See what the model is doing while it does it.
Spectre is an inference analysis sidecar. It sits beside a language model, reads the concepts present in each request and response, applies the policy you define, and records the result. It is a sensor, not a guardrail: it measures and reports, and enforcement is a policy you set.
PreviewSpectre is in active development. This page outlines what it does and what is planned. Capabilities marked as planned are not yet available, and no dates are given.
Four steps on every request.
Spectre runs in your own infrastructure, in line with inference, so the analysis covers the traffic itself and not a sample of it.
Intercept
Every request and response passes through the sidecar before it reaches its destination.
Analyze
The analysis pipeline scores the human-meaningful concepts present, such as deception, threat, data exfiltration and authorization.
Apply policy
Scores are checked against rules you define. A rule can alert, log, or hold a request for review.
Record
Each decision is written as a signed audit record suitable for a transparency log.
Two ways to read a model.
Open-weight models. Spectre attaches to the model’s layers and reads concepts from its internal state as it generates. This needs direct access to the model, so the model is hosted inside your boundary.
Models reached only through an API. Spectre falls back to behavioural analysis of the inputs and outputs. It cannot read internal state in this mode, and the two modes should not be compared as equivalent.
Spectre reads text. It does not analyze images, audio or video.
Built on published concept-geometry research.
Spectre applies the method described in Eigan’s research on how concepts form across a model’s depth and how to extract stable probes from them. The papers, code and data are open: see the research page.
A concept score is a measurement of what a model’s internal state contains, not a statement of intent. A high score is evidence for a person or a rule to weigh. Thresholds are calibrated for each model and each concept. This page makes no claims about detection accuracy; Eigan will publish evaluation methods and results as they are verified.
Interactive demo capabilities are planned.
A limited preview is available on request. The following are planned for the public demo. None is available yet, and none has a date.