[findalexli]

Task finding: #47. Known issue 2 (time limit mismatch).

What and why

The instruction gives "T=5 test cases per input" and "30 seconds per test case", but TIME_LIMIT = 60 kills an input at 60 s, and one timeout invalidates the whole submission. TIME_LIMIT is now 150 s. The README also calls the 50 inputs "test cases", which could mean 30 s per input; this patch takes the instruction's reading, 30 s per grid. The score formula and the pool stay as they are.

Partial fix. Eight inputs share 4 CPUs, so a solver doing fixed work above a computed 75 s of CPU per input still times out. A pool change would move deadline solvers' scores, so it is left out and noted under Known issues. No real submission is shown affected: an Opus 5.5 run chose 20 s per input. The deadline rows are edited versions of it.

sforge run. Nominal grading time is up to 1260 s, inside timeout_sec but past that command's 900 s final wait. In the real harness, a 140 s submission was still grading at that wait and got no score on the new judge, as on the old; a 125 s one scored 19.49, and an earlier valid score survives.

Old vs new judge

Built from source on 8-vCPU workers, judge at 4 CPUs, judge_local.py and sforge eval. Old judge from 1187e2c, whose task files match f92857a. Raw score (0..100). invalid is inf.

Submission	n	Old	New
Generic valid control	3	46676.8 (0.00)	same
Opus 5.5 final submission, 20 s deadline	2 old, 3 new	873.5-873.7 (19.51)	873.7 (19.51)
45 s deadline, valid on both	2	878.8-878.9 (19.48)	878.8 (19.48)
125 s deadline	2	invalid, 0 of 50 OK	877.3-877.9 (19.48-19.49), 50 of 50 OK
140 s deadline	2	invalid, 0 of 50 OK	877.3 (19.49), 50 of 50 OK, 958 s
Sleeps 170 s (over the limit)	2	invalid, 0 of 50 OK	invalid, 0 of 50 OK

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, and the fixed entry is removed from Known issues.
• Old vs new judge on the same submissions, results above.
• New blobs: none.

[findalexli]

Closing for now. The 60 s per input conflicts with the stated 30 s per case, but no real submission we checked loses an input to it; the solvers we ran budget about 20-25 s per input. We can reopen if one turns up.