## Update: validation boundary

After publication, an external technical review confirmed that ASDR is reproducible as a research prototype: the test suite passes, the reference scenario reproduces, and the composition-layer framing remains a valid AI safety research direction.

The review also identified an important limitation: the current reference scenario is primarily driven by the lexicon-based S_evasion layer. Therefore, the current result should be read as a reproducible breach-candidate under ASDR’s current scoring model, not as a statistically validated universal semantic phase transition.

This does not invalidate the project. It clarifies the next validation target.

The proposed grant will test whether ASDR’s composition-layer signal generalizes beyond:

- the seed scenario

- exact evasion keywords

- TF-IDF lexical scoring

- a single synthetic trace

ASDR should be understood as a reproducible research prototype seeking validation, not a completed production safety system.