Limits and evaluation¶
Use the product for human-reviewed drafts from bounded business text. Source references help you inspect a result; they do not certify its meaning.
Scope¶
The runtime does not retrieve documents, verify facts, execute tools, assign tasks, send messages, perform OCR or transcribe audio. It does not establish independent native-language customer accuracy. Inputs and outputs can contain private business information; secure access, logging and storage in your application.
Evidence limitations¶
The model selects segment IDs and the backend copies those segments. Exact quote/offset checks establish literal source location, not semantic entailment. A wrong or incomplete summary can still include an exact source quote. Owner/date postprocessing does not turn an inferred task into an authorized business action.
Actual qualification¶
The released runtime passed its 60-record fictional development evaluation on ml.g6e.4xlarge and actual English/Spanish/Portuguese certification batch on ml.g6e.2xlarge. The corpus is generated development material with no independent human/native acceptance. These technical checks do not measure customer factuality or generalization.
A measured cold deployment took about 21 minutes. Startup is separate from per-request latency. Allow sufficient startup time, inspect failed resources and clean up after timeouts. The qualified region is us-east-1; catalog availability elsewhere does not qualify every regional configuration.
Read the evaluation scope, model card and source/license notices. Evaluate your own authorized workload and reviewer process before relying on drafts.
Last updated November 15, 2026 · Suggest an improvement