Model Router · Intelligent model routing

# Core capabilities

# 1. Model Registry

Register full metadata for every AI model: type (NN/RL/GBDT/GNN/LLM), version, applicability (design family, PDK node, scenario type), training scope, known limits. Each model has a unique model_tag; routing uses Registry calibration data, not hard-coded rules.

# 2. Calibration

Learn calibrated probabilities and error correlation matrices per model. Track systematic bias patterns—not just accuracy (e.g., a model always overestimates congestion on small designs, another is optimistic on 7nm data). Correlation matrices guide ensemble: independent-error models combine better than co-linear ones.

# 3. OOD Detection

Detect whether inputs lie in the model's applicability domain. Dual check: Mahalanobis distance to training centroid (default threshold 3.0) and density-based blank regions mark OOD. OOD forces fallback to deterministic engines—no risky predictions.

# 4. F0-F4 Routing

Route by fidelity tier—F0 (estimate NN, fast coarse prediction) → F2 (in-design hybrid, mid-precision assist) → F4 (correlated deterministic, high-precision execution). Agents use F0 early to filter candidates, F2 for refinement, F4 deterministic tools for gold results.

# 5. Ensemble

Ensemble considers error correlation, not simple voting. Independent models (~0 correlation) outperform co-linear ones. Router prefers high-independence combinations to maximize ensemble gain.

# Agent invocation

Model Router exposes a unified inference entry—Agents describe requirements without knowing which specific model runs.

from ieda.router import ModelRouter

router = ModelRouter("http://localhost:9120")

# Register a new model
router.register(
    model_tag="congestion_nn_v3.1",
    model_type="NN",
    applicable_domain={
        "design_family": ["riscv", "soc"],
        "pdk": ["sky130", "nangate45"],
    },
)

# Route inference request
result = router.route(
    model_family="congestion_nn",
    input_features=design_features,
    required_fidelity="F2",
    fallback_policy="deterministic",
)

print(result.selected_model)      # "congestion_nn_v3.1"
print(result.confidence)         # 0.87
print(result.ood_flag)           # False
print(result.fallback_triggered) # False

// Route inference request
{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "router.route",
    "arguments": {
      "model_family": "congestion_nn",
      "required_fidelity": "F2",
      "fallback_policy": "deterministic"
    }
  }
}

// Response
{
  "selected_model": "congestion_nn_v3.1",
  "confidence": 0.87,
  "calibration_metadata": {"ece": 0.03, "bias": 0.01},
  "ood_flag": false,
  "fallback_triggered": false
}

# Register model
router register -tag congestion_nn_v3.1 \
  -type NN -domain {riscv soc} -pdk {sky130 nangate45}

# Route with fidelity F2
router route -family congestion_nn -fidelity F2 \
  -fallback deterministic

# Force recalibration
router calibrate -family congestion_nn

# Input/output contract

RouteRequest

RouteResult

# Related resources