differentiation_check.py writes layer_f_note into every report:
Layer F validates local sensitivity and FD consistency; it is not a public autograd guarantee.
- Layer F is internal validation:
scripts/differentiation_check.py, optional torch/jax toy probes,--shield-inter-simon licensed hosts. tests/vendor/diff/— engineering only, not product marketing.
- Production shield autograd (
enable_gradon the real shield QP). - “Fully validated differentiable runtime shield stack” in README, decks, or release notes.
Do not change public narrative until all pass on licensed hosts:
- Production-path autograd tests on shield objective
- Autograd vs FD within declared tolerances
- Stability around active-set changes
- Maintainer milestone + explicit doc update
Until then: ROADMAP.md item 2 stays deferred.