T4 mutation matrix on fa0fd90, 2026-10-16¶
| Report | Tier | Device | Code under test | Verdict |
|---|---|---|---|---|
mutants-cpu.json, .txt |
CPU (T0 + T1) | none | fa0fd90 (repo_commit) |
PASS: 43 killed, 2 equivalent, the 5 identity controls survive |
mutants-gpu.json, .txt |
GPU | RTX 5070 Ti (sm_120) | fa0fd90 | PASS: 7 killed, the identity control survives |
mutants-gpu-rtx4060.json, .txt |
GPU | RTX 4060 (sm_89) | fa0fd90 (see ../rtx-4060-sm89/README.md) |
PASS: 7 killed, the identity control survives |
mutants-cpu-7612fd8.json, .txt |
CPU (T0 + T1) | none | 7612fd8 | FAIL: identity_execute killed (below) |
63b7395, later, renamed the profiles; on its code every tier gave the same outcomes (../rtx-5070-ti-renamed/README.md, ../rtx-4060-sm89-renamed/README.md). Every mutant has the outcome it had on 2026-10-14 (../../2026-10-14/t4/README.md), and on 7612fd8 (GPU tiers) and 3fe4e9c (CPU tier) earlier the same day.
The CPU tier on 7612fd8 killed the identity control identity_execute with
tests/test_h100_mini.py::test_case[m6_040_adv_split] (K 8192, S 3, "contiguous", slow accumulation): the
mutation tool's own re-implementation of execute_plan (tools/mutants/impl.py) combined every scheme but
"serial" in the pairwise order, while mmaemu combines "contiguous" splits serially, and it had no +0 partial for a
contiguous split without k-tiles. The identity control did what it is for: it showed that the injection was unsound
for the hopper preset, whose tests this tier had not run before. 3fe4e9c fixes the tool and changes nothing under
mmaemu/ or tests/; the CPU tier passes from 3fe4e9c on. The GPU tier does not use this re-implementation (its
mutants edit the kernel sources).
Paths: <src> is the worktree, <out> the scratch output directory, ~/ the RTX 4060 host's home directory.