lookadev/cognitive-triage-benchmark

Decoupled Cognitive Triage: Evaluating TypeSafe System 1 (Jev) Against Google Gemini 3.8 Flash

This dataset contains 1,200 frozen, labeled real-world business transactions evaluated across dual-process AI architectures (TypeSafe System 1 / Jev + Google Gemini 3.8 Flash).

• Technical Report: LOOKADEV-TR-2026-004 (Preprint)
• GitHub Repository: https://github.com/lucasmartins-ai/cognitive-triage-benchmark
• Executive Engineering Report: https://lookadev.com/en/devlog/decoupled-cognitive-triage-jev-vs-gemini

Key Empirical Findings

• 58.2% of Gemini 3.8 Flash calls eliminated (resolved locally in 283ms at $0.075/1K).
• Median system latency dropped by 97.9% (16,240ms down to 333ms).
• Macro F1 retained at 102.25% of baseline (75.7% vs 74.0%, McNemar p = 0.624).
• Direct cost reduction of 58.0%, saving $15,761.26 per 1M decisions.

Categories

1. sales_high_intent
2. support_and_inquiry
3. spam_and_solicitation
4. security_and_injection
5. enterprise_bespoke
6. ambiguous_intent
7. compliance_and_edge

Reproduction

Clone the open repository:

git clone https://github.com/lucasmartins-ai/cognitive-triage-benchmark.git