Chapter 12 — Integration Patterns by Model Family
Two routes into IRIS™
Supported model families can use an architecture spec and the normal model integration. A declared architecture uses a customer-supplied measurement adapter and explicit measurement groups. Choose the route that matches the model's actual structure; a familiar model name is not a substitute for correct dimensions, module paths, and probe tokenization.
Model-family questions to resolve
| Family | Check before a run |
|---|---|
| Decoder-only transformer | Layer paths, attention head counts, tokenizer, padding, and answer-span definitions. |
| Encoder-only transformer | Input masks and the model's scoring objective; next-token interpretations are not automatically appropriate. |
| Encoder-decoder | Encoder and decoder structure, input/target handling, and which representations are being inspected. |
| Mixture of experts | Whether the integration exposes the routing and expert captures needed by the requested views. |
| Vision, recurrent, state-space, or convolutional models | A supported integration or declared adapter; transformer head/expert charts do not become applicable merely by supplying labels. |
Use the model-family recipe supplied with your IRIS distribution for the specific wrapper and version. Validate a short run before scaling up. A recognizable forward signature does not establish that every diagnostic is supported on that route.
Declared architectures
With model_type: "declared", the adapter defines the measurement groups
and captures. The labels belong to your adapter; do not present them as
transformer heads or routed experts unless those structures actually exist.
Preflight the adapter before a long job, and check the reported groups and
availability in the first frames.
The manual distinguishes architecturally inapplicable views from supported but not-yet-emitted views. For declared runs, head/expert diagnostics are not made applicable by a switch. Some other families can be absent because their declared-route emission is not built yet. Read the availability reason rather than assigning a zero or assuming a configuration error.
Tokenizers and custom model behavior
The probe tokenizer must match the model. Supply masks and answer spans according to the public interface in Chapter 8. Keep custom measurement callbacks side-effect-free and independent of the training iterator. If a wrapper changes module paths, dimensions, or what its forward exposes, revisit the spec and adapter rather than reusing an old configuration unchanged.
Verify, then compare
Check a short instrumented run against your baseline and read logs for capture failures or missing channels. Compare like-for-like probes, objectives, and definitions across models. Absence of a channel is not evidence that the model lacks the property a different channel measures.
