weirdness

README.md

@@ -38,6 +38,15 @@ We do not get the same results for sub_40FD86.  In fact, we don't create the sam
 
 It's unclear why a1, a2, and result are not listed.
 
+### `first_token` weirdness
+
+The [example
+inference](https://github.com/lt-asset/resym/blob/main/training_src/fielddecoder_inf.py)
+scripts get the first token of the output and include it in the prompt.
+Technically this is data leakage, but since the first token is usually part of
+the prompt (a variable name or field expression) it's probably OK?  But it's
+also pretty weird.
+
 ## Todo
 
 * Add field decoding (probably needs Docker)