[findalexli]

Task finding: #46.

What and why

problem.md allows prediction keys "convertible to strings", but the verifier looks up string IDs only, so a complete integer-keyed answer scores about 2 while marked valid. Keys are now converted only when none matches as given, the converted keys are exactly the test nodes, and every value converts with int(). Everything else takes the old path, so string-keyed work scores as before.

No natural affected submission is known. The reproducer changes only a real solver's key type, so please decide whether that meets the good-faith criterion. Mixed keys, and integer confidence keys with string prediction keys, stay as they are.

Old vs new judge

tools/judge_local.py, on two workers; old judge built from 1187e2c, whose task files match f92857a. n=4 per judge per row, reversing order, except the half-not-test-nodes row (n=2). Scores out of 100 move only through the solve-time term; overall excludes it.

Submission	Old	New
Opus 5.5 solution, string keys	90.54-90.63, overall 0.885740	90.53-90.60, overall 0.885740
Same predictions, integer keys	1.95-2.01, overall 0	90.52-90.57, overall 0.885740
Wrong classes, string keys	8.38-8.45, overall 0.064187	8.38-8.41, overall 0.064187
Wrong classes, integer keys	1.94-2.01, overall 0	8.33-8.41, overall 0.064187
Half string, half integer keys	90.31-90.43, overall 0.884238	90.38-90.43, overall 0.884238
Integer keys, half of them not test nodes	1.95-2.02, overall 0	1.95-1.97, overall 0
Integer keys, half the nodes	invalid	invalid
Integer keys, all values "x"	1.97-2.04, overall 0	1.95-2.00, overall 0

The starter and 103 real string-keyed submission archives, one grade per judge: verdicts unchanged, and quality identical on 102 of 104. The other two varied between grades; both use string keys and bypass the conversion. All rows: long text.

Checklist for task changes

• version in task.toml is bumped (any change that can move scores).
• NOTES.md: the change and its score effect are under Changes; no Known issue covered this.
• Old vs new judge on the same submissions, results above.
• New blobs: none.

[findalexli]

Closing for now. Integer keys are allowed and do score about 2, but no real submission we checked uses them (334 public graded entries and 26 fresh solver runs all use string keys). We can reopen if one turns up.