Defense in depth · v1.0
Anti-hallucination by structure.
Let's be honest: the NOE alone does not eliminate hallucinations. It is one of five layers. The strength of Crisol lies not in a magic filter at the end, but in a defense in depth where a false claim must cross five independent barriers — and the largest of them acts before the model reasons about anything.
Honesty is the argument.
No system can promise zero hallucinations with a single mechanism, and anyone who promises it is lying. Crisol does not sell a perfect shield: it sells a design where each layer covers what the previous one let through. The probability of a falsehood crossing all five is not the sum of the errors — it is their product.
That is why the sentence governing this entire page is the simplest and the strongest:
The expert never saw a falsehood → it does not repeat it.
How the defense is distributed
Five layers, five weights, one objective.
The percentages are not rhetoric: they are the relative contribution of each layer to the total robustness. The corpus weighs four times more than the invariants, and that has a strategic consequence you will see at the end.
Explore each layer: mouse, tab or tap.
Distilled corpus
The expert never saw a falsehood.
Each expert is trained on a distilled, verified and cleaned corpus. If the expert never encountered a false claim during its formation, it has nothing to repeat. It is the largest line of defense and it does not live in the model: it lives in what the model learns.
The NOE alone is just 10%. The distilled corpus (40%) is what pays off most. It is not a filter at the end: it is defense in depth — and the largest barrier acts before the model reasons about anything.
Before, during and after the reasoning.
Defense in depth is not stacking filters: it is distributing them in time. A falsehood that dodges prevention runs into constraint; one that dodges constraint runs into verification. Three windows, none of them redundant.
Before the forward
- · Distilled, verified corpus (40%)
- · NAR pre-trained on axiomatic axes
Prevented at the source. The falsehood is not filtered out — it simply never enters the system.
During the forward
- · Top-k routing with QSE (20%)
- · Axiomatic NAR constraining inferences (15%)
- · NOE checking invariants (10%)
Constrained at generation time. The space of possible outputs is already bounded by structure.
After the forward
- · VerificationSubagent over the output (15%)
Caught at the end. The final customs gate reviews what survived the earlier layers.
Strategic implication
Six months polishing the corpus pay off more than a thousand new invariants.
If the corpus delivers 40% of the robustness and the invariants 10%, the highest-return lever is obvious: invest in clean data, not in patches at the end. The quality of the knowledge the expert absorbs weighs four times more than the layer that checks the output.
It is the difference between building on solid foundations and adding props once the walls crack. Crisol chooses the foundations.
The next level
An answer that does not hallucinate is only the beginning. One that reasons causally is the goal.
The structure that prevents falsehoods is the same one that lets Crisol simulate counterfactuals, check causes and reason about the world. That is where the real engine begins.
See causal reasoning →