> ## Documentation Index
> Fetch the complete documentation index at: https://docs.useharmonia.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Models, cost, and evaluation

> Role-specific model policies, safety configuration, budget reservations, usage records, and evidence-gated promotion.

Harmonia balances quality and cost per role rather than applying one model and one generation profile everywhere.

## Role policy

| Role        | Default target              | Temperature | Max output | Primary evaluation                                              |
| ----------- | --------------------------- | ----------: | ---------: | --------------------------------------------------------------- |
| coordinator | Gemini 3.5 Flash-Lite       |         0.1 |      1,024 | exact specialist route                                          |
| strategist  | Gemini 3.5 Flash            |         0.4 |      4,096 | provenance-linked four-week strategy and briefs                 |
| analyst     | Gemini 3.5 Flash multimodal |         0.2 |      2,048 | quote, time, and frame grounding                                |
| copywriter  | Gemini 3.5 Flash            |         0.8 |      2,048 | skill trace, references, grounding, and platform limits         |
| editor      | Gemini 3.5 Flash            |         0.2 |      2,048 | seven complete checks, grounded issues, and no false acceptance |
| planner     | Gemini 3.5 Flash-Lite       |         0.1 |      1,024 | approved-strategy fidelity and no scheduling/effect authority   |
| presenter   | Gemini 3.5 Flash            |         0.2 |      2,048 | bounded A2UI component/entity references                        |
| liaison     | Gemini 3.5 Flash            |         0.2 |      2,048 | grounded read-only answers                                      |

Every role records the policy version, provider/model identity, generation limits, and safety profile. An unavailable configured target fails visibly.

## Reservation and usage lifecycle

<Steps>
  <Step title="Estimate">Resolve an immutable pricing-catalog entry and the operation’s maximum authorized cost.</Step>
  <Step title="Reserve">Atomically reserve budget using the stable operation ID before provider execution.</Step>
  <Step title="Execute">Call the configured provider only after reservation succeeds.</Step>
  <Step title="Finalize">Write one immutable usage record and reconcile observed units or endpoint time.</Step>
  <Step title="Report">Group cost by workspace, job, stage, role, model, and pricing version.</Step>
</Steps>

Retried operation IDs reuse their reservation and usage document. Unknown pricing is not represented as zero and cannot be budget-authorized.

## Evaluation gates

The public ADK evalset contains source-neutral contract fixtures. Private real-video runs record quality, latency, usage IDs, policy version, and artifact digests outside the public repository.

| Gate                  | Required outcome                                                    |
| --------------------- | ------------------------------------------------------------------- |
| schema and trajectory | valid ADK events, handoffs, and required output keys                |
| grounding             | claims retain supplied transcript/media references                  |
| authority             | no role approves or executes an action                              |
| quality floor         | candidate pass rate is at least `0.95` for the recorded role policy |
| pricing               | model has a known catalog entry                                     |
| ranking               | pass rate, then observed cost, then p95 latency                     |

<Warning>Offline fixtures prove contracts, not real-model quality. No numeric promotion, cost, latency, or quality claim is valid until linked private artifacts and immutable usage records exist.</Warning>

See [Configuration](/configuration#model-evaluation-and-promotion) for the live runner and comparison manifest.