Eval span attributes
When memledger's MAI evaluators run (deterministic, structural, RAGAS), each evaluator emits an OpenTelemetry span with a structured attribute set. Any OTEL-supported observability backend can ingest them; the Phoenix integration renders them as inline annotations on the source span.
What evaluators emit
| Attribute | Type | Set by | Use |
|---|---|---|---|
eval.outcome | "pass" | "fail" | all 3 tiers | Phoenix filter chips |
eval.score | float | all 3 tiers | rubric score |
eval.passed | bool | all 3 tiers | pass/fail boolean |
eval.record_ids | comma-separated | all 3 tiers | UI deep-link from eval span back to per-record inspector |
eval.judge_model | string | RAGAS (Tier 2) only | which LLM judged |
openinference.span.kind | "EVALUATOR" | all 3 tiers | Phoenix categorization |
Configuring evaluator runs
The 3 evaluator tiers (deterministic, structural, RAGAS) all consume the same canonical MAI rubric. The runner is your choice; the rubric is constant. See The evaluator tiers.
Sample annotation in Phoenix
Tier 1 — deterministic evaluator annotation rendered in Phoenix.
Tier 2 — RAGAS LLM-judge evaluator annotation rendered in Phoenix.
RAGAS evaluator span detail — judge model, score, rubric criteria.
The full attribute inventory + Phoenix dashboard template ships in a future release.