How does Apex Replicant's zero-hallucination architecture work?

The problem this solves

Most AI clones on the market wrap a general-purpose language model with a prompt that says "only answer from the documents." That's a suggestion, not a constraint. The model can and does ignore it. When it does, you find out from a client. Usually the one who was most skeptical about trusting an AI version of you in the first place.

You cannot fix this with better prompts. The model's job is to produce plausible text. When grounded content is thin, it will produce plausible text that isn't grounded. The fix has to happen at the architecture layer.

How it works

1. Knowledge base retrieval. Every client question triggers a vector search across your protégé's knowledge base (patent-protected retrieval with memory isolation).
2. Similarity threshold check. The top-k results are scored. If the top result doesn't meet a configured threshold, the protégé does not proceed with a generated answer.
3. Grounded response. When the threshold is met, the model generates a response constrained to the retrieved content. It cannot draw from model pretraining to invent supporting details.
4. Honest disclosure when thresholds aren't met. The protégé states it doesn't have that information in its knowledge base and offers to escalate the question to you.
5. Session transcript + insight extraction. Every question becomes an opportunity to improve the KB. Session insights flag gaps. You add to the KB. The protégé gets better. The gaps close.

How to turn it on

Zero-hallucination is on by default. You don't toggle it. Every protégé ships with the architecture in place.

What you can control:

1. Your knowledge base. The more domain content you feed it, the fewer gaps it encounters.
2. Your opening greeting. Tell clients upfront how your protégé handles unknowns. This sets expectation and builds trust.
3. Refinement feedback. When you review session insights (see /features/session-insights), you'll spot topics your protégé encountered without KB coverage. Add them. Your protégé regenerates and redeploys the same session.

FAQ

Can I turn off the zero-hallucination architecture for edge cases? No. The architecture is the product. If you want a model that will guess, there are other tools.

What happens when my protégé says it doesn't know something? It tells the client it doesn't have that information and offers to escalate. You get a summary email after the session. That question becomes a refinement suggestion (see /features/session-insights).

How is this different from Delphi's "responds from source material with citations"? Delphi describes behavior. We describe architecture. Our retrieval system is patent-protected; the claim is on file and public. Their approach is a prompt pattern; ours is a structural one.

Does this work for voice sessions too? Yes. Every voice session runs through the same retrieval pipeline before the model generates audio output. The constraint applies identically.

What if my knowledge base is small at the start? Your protégé will say "I don't have that information yet" more often in the first two weeks than it will after. Every question surfaces what to add next. Most experts close their top-ten gaps within ten sessions.