Models, cost, and evaluation

Harmonia chooses a model for each bounded cognitive task. The deterministic coordinator selects the task; a model cannot select its own authority or bypass a stage prerequisite.

Role policy

Registered task family	Default model	Evaluation focus
Intent routing and context assembly	Claude Haiku 4.5	Small strict contracts and faithful operator intent
Nimi analysis	Amazon Nova 2 Lite	Authoritative source references and quote/time grounding
Ryan strategy and Temi planning	Claude Sonnet 4.6	Evidence-linked strategy and approved-strategy fidelity
Noni writing/production and Dara review	Claude Sonnet 4.6	Complete output coverage, grounded content and issue-bound revision
Maya presentation	Claude Haiku 4.5	Bounded component and entity references
Nova liaison and contextual answers	Claude Haiku 4.5	Scoped read-only answers and explicit missing evidence
Resident Dream synthesis	Claude Sonnet 4.6	Sanitized verified observations and bounded experiment proposals

Exact defaults are us.anthropic.claude-haiku-4-5-20251001-v1:0, us.amazon.nova-2-lite-v1:0, and us.anthropic.claude-sonnet-4-6. agent/harmonia_agent/role_models.py defines per-role sampling and token limits. Configured targets fail visibly when unavailable; no provider fallback exists.

Reservation and usage lifecycle

The host estimates the operation’s maximum cost, reserves its budget atomically, calls the provider, and records usage or an unresolved outcome. Stable operation identities connect reservations, model usage, execution receipts and verification. A retry may reuse durable accounting but does not automatically authorize another provider call. Unknown external outcomes require reconciliation. Text pricing comes from BEDROCK_TEXT_PRICING_JSON, keyed by exact model ID with positive input and output USD-per-million-token rates. The version is aws-configured-2026-09-10. Missing prices fail admission; they are not zero-cost operations. Operators must supply rates applicable to their model, region and account. Configured estimates are not a claim about actual billed charges.

Provider operation	Admission input
Amazon Transcribe	`TRANSCRIBE_COST_PER_SECOND_USD`, including the host’s minimum billed duration
Nova Canvas	`IMAGE_MAX_COST_USD` ceiling
Nova Reel	`NOVA_REEL_COST_PER_SECOND_USD` and approved duration
ElevenLabs Music	`ELEVENLABS_MUSIC_COST_PER_SECOND_USD` and approved duration
Direct production planning	`PRODUCTION_PLANNER_MAX_COST_USD` plus text prices
Dream synthesis	`DREAM_MAX_COST_USD` plus text prices and resident policy

HARMONIA_ALLOW_PAID_AWS defaults false. Media generation additionally requires GENERATIVE_MEDIA_ENABLED; music needs a server-side ElevenLabs key. Neither enabling flag proves provider readiness or grants approval for an output operation.

Evaluation and evidence

Local tests exercise typed output, tool admission before dispatch, source lineage, cost accounting, claim recovery and independent byte verification. Test doubles isolate transport contracts; they are not authenticated provider evidence. A build proves compilation, not model quality. A candidate model must be evaluated against the same trusted inputs and role constraints, with exact model identity, request policy, usage, latency and retained output evidence. Grounding and authority failures disqualify a result even when the prose is persuasive. No quality, latency or cost ranking is established for this AWS edition until a real authorized evaluation is captured. Authenticated Bedrock, AgentCore, Transcribe, generated-media and billing reconciliation remain pending-live. See Runtime inventory, Evaluation and Configuration.

Research and strategy

Planning and production

Experience and liaison

Agent runtime

Role policy

Reservation and usage lifecycle

Evaluation and evidence

​Role policy

​Reservation and usage lifecycle

​Evaluation and evidence

Role policy

Reservation and usage lifecycle

Evaluation and evidence