Beginner-path validation

Executed on 2026-09-23 and last re-executed on 2026-09-30. The second review made kernel sizes runtime arguments and changed how the checker reports unfinished work. The course focuses on explanations, editable exercises and feedback; it makes no performance claims.

Check	Observed result	Evidence
Seven CPU reference exercises	40 named checks passed	cpu-solutions.log
Three Triton reference exercises	36 GPU checks passed on RTX 4080 SUPER, including two that launch the add kernel into a guarded buffer and fail if any lane stores past N	all-solutions.log, which includes CPU and GPU checks
Checker regression tests	15 tests passed (9 checker, 6 workshop runner)	checker-tests.log
Portfolio regression suite	48 of 49 passed; the two-GPU test skipped, and the interpreter-only class, skipped at setup, runs separately (5 passed); GPU tests also check that the add and fused-GEMM checkers reject deliberately broken kernels	portfolio-regression.log; interpreter run in interpreter.log
Unfinished starter behavior	Reports every TODO check and returns exit code 2	starter-example.log
Static review	Ruff passed across src, tests, scripts, examples and learning	Local Ruff execution
Worked trace commands	Indexing, strides, reduction, softmax and matmul all ran	python -m learning.explain TOPIC
Course source identity	Manifest of the 30 maintained course files (editable exercises and the journal are excluded), refreshed by scripts/course_manifest.py and checked against the files by tests/test_docs.py	source-sha256.json; workshop sources are identified by source_sha256 in their reports

CPU solutions and checker tests were executed with python -S, which disables site-package loading. This verifies that the first seven exercises do not need PyTorch, Triton, NumPy or a GPU. GPU reference solutions ran with the existing WSL Python 3.12/PyTorch 2.11/Triton 3.6 environment. They exercise actual GPU loads, reductions and stores, including ragged sizes, FP16/FP32, row gaps and extreme finite softmax values.

The nine checker tests cover good reference answers, unfinished code, a wrong answer, NaN comparison, hints without loading code, syntax errors, unavailable GPU handling, separate learner/reference paths, and an unfinished function that must not hide results for a later, finished one. They create temporary fake student files instead of depending on real starters remaining unfinished, so learners can freely complete their exercises without breaking the harness tests.

Reference execution is explicitly labeled not learner completion. The ten starter files intentionally contain TODOs; no personal progress is inferred from the reference checks. The minimal GPU exercise wrappers assume the checker's stated contracts; the production-style validations remain in src/kernel_portfolio/ops.py.

Reproduce from the repository root:

python -S -m learning.check all --solution
python -S -m unittest discover -s learning/tests -v
# In the configured GPU environment:
python -m learning.check all --include-gpu --solution
KERNEL_REQUIRE_GPU=1 python -m unittest discover -s tests -v

The local GitHub workflow now includes the learning tests. It has not been run on GitHub. Compiler, device and profiling limitations from the original validation record remain unchanged.