In an experimental feature exploration, I found some astounding behaviour when I try to load a Unix file that is UTF-8 on a mainframe system using Cobol and declaring the FD record as Unicode alphanumeric.
If my record length is 10 (ie. 10 chunk PIC U(10) , first 10 characters are correctly loaded. Then 30 (apparently 3x the length according to exploration) is skipped, then it reads again 10 characters. Then the next 10 characters are read into my next record.
Source Code of my program:

IDENTIFICATION DIVISION.
       PROGRAM-ID.   loadutf8.
       ENVIRONMENT DIVISION.
       INPUT-OUTPUT SECTION.
       FILE-CONTROL.
           SELECT XMLFILE ASSIGN TO "XML".
       DATA DIVISION.
       FILE SECTION.
       FD XMLFILE RECORDING MODE F.
          01  chunk PIC  U(10).
       WORKING-STORAGE SECTION.
          01  EOF   PIC  X.
       PROCEDURE DIVISION.
       START-PROGRAM.
           OPEN INPUT XMLFILE.
           PERFORM WITH TEST after  UNTIL EOF = 'T'
              READ XMLFILE
              AT END MOVE 'T' TO EOF
              NOT AT END
               DISPLAY FUNCTION DISPLAY-of (chunk)
              END-READ
           END-PERFORM.
           CLOSE XMLFILE.
           GOBACK.
       END PROGRAM loadutf8.

JOB-CARD:

//COBOL     EXEC IGYWCLG,LNGPRFX=IGY630 
//SYSIN       DD DISP=SHR,DSN=COB.SRC(loadutf8) 
//GO.XML      DD PATH='/u/utf8.xml'

My UTF-8 file:

<?xml  ?>
<!-- 0 --><!-- 1 --><!-- 2 --><!-- 3 --><!-- 4 --><!-- 5 --><x>???</x>

Output observed:

<?xml  ?>   
<!-- 3 -->

To me, it looks like consistently reading the chunk according to size, skip 3 times and then goes to next chunk size, etc..
What could be causing this?
Is there a best practice to it, how to load a Unix file to a XML and use a variable that has the usage UTF-8? Preferably w/o any hacks, just using 'standard' language features.
Just asking this out of curiosity, any idea on how to explain the observed outcome is appreciated.