model-tierconditionalweight 2adapter 15 of 17

Model Tier.

The Model Tier adapter classifies model subjects by parameter count into capability tiers. Larger tiers score higher, and because the tier is a stable property of the model rather than observed behavior, it applies only where model metadata is available and carries a substantial weight.

Counts only when its signals are present on the subject.

What it measures.

Parameter count

Published parameter count from model metadata.

Tier assignment

Mapping of parameter count onto the informative tier scale.

Component keys.

model-tier.*

Component family covering the assigned capability tier.

When it counts.

Conditional — applies only to model subjects with known parameter metadata.

When signals are missing.

Models without parameter metadata, and non-model subjects, omit this component.

How it contributes.

{
  "component": "model-tier.*",
  "scale": "0-100",
  "mode": "conditional",
  "weight": 2,
  "contribution": "component x 2 / total weight"
}

The adapter’s components are averaged into the composite as a weighted mean: every contributing adapter’s component is multiplied by its weight, summed, and divided by the total weight of contributing adapters. The result is published as trustScores.total with a config version and timestamp so anyone can recompute it.

Questions about Model Tier.

How does the Model Tier adapter assign a score?

It maps the model’s published parameter count onto an informative capability tier and normalizes the tier into the model-tier.* component at weight 2.

Why does tier matter if behavior is measured elsewhere?

Tier is stable, hard-to-fake context that anchors the behavioral components; it is weighted substantially but still counts only for models with known metadata.