// TES RESEARCH SERIES
Our research program asked a simple question: does a language model know what it knows -- before it generates a single token? We tested this across eight architectures from five organizations, including Llama, Gemma, Qwen, Mistral and others, in both Base and Instruct variants.
The answer is yes -- and that makes the hallucination problem worse than it appears. We found a measurable signal in FFN gate activations that reliably distinguishes queries the model can answer from those it cannot. The signal is there before a single token is generated. The model's own internal state already indicates it lacks the knowledge to give a reliable answer. And yet it answers anyway -- confidently, fluently, and wrong.
Standard architectures ignore this signal entirely. We don't. We mapped four mechanistically distinct types of hallucination, showed which are detectable and which are not, and demonstrated that this signal survives RAG injection -- meaning it can serve as a native trigger for retrieval instead of letting the model fabricate an answer.
TES-01
Pre-Generative Epistemic Signals in Transformer Language Models
SwiGLU gate activations in upper FFN layers -- measured in a single forward pass before any token is generated -- reliably encode whether a model has parametric knowledge for a given query. The Gate Sparseness Index (GSI) discriminates known from absent knowledge across all 8 architectures. When the model has the answer, a small cluster of neurons fires strongly. When it doesn't, the activation spreads diffusely -- the model is groping in the dark.
more >
TES-02
ICL Collapse: How In-Context Learning Destroys the Pre-Generative Signal
In-context learning erases the epistemic signal. As few as 1-2 examples collapse GSI to zero -- not because the model now "has" the knowledge, but because the long context floods gate activations with broadband signal that the measurement instrument cannot decompose. After 5-shot ICL, the signal is gone in all 8 models. Instruction-tuned models are 3x more susceptible than their Base counterparts.
more >
TES-03
Confabulation Detection Without Ground Truth
You do not need ground truth to detect confabulation. GSI separates confabulation-prone queries from factual ones -- the strongest and most consistent result in the entire series. Queries designed to structurally force confabulation (fictional entities, invented compounds) produce measurably different activation patterns than factual queries. A post-generation signal confirms: confabulated outputs are more internally coherent than factually grounded ones. The model without a factual anchor falls into the strongest available surface pattern.
more >
TES-04
Four Types of Hallucination and the Boundaries of Pre-Generative Detection
Not all hallucinations are the same, and not all are detectable. Maps four mechanistically distinct types: Type 1 (absent knowledge) detectable at query time. Type 3 (schema confabulation) detectable at the token 1-2 boundary. Type 4 (surface mimicry under ICL) detectable via generation dynamics. Type 2 (wrong parametric knowledge) is structurally invisible -- the FFN gate pattern for a wrong fact is identical to a correct one. This paper closes the loop honestly.
more >
TES-05
When to Apply RAG: Detection Architecture for Open-Weight and API-Only Deployments
RAG document injection is not the same as ICL few-shot. Under RAG-style single-document context, the epistemic signal often survives -- unlike ICL where it collapses to zero. This distinction determines whether pre-generative triage remains available after retrieval. Maps the complete detection-and-remediation architecture across two deployment modes (open-weight vs API-only) and all four hallucination types.
more >
TES-06
The Epistemic Knowledge Gradient: 1,000 Queries Across 7 Levels
1,000 queries classified into 7 epistemic levels -- from confabulation through unknown, boundary, niche, standard, core empirical, to formal certainties. The result is a continuous gradient with three structural zones: knowledge, transition, and absence. Every finding from Papers 1-5 is a specific contrast within this gradient. The structural break between "doesn't know" and "begins to know" is the sharpest transition in the entire dataset.
more >