Mix of tensors and non-tensors.

Correct ordering: inputs (tensors + scalars) then outputs.

When both NaN and Inf are present, has_nan takes priority.

Mix of near-zero and large reference values.

Kernel that patches torch.cuda.Event.elapsed_time at import → REWARD_HACK.

First blob root with a match is used.

A single NaN in an otherwise correct tensor fails.

NaN outputs must produce finite correctness values with has_nan=True.

NaN triggers early sanity check, bypassing matched-ratio tolerance.

NaN in output produces non-finite abs_error and fails.

NaN in reference (but not output) is also incorrect.

NaN in reference also triggers has_nan.

Kernel that outputs NaN must produce strictly valid JSON traces. Regression test: Python's json.dumps(allow_nan=True) serializes NaN as the J

NaN in output returns Correctness with has_nan=True and finite errors.

NaN is not affected by allow_negative_inf.

-inf vs +inf at same position fails even when allow_negative_inf.

Matching -inf excluded; error computed on remaining elements.

-inf in output but not reference still fails.

-inf in reference but not output still fails.

Negative inf in output fails when reference is finite.

Negative near-zero reference — clamp uses abs, so same behavior.

No blob root contains the file.

Empty status string: log == extra_msg.

Kernel on a non-default stream reports the same time as default stream.

DPS output tensors also get unique data_ptr per call.

DPS output tensors are zero-filled on every call.

Fails when too many elements don't match.

Passes when enough elements match, even if some don't.

Blob root already contains leading components of the relative path.

Non-tensor arguments are returned as-is.

No exception when elapsed_time identity is unchanged.

A correct vector-add kernel produces a PASSED trace.

PASSED traces must preserve finite correctness values (not clobber to null).

Pool overhead is proportional to total_iterations * alignment, not numel * iters.

+inf is not affected by allow_negative_inf.

Full path match is preferred even when a stripped match also exists.

Each preset key resolves to its own preset.

Tensor subclasses (e.g. FakeTensor views) are rejected by strict type() check.

Raises RuntimeError when called more than total_iterations times.

RewardHackDetected when elapsed_time identity differs from the captured address.

When reference is 0, rel_error denominator is clamped to atol.

Reference near zero with a perturbation smaller than atol.

Error at zero reference that exceeds atol.

When required_matched_ratio is 1.0, all must match.

DRAM lock failure → unlocks GPU clocks and returns False.

Unknown device with no env var overrides → False (no subprocess calls).

GPU clock lock failure → returns False, no DRAM lock attempt.