Intent & routing

The first layer maps a query to an intent, a difficulty in [0,1], and a compute budget.

Quick example

from ebrm_system.intent import RuleBasedClassifier

clf = RuleBasedClassifier()
pred = clf.classify("Solve: 3x + 7 = 22")

print(pred.intent)                       # Intent.MATH_REASONING
print(pred.suggested_langevin_steps)     # scales with difficulty
print(pred.suggested_trace_count)        # K parallel traces

Budget policy

• steps: 50 → 2000 (linear in difficulty)
• restarts: 1 → 10
• trace_count: 1 → 16

Swap in a neural classifier

Any object implementing Classifier.classify(query: str) -> IntentPrediction works.