A fourth read by a separately built reader covered all 60 disputed Verified tasks: 14 state the fix, 13 point to it, 33 have neither. The reader never saw the pipeline verdict and works in two blind steps. No tier changed and no headline count moved. The 14 that state it: django__django-11087, django__django-11138, django__django-13809, django__django-14725, django__django-14771, pydata__xarray-4687, pylint-dev__pylint-4661, pytest-dev__pytest-5840, scikit-learn__scikit-learn-13439, scikit-learn__scikit-learn-13779, scikit-learn__scikit-learn-25102, scikit-learn__scikit-learn-25232, sphinx-doc__sphinx-8035, sympy__sympy-13877. Read how the fourth read works →
Boundary sample, each group reported alone: Settled A where the earlier read disagreed: 1 of 4 (states the fix); Settled A by a third check (drawn at random): 5 of 6 (states the fix); Tier A, both first checks agreed (drawn at random): 9 of 9 (states the fix); Tier B, split checks (one check had A): 5 of 8 (points or states); Tier B, settled by a third check (one check had A, drawn at random): 8 of 9 (points or states). Open the table →
The Tier A headline rests on tasks where the first two checks agreed. The fourth read said “states” on all 9 drawn from them, so the headline figure stands. Read the validation →
The four settled-A tasks the earlier read disagreed with, as the fourth read found them: django__django-16145 states, django__django-7530 points, sphinx-doc__sphinx-10435 clean, sympy__sympy-18199 points. Their tiers are unchanged and they stay counted apart from the headline. Read the validation →
sympy__sympy-20322 re-read: the Tier C flag stands. The discussion holds a diff of Mul._eval_is_integer, the same function the gold patch edits, with the same numerators and denominators loop. The first read had compared a different diff (radsimp.fraction). Its stored label moved from unsure to agree. Open the task →
sphinx-doc__sphinx-8713 checked: 3 of its 5 matched lines are sibling code the issue quotes, so this Tier C flag is the thinnest of the 12 on Lite. The verdict is unchanged and no label was added. Open the task →
Lite run again through the audit tool: the same tier as the public table on 239 of 300 tasks, 76 of 93 table Tier A or C flags flagged again, 86 of 88 clear tasks still clear. The model is sampled again on each run; every differing task is listed. Read the validation →
Check your own benchmark’s tasks
This is the check Leaklens runs on every Verified and Lite task. Upload your benchmark and get the same per-task report, with the evidence quoted. Sign in with an email, no card.