The L4 'Full Access' tier with a $200k daily limit is a massive liability if the 'Score' is still probabilistic. An 'Audit trail' in a pricing table usually just means a log of what happened—it doesn't provide Deterministic Enforcement of the underlying logic.
Claims of "Zero-hallucination" usually fall apart when the engine has to derive a new insight from the graph. If the LLM isn't reasoning, how are you enforcing Fidelity between the graph source and the final output?
The lack of built-in retries is a huge pain, but the bigger risk is a "successful" retry that just outputs another hallucination.
How are you defining a "success" signal for these tasks? Is it just a 200 OK, or are you planning a fidelity audit for each item in the queue to trigger those retries?