20.6 A compositional account of AI safety

The foundry perspective does not replace alignment, robustness, privacy, security, or human-factors research. It contributes a structural layer that connects them. A safety requirement becomes operational when one can say which diagram should commute, which perturbations are meaningful, which variation is irrelevant, where failure can be localized, and what repair may be admitted.

Five recurring safety failures acquire compositional forms:

1. provenance loss: source restriction no longer commutes with claim transport;

2. scope drift: a claim is pushed into a population or use not covered by its qualifier;

3. context collapse: locally valid sections are glued across an obstructed overlap;

4. argument laundering: grounds survive summarization while the warrant, rebuttal, or uncertainty disappears; and

5. state contamination: a candidate or stale artifact is treated as admitted current knowledge.

These failures are related but not interchangeable. Their obstruction types determine whether the correct response is retrieval, qualification, cover refinement, argument repair, refresh, quarantine, or abstention. The system learns from the repair traces as well: repeated gate failures reveal weak workflow skills, incomplete covers, and brittle transports. Policy optimization can then improve the proposal and audit process while the admission boundary remains fixed.