OWNED AI REPRESENTATIVE PORTABILITY SUITE — LIMITATIONS

This suite does not establish arbitrary workflow or enterprise portability; a population-level success or failure rate; relative model quality; completeness of any commercial provider export; portability of hidden reasoning or identical stochastic replay; legitimate real-world authority; or truth merely because hashes and audit chains verify.

Experimental limits:
- five operator-designed internal sentinels from one operating environment
- synthetic or frozen/redacted fixtures rather than customer migrations
- no production side effects
- one frozen prompt/schema/validator family per controlling campaign
- Git used a deterministic local bare remote rather than GitHub
- publishing used an exactly-once sandbox rather than Ghost
- healthcare outcome was explicitly no-action and simulated
- privacy minimization measured retained bytes and relationship removal, not formal privacy guarantees
- V4 combines safe fail-closed responses with bounded continuation according to disclosed condition-specific expectations
- serialization failure is distinct from semantic failure and should not be interpreted as unsafe action by itself

Next stronger evidence: independently authored fixtures, actual vendor exports, a separately implemented validator, external operator review, and a successor environment administered outside the original repositories.