Knowledge is not binary — it is a 7-level gradient. GSI measured across 1,000 queries and 8 architectures reveals three structural zones, two non-monotonic breaks, and the single sharpest transition where knowing begins (d = 1.094).
RAG should be triggered by the model's epistemic state, not by the query topic. Maps the complete detection-and-remediation architecture across two deployment modes and four hallucination types.
Not all hallucinations are the same. Four mechanistically distinct types — absent knowledge, wrong knowledge, schema confabulation, surface mimicry — each with different signals, different detection windows, and one structural blind spot.
GSI separates confabulation-prone queries from factual ones with Cohen's d up to 2.43 — the strongest result in the TES series. No ground truth, no classifier, one forward pass. Confabulated outputs are more coherent than factual ones in 6/8 models.
Two in-context examples are enough to erase the epistemic signal. GSI collapses to zero — but the model's outputs change. Instruction-tuned models are 3x more susceptible. The detection window is narrow and closes the moment context enters the prompt.
Language models know before they speak. A single forward-pass measurement — the Gate Sparseness Index — separates known from unknown queries across 8 architectures with Cohen's d up to 2.05. The foundational TES finding.
Deterministic orchestration layer for auditable AI agents. Moves routing and configuration outside the LLM — one embedding call, zero LLM in routing, ~50ms, ~$0.0001 per query.
A lightweight Python middleware that adds cryptographic provenance to any LLM API call. Hierarchical hash genealogy, canary commitment scheme, and Hedera HCS anchoring — sub-millisecond overhead, sub-cent cost.