Session 15: budget aware routing for irregular travel operations

[kumariAstha]

Session 15: Budget-Aware Routing for Irregular Travel Operations

What this PR adds

• Workload: 15 airline disruption recovery tasks (5 trivial / 5 moderate / 5 hard) in proofs/tasks/travel_ops.jsonl
• Ladder: cerebras/zai-glm-4.7 (economy) → gemini/gemini-3.1-flash-lite (standard) → openrouter/openai/gpt-4.1 (frontier)
• Budget policy: $0.03 ceiling, downgrade at 50% pressure, refuse at 90%, max 40 calls/run, max 5 calls/node
• Adversarial proof: proofs/p8_adversarial_budget.py — ladder climb, budget exhaustion, per-node ceiling, provider failure gap

Key findings

Strategy	Resolved	Cost/Resolved Task
A: Always frontier	12/15 (80%)	$0.000653
B: Always cheapest	2/15 (13%)	$0.000172
C: Budget-aware cascade	13/15 (87%)	$0.000267

The cascade resolves the most tasks at 59% lower cost per resolved task than always-frontier. It also resolves one task (t15) that frontier fails — the standard tier produces more focused output that the judge scores higher.

Break-even resolution rate: 26.2%. Signature failure mode not observed — economy model was too weak (13% resolution) for the cheap-model trap to appear.

Where the policy fails

• t02, t09: failed across all strategies. Rate-limiting + empty answer extraction on bare calls. Same tasks succeed through the full agent runtime.
• t15: frontier fails, cascade succeeds — verbose frontier output scored lower than concise standard output.
• Known gap: provider-side token consumption on failed calls is not metered.

Evidence

Full analysis in the README evidence section. Proof outputs in proofs/out/ and evidence/.

Reproduce

All commands in the README. Requires API keys for Cerebras, Gemini, OpenRouter, Groq, Nvidia (all free tier).