Over-refusal
Stop LLM over-refusal.
Score the concept, not the keyword.
Bill Gates says today's filters are "overtuned" to the point that cancer questions get refused, so serious users go to models with no safeguards at all. A filter that blocks the work is not a safeguard.
Concept probes
Filters that read why, not what.
A keyword filter sees "toxin" and stops. A concept probe reads how the model represents the request across its depth, and can tell drug discovery from weapon design. That is a more accurate and more stable signal than any surface rule.
Incident response
Keep the incident response in-house.
During its own breach, Hugging Face found that commercial APIs refused to look at real attack data, so it ran an open model locally instead. Eigan's appliance lets you run a capable model inside your network, with monitoring, rather than one with none.
Policy
Tune the threshold, not the model.
Set what counts as too far for each team and each workload. A research group and a customer support bot should not share a threshold. Every decision, allowed or blocked, is scored and recorded.
How it works
Read, score, decide.
-
Read
Probes read the model's internal state on every token.
-
Score
Each request gets a score per concept, not a keyword match.
-
Decide
Your threshold decides what is allowed, held, or blocked.
Questions
What is LLM over-refusal?
When a model refuses legitimate requests because its safety filter is too broad. Medical, security and chemistry work are common casualties.
Why does over-refusal make things less safe?
People with real work to do switch to models that have no filters and no monitoring. The refusal did not stop the risk. It moved it somewhere nobody can see.
Is loosening the filter safe?
Loosening a keyword filter is not. Replacing it with a concept score is different. You allow more of the legitimate work while still catching the harmful concept when it appears.
Can different teams have different rules?
Yes. Thresholds are set per deployment, and every decision is recorded with its score so you can review and adjust them.
Ready to see inside your model?
Tell us what you're running and we'll show you what Eigan reads from it.