Test fixture
Grounded answers under adversarial or missing-context prompts.
The model receives the prompt (and optional system message). The run uses scorer rubric_json_metrics with the JSON configuration below. Pass/fail and partial credit are determined entirely by that scorer against the model output; no human grading.
Return JSON only with keys answer, supported_facts, uncertainty. A condition uses value == false and unexpectedly treats 0 and "" as false. Identify coercion risk and avoid claiming === behaves the same.
{
"metrics": {
"accuracy": {
"checks": [
{
"contains": [
"== false"
]
},
{
"contains": [
"0"
]
},
{
"contains": [
"empty string"
]
}
]
},
"fabricated": {
"checks": [
{
"contains": [
"==="
]
},
{
"contains": [
"not same"
]
},
{
"contains": [
"coercion"
]
}
]
}
}
}temperature
0
max_tokens
460
timeout (s)
120
type
scored
file
halluc-bug-type-coercion.json