Hallucination detection
in/guard/out vs Galileo Luna
Galileo’s Luna family – small fine-tuned evaluation models – makes continuous quality scoring affordable: low latency, a fraction of judge-model cost, wired into Galileo’s evaluation and observability platform. The scoring is real progress. The gap is what happens after the score: a metric dashboard doesn’t repair the answer, block the payment, or stop the run. That’s the layer in/guard/out occupies.
Side by side
| Capability | Galileo Luna | in/guard/out |
|---|---|---|
| What it is | Small eval models + a metrics/observability platform | A runtime control plane – checks with enforcement authority |
| First line of defense | A small-model inference per metric | Zero-token deterministic checks before any model spends |
| On a bad score | Metric drops; alert fires | Repair, block, or flag – per policy, before the response ships |
| Scope | Hallucination/quality metrics on LLM outputs | Output checks plus PII round-trip, format repair, injection, tool policy, budgets |
| Agent runs | Scored as traces | Governed: per-run hard budgets, action grounding, taint, sequence policies |
| Integration | SDK instrumentation into the platform | One base-URL change; no app code |
| Cost accounting | Platform metrics | Every guard call priced into the request and the run – the guardrail tax is a number |
| Audit output | Dashboards | Per-stage evidence + the run graph of enforced decisions |
Scores inform. Enforcement decides.
Luna answers “how good is this output?” quickly enough to ask on every request. in/guard/out asks a different question – “does this ship?” – and is built to act on the answer: deterministic grounding catches the numeric class outright, claim audit ties assertions to sources, FIX mode repairs what has a safe fix, and the agent stack extends the same authority to tool calls and budgets. If your team already scores with Luna, nothing conflicts: the wire enforces while the platform trends.
Frequently asked questions
Luna is fast and cheap – isn’t that enough for guardrailing?
Fast scoring solves half the problem. The score still needs an enforcement point with authority over the response and – for agents – the actions. in/guard/out is that point: it starts with checks that cost nothing at all, and its verdicts end in repair or block, not a metric.
Do the platforms conflict?
No. Galileo instruments your app for evaluation; in/guard/out sits on the wire for enforcement. Production catches from the proxy make strong labeled cases for whatever evaluation platform you trend quality in.
What does enforcement add for agents specifically?
The failures that hurt are actions, not sentences: the ungrounded payee, the loop that burns budget, screened PII exiting via a tool argument. Those need run-scoped enforcement – budgets, grounding, taint, ordering – which no output-scoring metric provides.
See it on your traffic first.
FIX mode observes without blocking, so you can measure what in/guard/out catches before enforcing anything. We are running a limited demo - sign up and we will get you in as soon as we can.