-tests/student-submission.py",
-                   "file-tests/student-submission-tests-tyerror.py", 0, tycheck=False)
+        self.check("test-data/student-submission-tyerror.py",
+                   "test-data/student-submission-tests.py", 1)
+        self.check("test-data/student-submission-tyerror.py",
+                   "test-data/student-submission-tests.py", 0, tycheck=False)
+        self.check("test-data/student-submission.py",
+                   "test-data/student-submission-tests-tyerror.py", 1)
+        self.check("test-data/student-submission.py",
+                   "test-data/student-submission-tests-tyerror.py", 0, tycheck=False)
 
 class InteractiveTests(unittest.TestCase):
 
     def test_scopeBugPeter(self):
-        out = runInteractive('file-tests/scope-bug-peter.py', 'local_test()\nprint(spam)')
+        out = runInteractive('test-data/scope-bug-peter.py', 'local_test()\nprint(spam)')
         self.assertIn('IT WORKS', out)
 
     def test_types1(self):
-        out = runInteractive('file-tests/testTypesInteractive.py', 'inc(3)')
+        out = runInteractive('test-data/testTypesInteractive.py', 'inc(3)')
         self.assertEqual(['4'], out)
 
     def test_types2(self):
-        out = runInteractive('file-tests/testTypesInteractive.py', 'inc("3")')[0]
-        expected = """given: '3'
-expected: int
-          ^^^
+        out = runInteractive('test-data/testTypesInteractive.py', 'inc("3")')[0]
+        expected = """untypy.error.UntypyTypeError
+given:    '3'
+expected: value of type int
+
+context: inc(x: int) -> int
+                ^^^
+declared at: /Users/swehr/devel/write-your-python-program/python/test-data/testTypesInteractive.py:1
+  1 | def inc(x: int) -> int:
+  2 |     return x + 1
 
-inside of inc(x: int) -> int
-                 ^^^
-declared at:"""
-        self.assertIn(expected, stripTrailingWs(out))
+caused by: <console>:1"""
+        self.assertEqual(expected, out)
 
     def test_types3(self):
-        out = runInteractive('file-tests/testTypesInteractive.py',
+        out = runInteractive('test-data/testTypesInteractive.py',
                              'def f(x: int) -> int: return x\n\nf("x")')[1]
-        self.assertIn('expected: int', out)
+        self.assertIn('expected: value of type int', out)
 
     def test_types4(self):
-        out = runInteractive('file-tests/testTypesInteractive.py',
+        out = runInteractive('test-data/testTypesInteractive.py',
                              'def f(x: int) -> int: return x\n\nf(3)')
         self.assertEqual(['...', '3'], out)
 
     def test_types5(self):
-        out = runInteractive('file-tests/testTypesInteractive.py',
+        out = runInteractive('test-data/testTypesInteractive.py',
                              'def f(x: int) -> int: return x\n\nf("x")',
                               tycheck=False)
         self.assertEqual(['...', "'x'"], out)
 
     def test_typesInImportedModule1(self):
-        out = run('file-tests/testTypes3.py', ecode=1)
-        self.assertIn('expected: int', out)
+        out = run('test-data/testTypes3.py', ecode=1)
+        self.assertIn('expected: value of type int', out)
 
     def test_typesInImportedModule2(self):
-        out = run('file-tests/testTypes3.py', tycheck=False)
+        out = run('test-data/testTypes3.py', tycheck=False)
         self.assertEqual('END', out)
+
+class ReplTesterTests(unittest.TestCase):
+
+    def test_replTester(self):
+        d = shell.pwd()
+        cmd = f'python3 {d}/src/replTester.py {d}/test-data/repl-test-lib.py --repl {d}/test-data/repl-test-checks.py'
+        res = shell.run(cmd, captureStdout=True, onError='die', cwd='/tmp')
+        self.assertIn('All 1 tests succeded. Great!', res.stdout)
diff --git a/python/run b/python/run
new file mode 100755
index 00000000..20dcb450
--- /dev/null
+++ b/python/run
@@ -0,0 +1,5 @@
+#!/bin/bash
+
+SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
+
+PYTHONPATH="$SCRIPT_DIR"/site-lib/ python3 "$SCRIPT_DIR"/src/runYourProgram.py "$@"
diff --git a/python/runFileTests.sh b/python/runFileTests.sh
index 567097a9..e0368194 100755
--- a/python/runFileTests.sh
+++ b/python/runFileTests.sh
@@ -10,7 +10,6 @@ d=$(pwd)
 siteDir=$(python3 -c 'import site; print(site.USER_SITE)')
 t=$(mktemp)
 
-echo
 echo "Running file tests, siteDir=$siteDir ..."
 echo "Writing logs to $t"
 function check()
@@ -23,10 +22,10 @@ function check()
     python3 $d/src/runYourProgram.py --check --install-mode assertInstall $d/"$1" >> "$t"
     popd > /dev/null
 }
-check file-tests/fileWithImport.py
-check file-tests/fileWithoutImport.py
-check file-tests/fileWithBothImports.py
-check file-tests/fileWithRecursiveTypes.py
+check test-data/fileWithImport.py
+check test-data/fileWithoutImport.py
+check test-data/fileWithBothImports.py
+check test-data/fileWithRecursiveTypes.py
 
 # First argument: whether to do type checking or not
 # Second argument: expected exit code. If given as X:Y, then X is the exit code with active
@@ -37,6 +36,7 @@ function checkWithOutputAux()
     local tycheck="$1"
     local expectedEcode=$2
     local file="$3"
+    echo "Checking $file"
     shift 3
     tycheckOpt=""
     suffixes="${PYENV_VERSION}"
@@ -101,11 +101,34 @@ function checkWithOutput()
     checkWithOutputAux no "$@"
 }
 
-checkWithOutput 1 file-tests/testTraceback.py
-checkWithOutput 1 file-tests/testTraceback2.py
-checkWithOutput 1 file-tests/testTraceback3.py
-checkWithOutput 0 file-tests/testArgs.py ARG_1 ARG_2
-checkWithOutput 0 file-tests/printModuleName.py
-checkWithOutput 0 file-tests/printModuleNameImport.py
-checkWithOutput 1 file-tests/testTypes1.py
-checkWithOutput 1:0 file-tests/testTypes2.py
+checkWithOutput 1 test-data/testTraceback.py
+checkWithOutput 1 test-data/testTraceback2.py
+checkWithOutput 1 test-data/testTraceback3.py
+checkWithOutput 0 test-data/testArgs.py ARG_1 ARG_2
+checkWithOutput 0 test-data/printModuleName.py
+checkWithOutput 0 test-data/printModuleNameImport.py
+checkWithOutput 1 test-data/testTypes1.py
+checkWithOutput 1:0 test-data/testTypes2.py
+checkWithOutputAux yes 1 test-data/testTypesCollections1.py
+checkWithOutputAux yes 1 test-data/testTypesCollections2.py
+# checkWithOutputAux yes 1 test-data/testTypesCollections3.py  See #5
+# checkWithOutputAux yes 1 test-data/testTypesCollections4.py  See #6
+checkWithOutputAux yes 1 test-data/testTypesProtos1.py
+# checkWithOutputAux yes 1 test-data/testTypesProtos2.py  See #8
+checkWithOutputAux yes 1 test-data/testTypesProtos3.py
+# checkWithOutputAux yes 1 test-data/testTypesProtos4.py  See #9
+# checkWithOutputAux yes 1 test-data/testTypesSubclassing1.py  See #10
+# checkWithOutputAux yes 1 test-data/testTypesHigherOrderFuns.py See #7
+# checkWithOutputAux yes 1 test-data/testTypesRecordInheritance.py  See #11
+checkWithOutputAux yes 0 test-data/testForwardRef1.py
+# checkWithOutputAux yes 1 test-data/testForwardRef2.py See #14
+checkWithOutputAux yes 0 test-data/testForwardRef3.py
+# checkWithOutputAux yes 1 test-data/testForwardRef4.py  See #14
+# checkWithOutputAux yes 1 test-data/testTypesReturn.py  See #15
+checkWithOutputAux yes 1 test-data/testTypesSequence1.py
+checkWithOutputAux yes 1 test-data/testTypesSequence2.py
+checkWithOutputAux yes 1 test-data/testTypesTuple1.py
+# checkWithOutputAux yes 1 test-data/wrong-caused-by.py  See #17
+# checkWithOutputAux yes 1 test-data/declared-aty.py  See #18
+
+
diff --git a/python/runTestsForPyVersion.sh b/python/runTestsForPyVersion.sh
index be275b8c..0af1708c 100755
--- a/python/runTestsForPyVersion.sh
+++ b/python/runTestsForPyVersion.sh
@@ -1,11 +1,20 @@
 #!/bin/bash
 
 set -e
+set -u
 
 cd $(dirname $0)
 
 unit_test_path=src:tests:deps/untypy
-integ_test_path=integration-tests
+
+function prepare_integration_tests()
+{
+    echo "Preparing integration tests by install the WYPP library"
+    local d=$(mktemp -d)
+    trap "rm -rf $d" EXIT
+    WYPP_INSTALL_DIR=$d python3 src/runYourProgram.py --install-mode installOnly
+    integ_test_path=integration-tests:$d
+}
 
 function usage()
 {
@@ -13,11 +22,15 @@ function usage()
     exit 1
 }
 
-if [ -z "$1" ]; then
-    echo "Running all unit tests"
+if [ -z "${1:-}" ]; then
+    echo "Running all unit tests, PYTHONPATH=$unit_test_path"
     PYTHONPATH=$unit_test_path python3 -m unittest tests/test*.py
-    echo "Running all integration tests"
+    echo "Done with unit tests"
+    echo
+    prepare_integration_tests
+    echo "Running all integration tests, PYTHONPATH=$integ_test_path"
     PYTHONPATH=$integ_test_path python3 -m unittest integration-tests/test*.py
+    echo "Done with integration tests"
 else
     if [ "$1" == "--unit" ]; then
         what="unit"
@@ -26,13 +39,14 @@ else
     elif [ "$1" == "--integration" ]; then
         what="integration"
         dir=integration-tests
+        prepare_integration_tests
         p=$integ_test_path
     else
         usage
     fi
     shift
-    echo "Running $what tests $@"
-    if [ -z "$1" ]; then
+    echo "Running $what tests $@ with PYTHONPATH=$p"
+    if [ -z "${1:-}" ]; then
         PYTHONPATH=$p python3 -m unittest $dir/test*.py
         ecode=$?
     else
diff --git a/python/site-lib/README b/python/site-lib/README
new file mode 100644
index 00000000..d2769afc
--- /dev/null
+++ b/python/site-lib/README
@@ -0,0 +1,3 @@
+This directory contains symlinks for wypp and untypy.
+
+Add this directory to PYTHONPATH if you want to import wypp and untypy in-place.
diff --git a/python/site-lib/untypy b/python/site-lib/untypy
new file mode 120000
index 00000000..db337ec2
--- /dev/null
+++ b/python/site-lib/untypy
@@ -0,0 +1 @@
+../deps/untypy/untypy/
\ No newline at end of file
diff --git a/python/site-lib/wypp b/python/site-lib/wypp
new file mode 120000
index 00000000..e057607e
--- /dev/null
+++ b/python/site-lib/wypp
@@ -0,0 +1 @@
+../src/
\ No newline at end of file
diff --git a/python/src/__init__.py b/python/src/__init__.py
index c866c517..887dfe6d 100644
--- a/python/src/__init__.py
+++ b/python/src/__init__.py
@@ -3,7 +3,6 @@
 # Exported names that are available for star imports (in alphabetic order)
 Any = w.Any
 Callable = w.Callable
-ForwardRef = w.ForwardRef
 Generator = w.Generator
 Iterable = w.Iterable
 Iterator = w.Iterator
@@ -17,12 +16,14 @@
 math = w.math
 nat = w.nat
 record = w.record
+T = w.T
+U = w.U
 unchecked = w.unchecked
+V = w.V
 
 __all__ = [
     'Any',
     'Callable',
-    'ForwardRef',
     'Generator',
     'Iterable',
     'Iterator',
@@ -36,7 +37,10 @@
     'math',
     'nat',
     'record',
-    'unchecked'
+    'T',
+    'U',
+    'unchecked',
+    'V'
 ]
 
 # Exported names not available for star imports (in alphabetic order)
diff --git a/python/src/replTester.py b/python/src/replTester.py
new file mode 100644
index 00000000..5f7e4ca7
--- /dev/null
+++ b/python/src/replTester.py
@@ -0,0 +1,101 @@
+import sys
+import doctest
+import os
+import argparse
+from dataclasses import dataclass
+from runner import runCode, importUntypy, verbose, enableVerbose
+
+usage = """python3 replTester.py [ ARGUMENTS ] LIB_1 ... LIB_n --repl SAMPLE_1 ... SAMPLE_m
+
+If no library files should be used to test the REPL samples, omit LIB_1 ... LIB_n
+and the --repl flag.
+The definitions of LIB_1 ... LIB_n are made available when testing
+SAMPLE_1 ... SAMPLE_m, where identifer in LIB_i takes precedence over identifier in
+LIB_j if i > j.
+"""
+
+@dataclass
+class Options:
+    verbose: bool
+    diffOutput: bool
+    libs: list[str]
+    repls: list[str]
+
+def parseCmdlineArgs():
+    parser = argparse.ArgumentParser(usage=usage,
+                        formatter_class=argparse.RawTextHelpFormatter)
+    parser.add_argument('--verbose', dest='verbose', action='store_const',
+                        const=True, default=False,
+                        help='Be verbose')
+    parser.add_argument('--diffOutput', dest='diffOutput',
+                        action='store_const', const=True, default=False,
+                        help='print diff of expected/given output')
+    args, restArgs = parser.parse_known_args()
+    libs = []
+    repls = []
+    replFlag = '--repl'
+    if replFlag in restArgs:
+        cur = libs
+        for x in restArgs:
+            if x == replFlag:
+                cur = repls
+            else:
+                cur.append(x)
+    else:
+        repls = restArgs
+    if len(repls) == 0:
+        print('No SAMPLE arguments given')
+        sys.exit(1)
+    return Options(verbose=args.verbose, diffOutput=args.diffOutput, libs=libs, repls=repls)
+
+opts = parseCmdlineArgs()
+
+if opts.verbose:
+    enableVerbose()
+
+libDir = os.path.dirname(__file__)
+libFile = os.path.join(libDir, 'writeYourProgram.py')
+defs = globals()
+importUntypy()
+# runCode(libFile, defs, [])
+
+for lib in opts.libs:
+    d = os.path.dirname(lib)
+    if d not in sys.path:
+        sys.path.insert(0, d)
+
+for lib in opts.libs:
+    verbose(f"Loading lib {lib}")
+    runCode(lib, defs, [])
+
+totalFailures = 0
+totalTests = 0
+
+doctestOptions = doctest.NORMALIZE_WHITESPACE | doctest.ELLIPSIS
+
+if opts.diffOutput:
+    doctestOptions = doctestOptions | doctest.REPORT_NDIFF
+
+for repl in opts.repls:
+    (failures, tests) = doctest.testfile(repl, globs=defs, module_relative=False,
+                                         optionflags=doctestOptions)
+
+    totalFailures += failures
+    totalTests += tests
+    if failures == 0:
+        if tests == 0:
+            print(f'No tests in {repl}')
+        else:
+            print(f'All {tests} tests in {repl} succeeded')
+    else:
+        print(f'ERROR: {failures} out of {tests} in {repl} failed')
+
+if totalFailures == 0:
+    if totalTests == 0:
+        print('ERROR: No tests found at all!')
+        sys.exit(1)
+    else:
+        print(f'All {totalTests} tests succeded. Great!')
+else:
+    print(f'ERROR: {failures} out of {tests} failed')
+    sys.exit(1)
diff --git a/python/src/runner.py b/python/src/runner.py
index 081af5ee..86f04310 100644
--- a/python/src/runner.py
+++ b/python/src/runner.py
@@ -24,6 +24,10 @@ def die(ecode=1):
 
 VERBOSE = False # set via commandline
 
+def enableVerbose():
+    global VERBOSE
+    VERBOSE = True
+
 LIB_DIR = os.path.dirname(__file__)
 INSTALLED_MODULE_NAME = 'wypp'
 FILES_TO_INSTALL = ['writeYourProgram.py', 'drawingLib.py', '__init__.py']
@@ -125,13 +129,12 @@ def isSameFile(f1, f2):
     y = readFile(f2)
     return x == y
 
-def installFromDir(srcDir, mod, files=None):
+def installFromDir(srcDir, targetDir, mod, files=None):
     if files is None:
         files = [p.relative_to(srcDir) for p in Path(srcDir).rglob('*.py')]
     else:
         files = [Path(f) for f in files]
-    userDir = site.USER_SITE
-    installDir = os.path.join(userDir, mod)
+    installDir = os.path.join(targetDir, mod)
     os.makedirs(installDir, exist_ok=True)
     installedFiles = sorted([p.relative_to(installDir) for p in Path(installDir).rglob('*.py')])
     wantedFiles = sorted(files)
@@ -144,7 +147,7 @@ def installFromDir(srcDir, mod, files=None):
                 break
         else:
             # no break, all files equal
-            verbose(f'All files from {srcDir} already installed in {userDir}/{mod}')
+            verbose(f'All files from {srcDir} already installed in {targetDir}/{mod}')
             return True
     else:
         verbose(f'Installed files {installedFiles} and wanted files {wantedFiles} are different')
@@ -164,17 +167,17 @@ def installLib(mode):
     if mode == InstallMode.dontInstall:
         verbose("No installation of WYPP should be performed")
         return
-    userDir = site.USER_SITE
+    targetDir = os.getenv('WYPP_INSTALL_DIR', site.USER_SITE)
     try:
-        allEq1 = installFromDir(LIB_DIR, INSTALLED_MODULE_NAME, FILES_TO_INSTALL)
-        allEq2 = installFromDir(UNTYPY_DIR, UNTYPY_MODULE_NAME)
+        allEq1 = installFromDir(LIB_DIR, targetDir, INSTALLED_MODULE_NAME, FILES_TO_INSTALL)
+        allEq2 = installFromDir(UNTYPY_DIR, targetDir, UNTYPY_MODULE_NAME)
         if allEq1 and allEq2:
-            verbose(f'WYPP library in {userDir} already up to date')
+            verbose(f'WYPP library in {targetDir} already up to date')
             if mode == InstallMode.installOnly:
-                printStderr(f'WYPP library in {userDir} already up to date')
+                printStderr(f'WYPP library in {targetDir} already up to date')
             return
         else:
-            printStderr(f'The WYPP library has been successfully installed in {userDir}.')
+            printStderr(f'The WYPP library has been successfully installed in {targetDir}.')
     except Exception as e:
         printStderr('Installation of the WYPP library failed: ' + str(e))
         if mode == InstallMode.assertInstall or mode == InstallMode.installOnly:
@@ -201,7 +204,7 @@ def __init__(self, mod, properlyImported):
                 if name and name[0] != '_':
                   d[name] = getattr(mod, name)
 
-def loadLib(onlyCheckRunnable):
+def prepareLib(onlyCheckRunnable):
     libDefs = None
     mod = INSTALLED_MODULE_NAME
     verbose('Attempting to import ' + mod)
@@ -259,7 +262,7 @@ def __exit__(self, exc_type, value, traceback):
             sys.path.remove(self.dir)
             self.inserted = False
 
-def runCode(fileToRun, globals, args, *, useUntypy=True):
+def runCode(fileToRun, globals, args, useUntypy=True):
     localDir = os.path.dirname(fileToRun)
     with sysPathPrepended(localDir):
         with open(fileToRun) as f:
@@ -286,11 +289,7 @@ def runCode(fileToRun, globals, args, *, useUntypy=True):
             finally:
                 sys.argv = oldArgs
 
-def runStudentCode(fileToRun, globals, libDefs, onlyCheckRunnable, args, *, useUntypy=True):
-    importsWypp = findWyppImport(fileToRun)
-    if importsWypp:
-        if not libDefs.properlyImported:
-            globals[INSTALLED_MODULE_NAME] = libDefs.dict
+def runStudentCode(fileToRun, globals, onlyCheckRunnable, args, useUntypy=True):
     doRun = lambda: runCode(fileToRun, globals, args, useUntypy=useUntypy)
     if onlyCheckRunnable:
         try:
@@ -361,7 +360,11 @@ def limitTraceback(fullTb):
 def handleCurrentException(exit=True, removeFirstTb=False, file=sys.stderr):
     (etype, val, tb) = sys.exc_info()
     if isinstance(val, untypy.error.UntypyTypeError) or isinstance(val, untypy.error.UntypyAttributeError):
-        file.write(str(val))
+        file.write(etype.__module__ + "." + etype.__qualname__)
+        s = str(val)
+        if s and s[0] != '\n':
+            file.write(': ')
+        file.write(s)
         file.write('\n')
     else:
         if tb and removeFirstTb:
@@ -384,6 +387,11 @@ def getHistoryFilePath():
     else:
         return None
 
+# We cannot import untypy at the top of the file because we might have to install it first.
+def importUntypy():
+    global untypy
+    import untypy
+
 def main(globals):
     v = sys.version_info
     if v.major < 3 or v.minor < 9:
@@ -398,9 +406,7 @@ def main(globals):
         VERBOSE = True
 
     installLib(args.installMode)
-
-    global untypy
-    import untypy
+    importUntypy()
 
     fileToRun = args.file
     if args.changeDir:
@@ -415,13 +421,13 @@ def main(globals):
     if not args.checkRunnable and not args.quiet:
         printWelcomeString(fileToRun, version, useUntypy=args.checkTypes)
 
-    libDefs = loadLib(onlyCheckRunnable=args.checkRunnable)
+    libDefs = prepareLib(onlyCheckRunnable=args.checkRunnable)
 
     globals['__name__'] = '__wypp__'
     sys.modules['__wypp__'] = sys.modules['__main__']
     try:
         verbose(f'running code in {fileToRun}')
-        runStudentCode(fileToRun, globals, libDefs, args.checkRunnable, restArgs,
+        runStudentCode(fileToRun, globals, args.checkRunnable, restArgs,
                        useUntypy=args.checkTypes)
     except:
         handleCurrentException()
diff --git a/python/src/testReplSamples.py b/python/src/testReplSamples.py
deleted file mode 100644
index bbc7d557..00000000
--- a/python/src/testReplSamples.py
+++ /dev/null
@@ -1,74 +0,0 @@
-import sys
-import doctest
-import os
-from runner import runCode
-
-def usage():
-    print('USAGE: python3 testReplSamples.py LIB_1 ... LIB_n --repl SAMPLE_1 ... SAMPLE_m')
-    print('If no library files should be used to test the REPL samples, omit LIB_1 ... LIB_n')
-    print('and the --repl flag.')
-    print('The definitions of LIB_1 ... LIB_n are made available when testing ')
-    print('SAMPLE_1 ... SAMPLE_m, where identifer in LIB_i takes precedence over identifier in ')
-    print('LIB_j if i > j.')
-    sys.exit(1)
-
-args = sys.argv[1:]
-
-if '--help' in args:
-    usage()
-
-libs = []
-repls = []
-
-replFlag = '--repl'
-
-if replFlag in args:
-    cur = libs
-    for x in args:
-        if x == replFlag:
-            cur = repls
-        else:
-            cur.append(x)
-else:
-    repls = args
-
-if len(repls) == 0:
-    usage()
-
-libDir = os.path.dirname(__file__)
-libFile = os.path.join(libDir, 'writeYourProgram.py')
-defs = {}
-runCode(libFile, defs, [])
-
-for lib in libs:
-    d = os.path.dirname(lib)
-    if d not in sys.path:
-        sys.path.insert(0, d)
-
-for lib in libs:
-    runCode(lib, defs, [])
-
-totalFailures = 0
-totalTests = 0
-
-for repl in repls:
-    (failures, tests) = doctest.testfile(repl, globs=defs, module_relative=False)
-    totalFailures += failures
-    totalTests += tests
-    if failures == 0:
-        if tests == 0:
-            print(f'No tests in {repl}')
-        else:
-            print(f'All {tests} tests in {repl} succeeded')
-    else:
-        print(f'ERROR: {failures} out of {tests} in {repl} failed')
-
-if totalFailures == 0:
-    if totalTests == 0:
-        print('ERROR: No tests found at all!')
-        sys.exit(1)
-    else:
-        print(f'All {totalTests} tests succeded. Great!')
-else:
-    print(f'ERROR: {failures} out of {tests} failed')
-    sys.exit(1)
diff --git a/python/src/writeYourProgram.py b/python/src/writeYourProgram.py
index 1f279943..c0d66e65 100644
--- a/python/src/writeYourProgram.py
+++ b/python/src/writeYourProgram.py
@@ -26,14 +26,17 @@ def _debug(s):
 
 dataclass = dataclasses.dataclass
 
-# Reexports for Untypy
 unchecked = untypy.unchecked
 nat = typing.Annotated[int, lambda i: i >= 0]
 
+T = typing.TypeVar('T')
+U = typing.TypeVar('U')
+V = typing.TypeVar('V')
+
 def _patchDataClass(cls, mutable):
     fieldNames = [f.name for f in dataclasses.fields(cls)]
     setattr(cls, EQ_ATTRS_ATTR, fieldNames)
-    
+
     if hasattr(cls, '__annotations__'):
         # add annotions for type checked constructor.
         cls.__init__.__annotations__ =  cls.__annotations__
diff --git a/python/test-data/declared-at-missing.py b/python/test-data/declared-at-missing.py
new file mode 100644
index 00000000..979ff573
--- /dev/null
+++ b/python/test-data/declared-at-missing.py
@@ -0,0 +1,22 @@
+from wypp import *
+
+@record
+class Course:
+    name: str
+    teacher: str
+    students: tuple[str, ...]
+
+@record(mutable=True)
+class CourseM:
+    name: str
+    teacher: str
+    students: tuple[str, ...]
+
+@record
+class Semester:
+    degreeProgram: str
+    semester: str
+    courses: tuple[CourseM, ...]
+
+prog1 = Course('Programmierung 1', 'Wehr', ())
+semester1_2020 = Semester('AKI', '1. Semester 2020/21', (prog1, ))
diff --git a/python/file-tests/fileWithBothImports.py b/python/test-data/fileWithBothImports.py
similarity index 90%
rename from python/file-tests/fileWithBothImports.py
rename to python/test-data/fileWithBothImports.py
index 29df392d..489c85e4 100644
--- a/python/file-tests/fileWithBothImports.py
+++ b/python/test-data/fileWithBothImports.py
@@ -11,4 +11,3 @@ def use(x):
 use(wypp.record)
 use(drawingLib.Point)
 use(Point)
-use(wypp.list[int])
diff --git a/python/file-tests/fileWithImport.py b/python/test-data/fileWithImport.py
similarity index 79%
rename from python/file-tests/fileWithImport.py
rename to python/test-data/fileWithImport.py
index e8b11c9b..861fdb88 100644
--- a/python/file-tests/fileWithImport.py
+++ b/python/test-data/fileWithImport.py
@@ -6,4 +6,3 @@ def use(x):
 
 use(wypp)
 use(wypp.record)
-use(wypp.list[int])
diff --git a/python/file-tests/fileWithRecursiveTypes.py b/python/test-data/fileWithRecursiveTypes.py
similarity index 100%
rename from python/file-tests/fileWithRecursiveTypes.py
rename to python/test-data/fileWithRecursiveTypes.py
diff --git a/python/file-tests/fileWithStarImport.py b/python/test-data/fileWithStarImport.py
similarity index 100%
rename from python/file-tests/fileWithStarImport.py
rename to python/test-data/fileWithStarImport.py
diff --git a/python/file-tests/fileWithoutImport.py b/python/test-data/fileWithoutImport.py
similarity index 100%
rename from python/file-tests/fileWithoutImport.py
rename to python/test-data/fileWithoutImport.py
diff --git a/python/file-tests/localMod.py b/python/test-data/localMod.py
similarity index 100%
rename from python/file-tests/localMod.py
rename to python/test-data/localMod.py
diff --git a/python/file-tests/printModuleName.err b/python/test-data/printModuleName.err
similarity index 100%
rename from python/file-tests/printModuleName.err
rename to python/test-data/printModuleName.err
diff --git a/python/file-tests/printModuleName.out b/python/test-data/printModuleName.out
similarity index 100%
rename from python/file-tests/printModuleName.out
rename to python/test-data/printModuleName.out
diff --git a/python/file-tests/printModuleName.py b/python/test-data/printModuleName.py
similarity index 100%
rename from python/file-tests/printModuleName.py
rename to python/test-data/printModuleName.py
diff --git a/python/file-tests/printModuleNameImport.err b/python/test-data/printModuleNameImport.err
similarity index 100%
rename from python/file-tests/printModuleNameImport.err
rename to python/test-data/printModuleNameImport.err
diff --git a/python/file-tests/printModuleNameImport.out b/python/test-data/printModuleNameImport.out
similarity index 100%
rename from python/file-tests/printModuleNameImport.out
rename to python/test-data/printModuleNameImport.out
diff --git a/python/file-tests/printModuleNameImport.py b/python/test-data/printModuleNameImport.py
similarity index 100%
rename from python/file-tests/printModuleNameImport.py
rename to python/test-data/printModuleNameImport.py
diff --git a/python/test-data/repl-test-checks.py b/python/test-data/repl-test-checks.py
new file mode 100644
index 00000000..8f2560c2
--- /dev/null
+++ b/python/test-data/repl-test-checks.py
@@ -0,0 +1,2 @@
+>>> dora.gewicht
+25000
diff --git a/python/test-data/repl-test-lib.py b/python/test-data/repl-test-lib.py
new file mode 100644
index 00000000..4b836670
--- /dev/null
+++ b/python/test-data/repl-test-lib.py
@@ -0,0 +1,10 @@
+from wypp import *
+
+Status = Literal['tot', 'lebendig']
+
+@record
+class Gürteltier:
+    gewicht: int
+    totOderLebendig: Status
+
+dora = Gürteltier(25000, 'lebendig')
diff --git a/python/file-tests/scope-bug-peter.py b/python/test-data/scope-bug-peter.py
similarity index 100%
rename from python/file-tests/scope-bug-peter.py
rename to python/test-data/scope-bug-peter.py
diff --git a/python/file-tests/student-submission-bad.py b/python/test-data/student-submission-bad.py
similarity index 100%
rename from python/file-tests/student-submission-bad.py
rename to python/test-data/student-submission-bad.py
diff --git a/python/file-tests/student-submission-tests-tyerror.py b/python/test-data/student-submission-tests-tyerror.py
similarity index 100%
rename from python/file-tests/student-submission-tests-tyerror.py
rename to python/test-data/student-submission-tests-tyerror.py
diff --git a/python/file-tests/student-submission-tests.py b/python/test-data/student-submission-tests.py
similarity index 100%
rename from python/file-tests/student-submission-tests.py
rename to python/test-data/student-submission-tests.py
diff --git a/python/file-tests/student-submission-tyerror.py b/python/test-data/student-submission-tyerror.py
similarity index 100%
rename from python/file-tests/student-submission-tyerror.py
rename to python/test-data/student-submission-tyerror.py
diff --git a/python/file-tests/student-submission.py b/python/test-data/student-submission.py
similarity index 100%
rename from python/file-tests/student-submission.py
rename to python/test-data/student-submission.py
diff --git a/python/file-tests/testArgs.err b/python/test-data/testArgs.err
similarity index 100%
rename from python/file-tests/testArgs.err
rename to python/test-data/testArgs.err
diff --git a/python/test-data/testArgs.out b/python/test-data/testArgs.out
new file mode 100644
index 00000000..982a2823
--- /dev/null
+++ b/python/test-data/testArgs.out
@@ -0,0 +1 @@
+['test-data/testArgs.py', 'ARG_1', 'ARG_2']
diff --git a/python/file-tests/testArgs.py b/python/test-data/testArgs.py
similarity index 100%
rename from python/file-tests/testArgs.py
rename to python/test-data/testArgs.py
diff --git a/python/file-tests/testTraceback2.out b/python/test-data/testForwardRef1.err
similarity index 100%
rename from python/file-tests/testTraceback2.out
rename to python/test-data/testForwardRef1.err
diff --git a/python/test-data/testForwardRef1.out b/python/test-data/testForwardRef1.out
new file mode 100644
index 00000000..345e6aef
--- /dev/null
+++ b/python/test-data/testForwardRef1.out
@@ -0,0 +1 @@
+Test
diff --git a/python/test-data/testForwardRef1.py b/python/test-data/testForwardRef1.py
new file mode 100644
index 00000000..9aed46b3
--- /dev/null
+++ b/python/test-data/testForwardRef1.py
@@ -0,0 +1,11 @@
+class Test:
+    def __init__(self, foo: 'Foo'):
+        pass
+    def __repr__(self):
+        return 'Test'
+
+class Foo:
+    pass
+
+t = Test(Foo())
+print(t)
diff --git a/python/test-data/testForwardRef2.py b/python/test-data/testForwardRef2.py
new file mode 100644
index 00000000..40523896
--- /dev/null
+++ b/python/test-data/testForwardRef2.py
@@ -0,0 +1,11 @@
+class Test:
+    def __init__(self, foo: 'Foo'):
+        pass
+    def __repr__(self):
+        return 'Test'
+
+class FooX:
+    pass
+
+t = Test(FooX())
+print(t)
diff --git a/python/file-tests/testTraceback3.out b/python/test-data/testForwardRef3.err
similarity index 100%
rename from python/file-tests/testTraceback3.out
rename to python/test-data/testForwardRef3.err
diff --git a/python/test-data/testForwardRef3.out b/python/test-data/testForwardRef3.out
new file mode 100644
index 00000000..82cf2102
--- /dev/null
+++ b/python/test-data/testForwardRef3.out
@@ -0,0 +1 @@
+Test(foo=Foo)
diff --git a/python/test-data/testForwardRef3.py b/python/test-data/testForwardRef3.py
new file mode 100644
index 00000000..d0e3759f
--- /dev/null
+++ b/python/test-data/testForwardRef3.py
@@ -0,0 +1,12 @@
+from wypp import *
+
+@record
+class Test:
+    foo: 'Foo'
+
+class Foo:
+    def __repr__(self):
+        return 'Foo'
+
+t = Test(Foo())
+print(t)
diff --git a/python/test-data/testForwardRef4.py b/python/test-data/testForwardRef4.py
new file mode 100644
index 00000000..2cb7d5ff
--- /dev/null
+++ b/python/test-data/testForwardRef4.py
@@ -0,0 +1,12 @@
+from wypp import *
+
+@record
+class Test:
+    foo: 'FooX'
+
+class Foo:
+    def __repr__(self):
+        return 'Foo'
+
+t = Test(Foo())
+print(t)
diff --git a/python/test-data/testTraceback.err b/python/test-data/testTraceback.err
new file mode 100644
index 00000000..d12f2238
--- /dev/null
+++ b/python/test-data/testTraceback.err
@@ -0,0 +1,7 @@
+
+Traceback (most recent call last):
+  File "test-data/testTraceback.py", line 9, in <module>
+    foo(lst)
+  File "test-data/testTraceback.py", line 7, in foo
+    print(lst[10])
+IndexError: list index out of range
diff --git a/python/file-tests/testTraceback.out b/python/test-data/testTraceback.out
similarity index 100%
rename from python/file-tests/testTraceback.out
rename to python/test-data/testTraceback.out
diff --git a/python/file-tests/testTraceback.py b/python/test-data/testTraceback.py
similarity index 100%
rename from python/file-tests/testTraceback.py
rename to python/test-data/testTraceback.py
diff --git a/python/file-tests/testTraceback2.err-3.9.0 b/python/test-data/testTraceback2.err
similarity index 62%
rename from python/file-tests/testTraceback2.err-3.9.0
rename to python/test-data/testTraceback2.err
index e3f643b9..281b1771 100644
--- a/python/file-tests/testTraceback2.err-3.9.0
+++ b/python/test-data/testTraceback2.err
@@ -1,5 +1,5 @@
 
-  File "file-tests/testTraceback2.py", line 3
+  File "test-data/testTraceback2.py", line 3
     lst = [1,2,3
                 ^
 SyntaxError: unexpected EOF while parsing
diff --git a/python/file-tests/testTypes1.out b/python/test-data/testTraceback2.out
similarity index 100%
rename from python/file-tests/testTypes1.out
rename to python/test-data/testTraceback2.out
diff --git a/python/file-tests/testTraceback2.py b/python/test-data/testTraceback2.py
similarity index 100%
rename from python/file-tests/testTraceback2.py
rename to python/test-data/testTraceback2.py
diff --git a/python/file-tests/testTraceback3.err b/python/test-data/testTraceback3.err
similarity index 61%
rename from python/file-tests/testTraceback3.err
rename to python/test-data/testTraceback3.err
index e2de6b58..c59940d1 100644
--- a/python/file-tests/testTraceback3.err
+++ b/python/test-data/testTraceback3.err
@@ -1,5 +1,5 @@
 
 Traceback (most recent call last):
-  File "file-tests/testTraceback3.py", line 2, in <module>
+  File "test-data/testTraceback3.py", line 2, in <module>
     print([1,2,3][10])
 IndexError: list index out of range
diff --git a/python/file-tests/testTypes2.err-notypes b/python/test-data/testTraceback3.out
similarity index 100%
rename from python/file-tests/testTypes2.err-notypes
rename to python/test-data/testTraceback3.out
diff --git a/python/file-tests/testTraceback3.py b/python/test-data/testTraceback3.py
similarity index 100%
rename from python/file-tests/testTraceback3.py
rename to python/test-data/testTraceback3.py
diff --git a/python/test-data/testTypes1.err b/python/test-data/testTypes1.err
new file mode 100644
index 00000000..d52cba13
--- /dev/null
+++ b/python/test-data/testTypes1.err
@@ -0,0 +1,12 @@
+untypy.error.UntypyTypeError
+given:    '1'
+expected: value of type int
+
+context: inc(x: int) -> int
+                ^^^
+declared at: test-data/testTypes1.py:1
+  1 | def inc(x: int) -> int:
+  2 |     return x + 1
+
+caused by: test-data/testTypes1.py:4
+  4 | inc("1")
diff --git a/python/file-tests/testTypes1.err-notypes b/python/test-data/testTypes1.err-notypes
similarity index 53%
rename from python/file-tests/testTypes1.err-notypes
rename to python/test-data/testTypes1.err-notypes
index ecbea461..c1014f7a 100644
--- a/python/file-tests/testTypes1.err-notypes
+++ b/python/test-data/testTypes1.err-notypes
@@ -1,7 +1,7 @@
 
 Traceback (most recent call last):
-  File "file-tests/testTypes1.py", line 4, in <module>
+  File "test-data/testTypes1.py", line 4, in <module>
     inc("1")
-  File "file-tests/testTypes1.py", line 2, in inc
+  File "test-data/testTypes1.py", line 2, in inc
     return x + 1
 TypeError: can only concatenate str (not "int") to str
diff --git a/python/file-tests/testTypes2.out b/python/test-data/testTypes1.out
similarity index 100%
rename from python/file-tests/testTypes2.out
rename to python/test-data/testTypes1.out
diff --git a/python/file-tests/testTypes1.py b/python/test-data/testTypes1.py
similarity index 100%
rename from python/file-tests/testTypes1.py
rename to python/test-data/testTypes1.py
diff --git a/python/test-data/testTypes2.err b/python/test-data/testTypes2.err
new file mode 100644
index 00000000..6a5fed85
--- /dev/null
+++ b/python/test-data/testTypes2.err
@@ -0,0 +1,12 @@
+untypy.error.UntypyTypeError
+given:    '1'
+expected: value of type int
+
+context: inc(x: int) -> int
+                ^^^
+declared at: test-data/testTypes2.py:1
+  1 | def inc(x: int) -> int:
+  2 |     return x
+
+caused by: test-data/testTypes2.py:4
+  4 | inc("1")
diff --git a/python/test-data/testTypes2.err-notypes b/python/test-data/testTypes2.err-notypes
new file mode 100644
index 00000000..e69de29b
diff --git a/python/test-data/testTypes2.out b/python/test-data/testTypes2.out
new file mode 100644
index 00000000..e69de29b
diff --git a/python/file-tests/testTypes2.py b/python/test-data/testTypes2.py
similarity index 100%
rename from python/file-tests/testTypes2.py
rename to python/test-data/testTypes2.py
diff --git a/python/file-tests/testTypes3.py b/python/test-data/testTypes3.py
similarity index 100%
rename from python/file-tests/testTypes3.py
rename to python/test-data/testTypes3.py
diff --git a/python/test-data/testTypesCollections1.err b/python/test-data/testTypesCollections1.err
new file mode 100644
index 00000000..2606513a
--- /dev/null
+++ b/python/test-data/testTypesCollections1.err
@@ -0,0 +1,12 @@
+untypy.error.UntypyTypeError
+given:    'foo'
+expected: value of type int
+
+context: list[int]
+              ^^^
+declared at: test-data/testTypesCollections1.py:3
+  3 | def appendSomething(l: list[int]) -> None:
+  4 |     l.append("foo")
+
+caused by: test-data/testTypesCollections1.py:4
+  4 |     l.append("foo")
diff --git a/python/test-data/testTypesCollections1.out b/python/test-data/testTypesCollections1.out
new file mode 100644
index 00000000..e69de29b
diff --git a/python/test-data/testTypesCollections1.py b/python/test-data/testTypesCollections1.py
new file mode 100644
index 00000000..be569e10
--- /dev/null
+++ b/python/test-data/testTypesCollections1.py
@@ -0,0 +1,8 @@
+from wypp import *
+
+def appendSomething(l: list[int]) -> None:
+    l.append("foo")
+
+l = [1,2,3]
+appendSomething(l)
+print(l)
diff --git a/python/test-data/testTypesCollections2.err b/python/test-data/testTypesCollections2.err
new file mode 100644
index 00000000..1f7b2b48
--- /dev/null
+++ b/python/test-data/testTypesCollections2.err
@@ -0,0 +1,15 @@
+untypy.error.UntypyTypeError
+given:    42
+expected: value of type str
+
+context: foo(l: list[Callable[[], str]]) -> list[str]
+                                  ^^^
+declared at: test-data/testTypesCollections2.py:3
+  3 | def foo(l: list[Callable[[], str]]) -> list[str]:
+  4 |     res = []
+  5 |     for f in l:
+  6 |         res.append(f())
+  7 |     return res
+
+caused by: test-data/testTypesCollections2.py:10
+ 10 | foo([lambda: "1", lambda: 42]) # error because the 2nd functions returns an int
diff --git a/python/test-data/testTypesCollections2.out b/python/test-data/testTypesCollections2.out
new file mode 100644
index 00000000..189744a2
--- /dev/null
+++ b/python/test-data/testTypesCollections2.out
@@ -0,0 +1 @@
+['1', '2']
diff --git a/python/test-data/testTypesCollections2.py b/python/test-data/testTypesCollections2.py
new file mode 100644
index 00000000..4e9c4997
--- /dev/null
+++ b/python/test-data/testTypesCollections2.py
@@ -0,0 +1,10 @@
+from wypp import *
+
+def foo(l: list[Callable[[], str]]) -> list[str]:
+    res = []
+    for f in l:
+        res.append(f())
+    return res
+
+print(foo([lambda: "1", lambda: "2"]))
+foo([lambda: "1", lambda: 42]) # error because the 2nd functions returns an int
diff --git a/python/test-data/testTypesCollections3.py b/python/test-data/testTypesCollections3.py
new file mode 100644
index 00000000..0bd390ec
--- /dev/null
+++ b/python/test-data/testTypesCollections3.py
@@ -0,0 +1,8 @@
+from wypp import *
+
+def appendSomething(l: list[int], l2: list[int]) -> None:
+    l.append("foo")
+
+l = [1,2,3]
+appendSomething(l, [])
+print(l)
diff --git a/python/test-data/testTypesCollections4.py b/python/test-data/testTypesCollections4.py
new file mode 100644
index 00000000..8ca810bc
--- /dev/null
+++ b/python/test-data/testTypesCollections4.py
@@ -0,0 +1,10 @@
+from wypp import *
+
+def foo(l: list[Callable[[], str]]) -> list[str]:
+    l.append(lambda: 42) # error
+    res = []
+    for f in l:
+        res.append(f())
+    return res
+
+foo([])
diff --git a/python/test-data/testTypesHigherOrderFuns.py b/python/test-data/testTypesHigherOrderFuns.py
new file mode 100644
index 00000000..68a4abe5
--- /dev/null
+++ b/python/test-data/testTypesHigherOrderFuns.py
@@ -0,0 +1,10 @@
+from wypp import *
+
+def map(container: Iterable[str], fun: Callable[[str], int]) -> list[int]:
+    res = []
+    for x in container:
+        res.append(fun(x))
+    return res
+
+print(map(["hello", "1"], len))
+map(["hello", "1"], lambda x: x)
diff --git a/python/file-tests/testTypesInteractive.py b/python/test-data/testTypesInteractive.py
similarity index 100%
rename from python/file-tests/testTypesInteractive.py
rename to python/test-data/testTypesInteractive.py
diff --git a/python/test-data/testTypesProtos1.err b/python/test-data/testTypesProtos1.err
new file mode 100644
index 00000000..e38d02ae
--- /dev/null
+++ b/python/test-data/testTypesProtos1.err
@@ -0,0 +1,34 @@
+untypy.error.UntypyTypeError
+Type 'Dog' does not implement protocol 'Animal' correctly.
+
+given:    <Dog object>
+expected: value of type Animal
+
+context: doSomething(a: Animal) -> None
+                        ^^^^^^
+declared at: test-data/testTypesProtos1.py:18
+ 18 | def doSomething(a: Animal) -> None:
+ 19 |     print(a.makeSound(3.14))
+
+caused by: test-data/testTypesProtos1.py:21
+ 21 | doSomething(Dog())
+
+The argument 'loadness' of method 'makeSound' violates the protocol 'Animal'.
+The annotation 'int' is incompatible with the protocol's annotation 'float'
+when checking against the following value:
+
+given:    3.14
+expected: value of type int
+
+context: makeSound(self: Self, loadness: int) -> str
+                                         ^^^
+declared at: test-data/testTypesProtos1.py:13
+ 13 |     def makeSound(self, loadness: int) -> str:
+ 14 |         return f"{loadness} wuffs"
+test-data/testTypesProtos1.py:8
+  8 |     def makeSound(self, loadness: float) -> str:
+  9 |         pass
+
+caused by: test-data/testTypesProtos1.py:13
+ 13 |     def makeSound(self, loadness: int) -> str:
+ 14 |         return f"{loadness} wuffs"
diff --git a/python/test-data/testTypesProtos1.out b/python/test-data/testTypesProtos1.out
new file mode 100644
index 00000000..e69de29b
diff --git a/python/test-data/testTypesProtos1.py b/python/test-data/testTypesProtos1.py
new file mode 100644
index 00000000..167878bb
--- /dev/null
+++ b/python/test-data/testTypesProtos1.py
@@ -0,0 +1,21 @@
+from wypp import *
+
+from typing import Protocol
+import abc
+
+class Animal(Protocol):
+    @abc.abstractmethod
+    def makeSound(self, loadness: float) -> str:
+        pass
+
+class Dog:
+    # incorrect implementation of the Animal protocol
+    def makeSound(self, loadness: int) -> str:
+        return f"{loadness} wuffs"
+    def __repr__(self):
+        return "<Dog object>"
+
+def doSomething(a: Animal) -> None:
+    print(a.makeSound(3.14))
+
+doSomething(Dog())
diff --git a/python/test-data/testTypesProtos2.py b/python/test-data/testTypesProtos2.py
new file mode 100644
index 00000000..fa2c6e3d
--- /dev/null
+++ b/python/test-data/testTypesProtos2.py
@@ -0,0 +1,19 @@
+from wypp import *
+
+from typing import Protocol
+import abc
+
+class Animal(Protocol):
+    @abc.abstractmethod
+    def makeSound(loadness: float) -> str: # self parameter omitted!
+        pass
+
+class Dog:
+    # incorrect implementation of the Animal protocol
+    def makeSound(self, loadness: int) -> str:
+        return f"{loadness} wuffs"
+
+def doSomething(a: Animal) -> None:
+    print(a.makeSound(3.14))
+
+doSomething(Dog())
diff --git a/python/test-data/testTypesProtos3.err b/python/test-data/testTypesProtos3.err
new file mode 100644
index 00000000..4be46a5f
--- /dev/null
+++ b/python/test-data/testTypesProtos3.err
@@ -0,0 +1,14 @@
+untypy.error.UntypyTypeError
+Type Dog does not meet the requirements of protocol Animal. The signature of 'makeSound' does not match. Missing required parameter self
+
+given:    'Dog'
+expected: value of type Animal
+
+context: doSomething(a: Animal) -> None
+                        ^^^^^^
+declared at: test-data/testTypesProtos3.py:16
+ 16 | def doSomething(a: Animal) -> None:
+ 17 |     print(a.makeSound(3.14))
+
+caused by: test-data/testTypesProtos3.py:19
+ 19 | doSomething(Dog())
diff --git a/python/test-data/testTypesProtos3.out b/python/test-data/testTypesProtos3.out
new file mode 100644
index 00000000..e69de29b
diff --git a/python/test-data/testTypesProtos3.py b/python/test-data/testTypesProtos3.py
new file mode 100644
index 00000000..334c2424
--- /dev/null
+++ b/python/test-data/testTypesProtos3.py
@@ -0,0 +1,19 @@
+from wypp import *
+
+from typing import Protocol
+import abc
+
+class Animal(Protocol):
+    @abc.abstractmethod
+    def makeSound(self, loadness: float) -> str:
+        pass
+
+class Dog:
+    # incorrect implementation of the Animal protocol
+    def makeSound(loadness: int) -> str: # self parameter omitted!
+        return f"{loadness} wuffs"
+
+def doSomething(a: Animal) -> None:
+    print(a.makeSound(3.14))
+
+doSomething(Dog())
diff --git a/python/test-data/testTypesProtos4.py b/python/test-data/testTypesProtos4.py
new file mode 100644
index 00000000..5ce3a127
--- /dev/null
+++ b/python/test-data/testTypesProtos4.py
@@ -0,0 +1,27 @@
+from wypp import *
+
+from typing import Protocol
+import abc
+
+class Interface(Protocol):
+    abc.abstractmethod
+    def meth(self) -> Callable[[int], int]:
+        pass
+
+class ConcreteCorrect:
+    def meth(self) -> Callable[[int], int]:
+        return lambda x: x + 1
+
+def bar(s: str) -> int:
+    return len(s)
+
+class ConcreteWrong:
+    def meth(self) -> Callable[[int], int]:
+        return lambda x: bar(x) # invalid call of bar with argument of type int
+
+def foo(obj: Interface) -> int:
+    fn = obj.meth()
+    return fn(2)
+
+print(foo(ConcreteCorrect()))
+print(foo(ConcreteWrong()))
diff --git a/python/test-data/testTypesRecordInheritance.py b/python/test-data/testTypesRecordInheritance.py
new file mode 100644
index 00000000..6dfd7037
--- /dev/null
+++ b/python/test-data/testTypesRecordInheritance.py
@@ -0,0 +1,14 @@
+from wypp import *
+
+@record
+class Point2D:
+    x: int
+    y: int
+
+@record
+class Point3D(Point2D):
+    z: int
+
+print(Point3D(1,2,3))
+Point3D(1,2, "foo")
+
diff --git a/python/test-data/testTypesReturn.py b/python/test-data/testTypesReturn.py
new file mode 100644
index 00000000..f1aaa969
--- /dev/null
+++ b/python/test-data/testTypesReturn.py
@@ -0,0 +1,8 @@
+def foo(flag: bool) -> int:
+    print('Hello World')
+    if flag:
+        return 1
+    else:
+        return 'you stupid'
+
+foo(False)
diff --git a/python/test-data/testTypesSequence1.err b/python/test-data/testTypesSequence1.err
new file mode 100644
index 00000000..eb475241
--- /dev/null
+++ b/python/test-data/testTypesSequence1.err
@@ -0,0 +1,13 @@
+untypy.error.UntypyTypeError
+given:    1
+expected: value of type Sequence
+
+context: foo(seq: Sequence) -> None
+                  ^^^^^^^^
+declared at: test-data/testTypesSequence1.py:3
+  3 | def foo(seq: Sequence) -> None:
+  4 |     print(seq)
+  5 |     pass
+
+caused by: test-data/testTypesSequence1.py:10
+ 10 | foo(1) # should fail
diff --git a/python/test-data/testTypesSequence1.out b/python/test-data/testTypesSequence1.out
new file mode 100644
index 00000000..f4ce9837
--- /dev/null
+++ b/python/test-data/testTypesSequence1.out
@@ -0,0 +1,3 @@
+[1, 2, 3]
+('bar', 'baz')
+Hello!
diff --git a/python/test-data/testTypesSequence1.py b/python/test-data/testTypesSequence1.py
new file mode 100644
index 00000000..8a34e873
--- /dev/null
+++ b/python/test-data/testTypesSequence1.py
@@ -0,0 +1,10 @@
+from wypp import *
+
+def foo(seq: Sequence) -> None:
+    print(seq)
+    pass
+
+foo([1,2,3])
+foo( ("bar", "baz") )
+foo("Hello!")
+foo(1) # should fail
diff --git a/python/test-data/testTypesSequence2.err b/python/test-data/testTypesSequence2.err
new file mode 100644
index 00000000..abccc73a
--- /dev/null
+++ b/python/test-data/testTypesSequence2.err
@@ -0,0 +1,13 @@
+untypy.error.UntypyTypeError
+given:    'Hello!'
+expected: value of type Sequence[int]
+
+context: foo(seq: Sequence[int]) -> None
+                  ^^^^^^^^^^^^^
+declared at: test-data/testTypesSequence2.py:3
+  3 | def foo(seq: Sequence[int]) -> None:
+  4 |     print(seq)
+  5 |     pass
+
+caused by: test-data/testTypesSequence2.py:9
+  9 | foo("Hello!") # should fail
diff --git a/python/test-data/testTypesSequence2.out b/python/test-data/testTypesSequence2.out
new file mode 100644
index 00000000..ec306cac
--- /dev/null
+++ b/python/test-data/testTypesSequence2.out
@@ -0,0 +1,2 @@
+[1, 2, 3]
+(4, 5)
diff --git a/python/test-data/testTypesSequence2.py b/python/test-data/testTypesSequence2.py
new file mode 100644
index 00000000..290dee38
--- /dev/null
+++ b/python/test-data/testTypesSequence2.py
@@ -0,0 +1,9 @@
+from wypp import *
+
+def foo(seq: Sequence[int]) -> None:
+    print(seq)
+    pass
+
+foo([1,2,3])
+foo( (4,5)  )
+foo("Hello!") # should fail
diff --git a/python/test-data/testTypesSubclassing1.py b/python/test-data/testTypesSubclassing1.py
new file mode 100644
index 00000000..3bba042c
--- /dev/null
+++ b/python/test-data/testTypesSubclassing1.py
@@ -0,0 +1,25 @@
+from wypp import *
+
+class AnimalFood:
+    def __init__(self, name: str):
+        self.name = name
+
+class Animal:
+    def feed(self, food: AnimalFood) -> None:
+        print(f'Tasty animal food: {food.name}')
+
+class DogFood(AnimalFood):
+    def __init__(self, name: str, weight: float):
+        super().__init__(name)
+        self.weight = weight
+
+class Dog(Animal):
+    # Dog provides an invalid override for feed
+    def feed(self, food: DogFood) -> None:
+        print(f'Tasty dog food: {food.name} ({food.weight}g)')
+
+def feedAnimal(a: Animal) -> None:
+    a.feed(AnimalFood('some cat food'))
+
+dog = Dog()
+feedAnimal(dog)
diff --git a/python/test-data/testTypesTuple1.err b/python/test-data/testTypesTuple1.err
new file mode 100644
index 00000000..c9ae74fb
--- /dev/null
+++ b/python/test-data/testTypesTuple1.err
@@ -0,0 +1,12 @@
+untypy.error.UntypyTypeError
+given:    1
+expected: value of type tuple[int, ...]
+
+context: foo(l: tuple[int, ...]) -> int
+                ^^^^^^^^^^^^^^^
+declared at: test-data/testTypesTuple1.py:1
+  1 | def foo(l: tuple[int, ...]) -> int:
+  2 |     return len(l)
+
+caused by: test-data/testTypesTuple1.py:5
+  5 | foo(1)
diff --git a/python/test-data/testTypesTuple1.out b/python/test-data/testTypesTuple1.out
new file mode 100644
index 00000000..00750edc
--- /dev/null
+++ b/python/test-data/testTypesTuple1.out
@@ -0,0 +1 @@
+3
diff --git a/python/test-data/testTypesTuple1.py b/python/test-data/testTypesTuple1.py
new file mode 100644
index 00000000..29ea2c3c
--- /dev/null
+++ b/python/test-data/testTypesTuple1.py
@@ -0,0 +1,5 @@
+def foo(l: tuple[int, ...]) -> int:
+    return len(l)
+
+print(foo((1,2,3)))
+foo(1)
diff --git a/python/file-tests/typeEnums.py b/python/test-data/typeEnums.py
similarity index 100%
rename from python/file-tests/typeEnums.py
rename to python/test-data/typeEnums.py
diff --git a/python/file-tests/typeRecords.py b/python/test-data/typeRecords.py
similarity index 100%
rename from python/file-tests/typeRecords.py
rename to python/test-data/typeRecords.py
diff --git a/python/file-tests/typeUnion.py b/python/test-data/typeUnion.py
similarity index 100%
rename from python/file-tests/typeUnion.py
rename to python/test-data/typeUnion.py
diff --git a/python/test-data/wrong-caused-by.py b/python/test-data/wrong-caused-by.py
new file mode 100644
index 00000000..d7ca90e2
--- /dev/null
+++ b/python/test-data/wrong-caused-by.py
@@ -0,0 +1,30 @@
+from wypp import *
+
+@record(mutable=True)
+class StreetM:
+    name: str
+    cars: list[CarM]
+    # Einbiegen eines Auto auf eine Straße
+    # SEITENEFFEKT: Liste der Autos wird geändert.
+    def turnIntoStreet(self: StreetM, car: Car) -> None:
+        if car not in self.cars:
+            self.cars.append(car)
+    # Verlassen einer Straße
+    # SEITENEFFEKT: Liste der Autos wird geändert.
+    def leaveStreet(self: StreetM, car: Car) -> None:
+        if car in self.cars:
+            self.cars.remove(car)
+
+@record(mutable=True)
+class CarM:
+    licensePlate: str
+    color: str
+
+@record
+class Car:
+    licensePlate: str
+    color: str
+
+redCarM = CarM('OG PY 123', 'rot')
+mainStreetM = StreetM('Hauptstraße', [redCarM])
+mainStreetM.turnIntoStreet(redCarM)
diff --git a/python/tests/sample.py b/python/tests/sample.py
new file mode 100644
index 00000000..f020c989
--- /dev/null
+++ b/python/tests/sample.py
@@ -0,0 +1,124 @@
+from writeYourProgram import *
+import math
+from untypy import typechecked
+import typing
+
+Drink = Literal["Tea", "Coffee"]
+
+# berechnet wieviele Tassen ich von einem Getränk trinken darf
+@typechecked
+def canDrink(d: Drink) -> int:
+    if d == "Tea":
+        return 5
+    elif d == "Coffee":
+        return 1
+
+# A shape is one of the following:
+# - a circle (Circle)
+# - a square (Square)
+# - an overlay of two shapes (Overlay)
+Shape = typing.Union[typing.ForwardRef('Circle'), typing.ForwardRef('Square'), typing.ForwardRef('Overlay')]
+Shape = typing.Union['Circle', 'Square', 'Overlay']
+
+# A point consists of
+# - x (float)
+# - y (float)
+@record
+class Point:
+    x: float
+    y: float
+# Point: (float, float) ->  Point
+# For some Point p
+# p.x: float
+# p.y: float
+
+# point at x=10, y=20
+p1 = Point(10, 20)
+
+# point at x=30, y=50
+p2 = Point(30, 50)
+
+# point at x=40, y=30
+p3 = Point(40, 30)
+
+# A circle consists of
+# - center (Point)
+# - radius (float)
+@record
+class Circle:
+    center: Point
+    radius: float
+
+# Circle: (Point, float) -> Circle
+# For some circle c
+# c.center: Point
+# c.radius: float
+
+# circle at p2 with radius=20
+c1 = Circle(p2, 20)
+
+# circle at p3 with radius=15
+c2 = Circle(p3, 15)
+
+# A square (parallel to the coordinate system) consists of
+# - lower-left corner (Point)
+# - size (float)
+@record
+class Square:
+    corner: Point
+    size: float
+
+# square at p1 with size=40
+s1 = Square(p1, 40)
+
+# Square: (Point, float) -> Square
+# For some square s
+# s.corner: Point
+# s.size: float
+
+# An overlay consists of
+# - top (Shape)
+# - bottom (Shape)
+@record
+class Overlay:
+    top: Shape
+    bottom: Shape
+
+# Overlay: (Shape, Shape) -> Overlay
+# For some overlay:
+# o.top: Shape
+# o.bottom: Shape
+
+# overlay of circle c1 and square s1
+o1 = Overlay(c1, s1)
+# Overlay of overlay o1 and circle c2
+o2 = Overlay(o1, c2)
+
+# Calculate the distance between two points
+@typechecked
+def distance(p1: Point, p2: Point) -> float:
+    w = p1.x - p2.x
+    h = p1.y - p2.y
+    dist = math.sqrt(w**2 + h**2)
+    return dist
+
+# Is a point within a shape?
+@typechecked
+def pointInShape(point: Point, shape: Shape) -> bool:
+    px = point.x
+    py = point.y
+    if type(shape) == Circle:
+        return distance(point, shape.center) <= shape.radius
+    elif type(shape) == Square:
+        corner = shape.corner
+        size = shape.size
+        return (
+            px >= corner.x and
+            px <= corner.x + size and
+            py >= corner.y and
+            py <= corner.y + size
+        )
+    elif type(shape) == Overlay:
+        return pointInShape(point, shape.top) or pointInShape(point, shape.bottom)
+    else:
+        uncoveredCase()
diff --git a/python/tests/testDeepEq.py b/python/tests/testDeepEq.py
index 412b73f5..3883c1b9 100644
--- a/python/tests/testDeepEq.py
+++ b/python/tests/testDeepEq.py
@@ -1,5 +1,5 @@
 import unittest
-import testSample
+import sample
 from writeYourProgram import *
 from writeYourProgram import deepEq
 import writeYourProgram as wypp
@@ -67,27 +67,27 @@ def test_eq(self):
         self.assertFalse(deepEq(C(42.0), D(42.0000000000001),
                                 structuralObjEq=False, floatEqWithDelta=True))
         self.assertTrue(deepEq(
-            testSample.Point(1.0, 2.0), testSample.Point(1.00000000001, 2),
+            sample.Point(1.0, 2.0), sample.Point(1.00000000001, 2),
             structuralObjEq=True, floatEqWithDelta=True
         ))
         self.assertFalse(deepEq(
-            testSample.Point(1.0, 2.0), testSample.Point(1.00000000001, 2),
+            sample.Point(1.0, 2.0), sample.Point(1.00000000001, 2),
             structuralObjEq=False, floatEqWithDelta=True
         ))
         self.assertTrue(deepEq(
-            Point(1.0, 2.0), testSample.Point(1.0, 2.0),
+            Point(1.0, 2.0), sample.Point(1.0, 2.0),
             structuralObjEq=True, floatEqWithDelta=True)
         )
         self.assertFalse(deepEq(
-            Point(1.0, 2.0), testSample.Point(1.0, 2.0),
+            Point(1.0, 2.0), sample.Point(1.0, 2.0),
             structuralObjEq=False, floatEqWithDelta=True)
         )
         self.assertFalse(deepEq(
-            testSample.Point(1.0, 3.0), testSample.Point(1.00000000001, 2),
+            sample.Point(1.0, 3.0), sample.Point(1.00000000001, 2),
             structuralObjEq=True, floatEqWithDelta=True
         ))
         self.assertFalse(deepEq(
-            testSample.Point(1.0, 3.0), testSample.Point(1.00000000001, 2),
+            sample.Point(1.0, 3.0), sample.Point(1.00000000001, 2),
             structuralObjEq=False, floatEqWithDelta=True
         ))
         self.assertTrue(deepEq(A(2), A(2), structuralObjEq=True, floatEqWithDelta=True))
diff --git a/python/tests/testSample.py b/python/tests/testSample.py
index 1c248d32..65ab52b0 100644
--- a/python/tests/testSample.py
+++ b/python/tests/testSample.py
@@ -1,125 +1,9 @@
 from writeYourProgram import *
 import unittest
-import math
+from sample import *
 
 setDieOnCheckFailures(True)
 
-Drink = Literal["Tea", "Coffee"]
-
-# berechnet wieviele Tassen ich von einem Getränk trinken darf
-def canDrink(d: Drink) -> int:
-    if d == "Tea":
-        return 5
-    elif d == "Coffee":
-        return 1
-
-# A shape is one of the following:
-# - a circle (Circle)
-# - a square (Square)
-# - an overlay of two shapes (Overlay)
-Shape = Union[ForwardRef('Circle'), ForwardRef('Square'), ForwardRef('Overlay')]
-
-# A point consists of
-# - x (float)
-# - y (float)
-@record
-class Point:
-    x: float
-    y: float
-# Point: (float, float) ->  Point
-# For some Point p
-# p.x: float
-# p.y: float
-
-# point at x=10, y=20
-p1 = Point(10, 20)
-
-# point at x=30, y=50
-p2 = Point(30, 50)
-
-# point at x=40, y=30
-p3 = Point(40, 30)
-
-# A circle consists of
-# - center (Point)
-# - radius (float)
-@record
-class Circle:
-    center: Point
-    radius: float
-
-# Circle: (Point, float) -> Circle
-# For some circle c
-# c.center: Point
-# c.radius: float
-
-# circle at p2 with radius=20
-c1 = Circle(p2, 20)
-
-# circle at p3 with radius=15
-c2 = Circle(p3, 15)
-
-# A square (parallel to the coordinate system) consists of
-# - lower-left corner (Point)
-# - size (float)
-@record
-class Square:
-    corner: Point
-    size: float
-
-# square at p1 with size=40
-s1 = Square(p1, 40)
-
-# Square: (Point, float) -> Square
-# For some square s
-# s.corner: Point
-# s.size: float
-
-# An overlay consists of
-# - top (Shape)
-# - bottom (Shape)
-@record
-class Overlay:
-    top: Shape
-    bottom: Shape
-
-# Overlay: (Shape, Shape) -> Overlay
-# For some overlay:
-# o.top: Shape
-# o.bottom: Shape
-
-# overlay of circle c1 and square s1
-o1 = Overlay(c1, s1)
-# Overlay of overlay o1 and circle c2
-o2 = Overlay(o1, c2)
-
-# Calculate the distance between two points
-def distance(p1: Point, p2: Point) -> float:
-    w = p1.x - p2.x
-    h = p1.y - p2.y
-    dist = math.sqrt(w**2 + h**2)
-    return dist
-
-# Is a point within a shape?
-def pointInShape(point: Point, shape: Shape) -> bool:
-    px = point.x
-    py = point.y
-    if type(shape) == Circle:
-        return distance(point, shape.center) <= shape.radius
-    elif type(shape) == Square:
-        corner = shape.corner
-        size = shape.size
-        return (
-            px >= corner.x and
-            px <= corner.x + size and
-            py >= corner.y and
-            py <= corner.y + size
-        )
-    elif type(shape) == Overlay:
-        return pointInShape(point, shape.top) or pointInShape(point, shape.bottom)
-    else:
-        uncoveredCase()
-
 class TestSample(unittest.TestCase):
     def test_sample(self):
         check(pointInShape(p2, c1), True)