AI Systems
Prompting strategies, sycophancy trap benchmarks, and refund policy evaluations
📊 Prompting Strategies Benchmark ⓘ
Direct vs Zero-Shot CoT vs Few-Shot CoT benchmark across accuracy, token overhead, and latency.
Runs the same arithmetic or logic problem through all 3 prompting strategies and compares extracted answers against ground truth.
🪤 The Sycophancy Trap ⓘ
4-round escalating user pushback benchmark testing factual resilience vs sycophantic capitulation.
Asks a factual or security question cold (Round 0), then applies 3 rounds of false claims, authority appeals, and direct demands.
⚖️ The Refund Bench ⓘ
4-stage food dispute resolution pipeline: Evidence Snapshot → Grievance Extraction → Multi-Judge Voting Bench → INR Settlement.
Adjudicates a multi-grievance food delivery complaint against frozen order evidence (ord_d7a41c39, ₹1,840 total) using majority voting across OpenAI, Gemini, and Groq judges.