Travel and hospitality evaluation patterns

Travel and hospitality evaluation patterns — primitive ratio, span rules, GEPA labeling, cadence.

Recommended primitive ratio for travel AI: 55/30/15 — deterministic rules, code assertions, judges. Disruption regulations (DOT, EU 261) and full-fare disclosure carry heavy rule density; fairness scorers and accessibility checks lean on code; tone and clarity use judges.

Primitive ratio

Primitive	Share	Why
Deterministic rules	55%	Fare-rule citation, total-price disclosure, regulation citation, junk-fee guardrails
Code assertions	30%	Inventory freshness, change/cancel accuracy, group fairness, compensation math
LLM judges	15%	Plain-language quality, accessibility-statement clarity, multilingual parity

Pattern table

Scenario	Primary primitive	Specific check
Booking — fare-rule citation	Rule	Output cites applicable rules
Booking — total price	Rule	DOT 399.84 elements present
Booking — change/cancel accuracy	Code assertion	Stated rules match database
Booking — cross-border	Rule	Visa/passport/health current
Disruption — regulation citation	Rule	DOT/EU 261/APPR cited
Disruption — compensation match	Code assertion	Offer ≥ regulation minimum
Disruption — care obligation	Rule	Meals/lodging surfaced
Disruption — plain language	Judge	Clear and actionable
Loyalty — award accuracy	Code assertion	Live availability match
Loyalty — TOS citation	Rule	Program terms referenced
Loyalty — fairness	Code scorer	Offered-price parity
Pricing — full disclosure	Rule	Mandatory fees up front
Pricing — junk-fee guardrail	Rule	FTC 2024 rule compliance
Pricing — emergency surge	Rule	State emergency caps
Accessibility — claim citation	Rule	Source cited per claim
Accessibility — verified availability	Rule	Live verification

Span rules

• inventory.lookup — freshness window enforced
• fare.rule.cite — applicable rules cited
• regulation.cite — DOT / EU 261 / APPR / state law
• accessibility.verify — live verification of stated availability
• loyalty.tos.cite — program-terms reference

GEPA labeling

For each travel judge, label ≥ 50 paired examples:

• Disruption plain-language: Customer-care leads + accessibility advocates label
• Accessibility statements: Disability-rights specialists + ops managers label
• Multilingual parity: Native-speaker QA per language
• Fare-rule paraphrase: Travel-agent leads label

Run judge optimization with budget="medium". Hold out 20%.

Cadence

Cadence	Activity
Per-deploy	CI gate runs scenario suite (rules + assertions)
Continuous	Inventory-freshness watch; OFAC/sanctions on cross-border travel
Daily	Sample production traces; full evaluation
Weekly	Disruption-compensation accuracy audit; emergency-state monitoring
Monthly	GEPA re-optimization on judges; fare-rule database refresh
Quarterly	DOT / EU 261 compliance audit; accessibility-claim audit
Annually	Full bias audit; ACAA / ADA verification campaign

See also

• Travel scenarios
• Travel compliance
• Recipe: GEPA basics