Evidence · first-party tested

When the prompt asks for angry and guarded, the tool can instead produce a neutral, emotionless expression, flattening the intended mood.

✗ FailedGemini
Output — unretouched
tool output
The observation

When the prompt asks for angry and guarded, the tool can instead produce a neutral, emotionless expression, flattening the intended mood.

Criterion: Expression accuracy · Scenario: Three-quarter face portrait reference (group: reference-portrait-variations)
Query this
get_evidence({
  tool: "gemini",
  scenario: "reference-portrait-variations"
})
MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Real outputs, no retouching · every cell queryable via API & MCP · aidemos.com