openspec/changes/fix-traced-bugs/specs/domain-classification/spec.md

MODIFIED Requirements

Requirement: Telecom domain classification covers SIM and prepaid recharge scenarios

The DomainClassifier SHALL classify complaints about SIM card purchase, SIM activation failure, prepaid mobile recharge not reflected, and mobile number portability as domain telecom with low_confidence=False when a trained checkpoint exists.

The synthetic supplement dataset SHALL include at least 10 sentence templates specifically covering SIM activation failure and prepaid recharge failure scenarios in addition to the existing 20 templates.

The supplement_per_class training parameter SHALL default to 5,000 (increased from 2,000) to reduce the imbalance ratio between CFPB-sourced banking rows and synthetic telecom rows.

Scenario: BSNL SIM activation complaint classified as telecom

• GIVEN the complaint text is "I bought a new sim card from BSNL for rs 450 on 11th january 2026. but sim card is not yet activated."
• WHEN classify(text) is called with a retrained checkpoint
• THEN result.domain == "telecom" and result.low_confidence == False

Scenario: Airtel prepaid recharge not reflected

• GIVEN the complaint text contains "recharged for 5450 rupee in airtel for 2 years" and "recharge was successful but it's not reflected"
• WHEN classify(text) is called with a retrained checkpoint
• THEN result.domain == "telecom"

Scenario: Banking complaint still classified correctly

• GIVEN the complaint text is "Axis bank charged 1000 rupee for late fee of not paying loan"
• WHEN classify(text) is called
• THEN result.domain == "banking"

Scenario: Low-confidence result triggers agent clarification

• GIVEN the classifier returns low_confidence=True for any domain
• WHEN the agent processes the classification result
• THEN the agent asks the user a clarifying domain question before proceeding with entity extraction