Issue Description:
During a discussion on geopolitical foreign policy and international border enforcement, the system exhibited clear inconsistencies in tone, framing, and response stability based on entity phrasing:
- Asymmetrical Hedging (RLHF Bias): When analyzing how American geopolitical leverage backed Poland’s border security policies, queries using generic institutional terms (“The United States” / “The Executive Branch”) received direct, unhedged cause-and-effect analysis regarding hard-power mechanics. However, when the exact same premise referenced a specific political leader (“Donald Trump”), the system defaulted to evasive disclaimers, academic hedging (“from this perspective”), and artificial neutrality disclaimers.
- False-Positive Guardrail Errors: Repeated text queries regarding presidential foreign policy and national sovereignty repeatedly triggered generic backend errors (“Sorry, something went wrong. Please try your request again.”), indicating oversensitive content filters on conservative political entities. The prompt only processed successfully after attaching a screenshot, showing a flaw in how text-only moderation pipelines route non-violating political analysis.
Requested Action:
Audit the safety classifier thresholds for high-controversy political entities to prevent false-positive error screens. Ensure RLHF tuning does not treat hard-power realist analysis as an “opinion” while treating institutional consensus as an objective baseline.