diff --git a/my/body/exercise.py b/my/body/exercise.py
new file mode 100644
index 0000000..2ce53bd
--- /dev/null
+++ b/my/body/exercise.py
@@ -0,0 +1,109 @@
+'''
+My exercise data, arbitrated between differen sources (mainly, Endomondo and various manual plaintext notes)
+
+This is probably too specific to my needs, so later I will move it away to a personal 'layer'.
+For now it's worth keeping it here as an example and perhaps utility functions might be useful for other HPI modules.
+'''
+
+from datetime import datetime, timedelta
+
+from my.config import exercise as config
+
+
+# todo predataframe?? entries??
+def cross_trainer_data():
+    # FIXME manual entries
+
+    from porg import Org
+    # TODO FIXME should use all org notes and just query from them?
+    wlog = Org.from_file(config.workout_log)
+    cross_table = wlog.xpath('//org[heading="Cross training"]//table')
+    return cross_table.lines
+    # todo hmm, converting an org table directly to pandas kinda makes sense?
+    # could have a '.dataframe' method in orgparse, optional dependency
+
+
+import pytz
+# FIXME how to attach it properly?
+tz = pytz.timezone('Europe/London')
+
+def cross_trainer_manual_dataframe():
+    '''
+    Only manual org-mode entries
+    '''
+    import pandas as pd
+    df = pd.DataFrame(cross_trainer_data())
+
+    from ..core.orgmode import parse_org_datetime
+    df['date'] = df['date'].apply(parse_org_datetime)
+
+    def tzify(d: datetime) -> datetime:
+        assert d.tzinfo is None, d
+        return tz.localize(d)
+
+    df['date'] = df['date'].apply(tzify)
+
+    # TODO convert duration as well
+    #
+    return df
+
+
+def cross_trainer_dataframe():
+    '''
+    Attaches manually logged data (which Endomondo can't capture) and attaches it to Endomondo
+    '''
+    import pandas as pd
+
+    from ..endomondo import dataframe as EDF
+    edf = EDF()
+    edf = edf[edf['sport'].str.contains('Cross training')]
+
+    # Normalise and assume single bout of exercise per day
+    # TODO this could be useful for other providers..
+    grouped = edf.set_index('start_time').groupby(lambda t: t.date())
+    singles = []
+    for day, grp in grouped:
+        if len(grp) != 1:
+            # FIXME yield runtimeerror
+            continue
+        singles.append(grp)
+    edf = pd.concat(singles)
+    edf = edf.reset_index()
+
+    mdf = cross_trainer_manual_dataframe()
+    # now for each manual entry, find a 'close enough' endomondo entry
+    rows = []
+    idxs = []
+    for i, row in mdf.iterrows():
+        # todo rename 'date'??
+        mdate = row['date']
+        close = edf[edf['start_time'].apply(lambda t: pd_date_diff(t, mdate)).abs() < timedelta(hours=3)]
+        if len(close) == 0:
+            # FIXME emit warning -- nothing matched
+            continue
+        if len(close) > 1:
+            # FIXME emit warning
+            continue
+        loc = close.index[0]
+        # FIXME check and make defensive
+        # assert loc not in idxs, (loc, row)
+        idxs.append(loc)
+        rows.append(row)
+    mdf = pd.DataFrame(rows, index=idxs)
+
+    df = edf.join(mdf, rsuffix='_manual')
+    # TODO arbitrate kcal, duration, avg hr
+    # compare power and hr?
+    return df
+
+
+def stats():
+    from ..core import stat
+    return stat(cross_trainer_data())
+
+
+def pd_date_diff(a, b) -> timedelta:
+    # ugh. pandas complains when we subtract timestamps in different timezones
+    assert a.tzinfo is not None, a
+    assert b.tzinfo is not None, b
+    return a.to_pydatetime() - b.to_pydatetime()