Data Oracle · Paired data & active learning

# Core capabilities

# 1. Paired Data Generation

Record both AI predictions and deterministic gold labels for the same design state. When an Agent calls an AI model via Model Router, Oracle asynchronously runs the deterministic engine on the same state to produce ground truth. Each pair includes input (design-state features), ai_prediction, gold_result, and delta (error analysis).

# 2. Active Learning

Select the most informative samples for deterministic labeling. Three strategies: uncertainty (lowest AI confidence), diversity (underrepresented design regions), hybrid (uncertainty + diversity). Active learning minimizes deterministic engine cost—gold labels only for the most valuable samples.

# 3. Lineage Tracking

Record design source, tool version, process conditions, and scenario parameters for each sample. Full lineage: design source (design_id + commit) → tool version (tool + version + config_hash) → process (PDK node + corner + temperature) → scenario (utilization, frequency target, etc.). Dataset provenance is fully transparent.

# 4. Applicability

Tag applicability for each sample: design_family (riscv/vliw/soc/...), PDK (sky130/nangate45/asap7/...), scenario_type (congestion_fix/timing_opt/power_reduction/...). Model Router uses applicability tags for domain checks—domains not covered by training data trigger OOD detection.

# 5. Quality Gate

Data must pass coverage and correctness gates to enter the training set. Coverage gate: ensure new design/scenario regions are covered (avoid redundant homogeneity). Correctness gate: gold labels must come from trusted deterministic engine versions (stale versions marked degraded). Data passing both gates receives the certified label.

# Agent invocation

Data Oracle provides two core operations—paired data generation and active learning sample selection.

from ieda.oracle import DataOracle

oracle = DataOracle("http://localhost:9130")

# Generate paired data: record AI prediction + gold result
samples = oracle.generate(
    design_ref="snap_7f3a",
    tools=["iPL", "iSTA"],
    scenarios=["congestion_fix", "timing_opt"],
)
print(len(samples))             # 42 paired samples

# Active learning: select most informative samples
selected = oracle.select(
    selection_strategy="uncertainty",
    top_k=100,
    model_family="congestion_nn",
)

# Check data lineage
lineage = oracle.get_lineage(sample_id="pair_7f3a_042")
print(lineage.design_source)    # "snap_7f3a (commit a1b2c3)"
print(lineage.tool_version)    # "iPL v2.1.0 (config hash: f3a9)"

// Generate paired data
{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "oracle.generate",
    "arguments": {
      "design_ref": "snap_7f3a",
      "tools": ["iPL", "iSTA"],
      "scenarios": ["congestion_fix", "timing_opt"]
    }
  }
}

// Active learning selection
{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "oracle.select",
    "arguments": {
      "selection_strategy": "uncertainty",
      "top_k": 100,
      "model_family": "congestion_nn"
    }
  }
}

# Generate paired data
oracle generate -design snap_7f3a \
  -tools {iPL iSTA} \
  -scenarios {congestion_fix timing_opt}

# Active learning: select top-100 most uncertain
oracle select -strategy uncertainty \
  -top_k 100 -model_family congestion_nn

# Check sample lineage
oracle lineage -sample pair_7f3a_042

# Input/output contract

OracleRequest

OracleResult

# Related resources