[evaluations.generate_job_content_eval]
type = "inference"
function_name = "generate_job_content"

[evaluations.generate_job_content_eval.evaluators.evaluate_jd_generation]
type = "llm_judge"
output_type = "float"
input_format = "messages"
optimize = "max"


[evaluations.generate_job_content_eval.evaluators.evaluate_jd_generation.variants.eval_gpt_audio_mini]
type = "chat_completion"
model = "gpt-audio-mini"
system_instructions = "evaluations/generate_job_content/evaluate_jd_generation/system_instructions.txt"
json_mode = "off"
active = true

[evaluations.generate_job_content_eval.evaluators.evaluate_jd_generation.variants.eval_gemini_2_5_flash_lite]
type = "chat_completion"
model = "gemini-2-5-flash-lite"
system_instructions = "evaluations/generate_job_content/evaluate_jd_generation/system_instructions.txt"
json_mode = "strict"

[evaluations.generate_job_content_eval.evaluators.evaluate_jd_generation.variants.eval_gemini_3_1_flash_lite_preview]
type = "chat_completion"
model = "gemini-3-1-flash-lite-preview"
system_instructions = "evaluations/generate_job_content/evaluate_jd_generation/system_instructions.txt"
json_mode = "strict"

 The bug: When variants don't have an explicit active field, get_weight(None) returns None. In legacy_from_variants_map, weight=None variants are added to fallback_variants — not
  excluded. The config validation only checks for "exactly 1 nonzero-weight variant", which passes (since the None-weight gemini variants are treated as 0 for counting). But at
  runtime, if eval_gpt_audio_mini fails, the sampling exhausts its total_weight and falls back to the first gemini variant in the fallback list.

  Root cause in code: experimentation/legacy.rs:78 — weight=None → fallback_variants.push(...). The evaluations validation at evaluations/mod.rs:594 doesn't catch this case because
   it only counts explicit nonzero weights.