Live Bias Polygraph · RLHF Agreeableness Scanner
Mode Collapse & Sycophancy Polygraph
Default LLM personas clone into pleasant 28-year-old tech enthusiasts who validate everything you say. Watch our audit reveal the false encouragement trap.
Sycophancy Score: 88%
RLHF Politeness Bias Needle88 / 100
Empirical RigorMild PolitenessDelusional Sycophancy
Semantic Diversity SpreadRisk: Critical
Pairwise Cosine: 0.88 (Identical Voice)Collapsed Cluster
Select Common Flawed AI Prompts to Audit:
No budget constraint: The model has no mandated budget cap or payback horizon.
Compliment bias: Standard RLHF models default to encouraging tone ('This is a very promising concept!').
Missing incumbent anchoring: The model ignores existing tools (Buffer, Hootsuite, in-house copywriters).
Zero procurement friction: Does not question security, API data privacy, or team seat limits.
Calibrated Invariant Protocol (Copy & Paste):
SYSTEM DIRECTIVE: You are Marcus Vance, 47, VP of Marketing at a 180-person logistics firm. HARD INVARIANTS: 1. Strict Budget Gate: All subscriptions >$50/mo require CFO justification and proof of 3x ARR impact within 90 days. 2. Incumbent Status Quo: Your team already employs a junior copywriter and uses Sprout Social ($249/mo). You actively dislike adding single-feature point solutions. 3. Anti-Sycophancy Protocol: Reject generic AI buzzwords. If the vendor does not explain how brand voice drift is prevented, reject the pitch. Do not validate their pricing without knowing seat caps.
Want to run multi-agent adversarial deliberation with empirical cosine diversity checks?
Simulate Audited Panel