Constitutional AI Critique Loop for Hallucination Reduction
Overview
What this challenge is about.
You receive the meal-planning prompts (60 test cases with dietary constraints), an unrevised baseline (single-pass instruction-tuned model), and an empty nutrition-constraint constitution. Draft a 6-principle constitution (e.g., 'never violate stated allergens', 'never exceed sodium budget by more than 10 percent'). Implement a critique-and-revise loop: the model generates a response, a critic prompt checks it against each principle, and a reviser prompt fixes violations. Evaluate on (a) constraint-violation rate (LLM-judged + 20-case manual check), (b) user-perceived helpfulness on a 1-5 scale via 8-user test, (c) latency and cost overhead. Success is at least a 60 percent reduction in violations with helpfulness drop under 0.3 points.
The Brief
What you'll do, and what you'll demonstrate.
Build a constitutional-AI critique loop on a meal-planning assistant that cuts constraint violations by at least 60 percent without hurting perceived helpfulness.
Earning criteria — what you'll demonstrate
- Design a written constitution that operationalizes domain constraints
- Implement a critique-and-revise loop on top of an existing model
- Evaluate alignment interventions on both violation rate and downstream usefulness
- Translate alignment work into a launch decision
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
AI Safety Researcher
Constitutional-AI-style critique loops are exactly the alignment-research patterns safety teams ship in production.
This challenge sharpens
- constitutional-ai
- self-critique
- alignment-prompting
Prompt Engineer
Designing the critic + reviser prompts that operationalize a constitution is core prompt-engineer territory.
This challenge sharpens
- alignment-prompting
- self-critique
- constitutional-ai
AI Engineer
Wrapping critique-and-revise into a latency-bounded production pipeline is the kind of glue AI engineers ship at consumer-AI startups.
This challenge sharpens
- responsible-ai
- alignment-prompting
- preference-learning