How Harmonia uses it
Strands cognition uses native Bedrock model IDs. The host selects a role policy, reserves the configured cost, dispatches the request and records the actual model and usage. There is no implicit provider fallback.
Amazon Transcribe owns timed speech transcription; it is not a Bedrock text-model call. ElevenLabs owns instrumental music. See generated media.
Policy and pricing
role_models.py defines role-specific token limits, sampling and deadlines. Model overrides use the corresponding role environment variable. BEDROCK_TEXT_PRICING_JSON supplies rates keyed by exact model ID, in USD per million input/output tokens. Missing rates fail closed; no guessed rate is substituted.
HARMONIA_ALLOW_PAID_AWS defaults false. Generated images, video and music also require GENERATIVE_MEDIA_ENABLED. A true configuration flag does not prove provider readiness or an authenticated invocation.
Evidence status
agent/harmonia_agent/role_models.py, model_catalog.py, usage.py and team_runtime.py. Local contract verification is separate from billed usage reconciliation and live provider proof, which remain pending for this AWS edition.