01
What is a hallucination, and why can a model be confidently wrong?
A hallucination is a plausible but incorrect answer delivered with confidence. A language model predicts what is likely to follow in text, not what is necessarily true in the world. Left alone, it can invent a fact, a source, or a number—and present it without hesitation.
The danger is not merely error; it is persuasive error. It can look indistinguishable from truth. The answer is not to ask a model to be honest—that is insufficient—but to build an external verification system around it that catches unsupported claims before they reach a user.
In short: the safeguard is not the model alone; it is the architecture around it—separated roles, source grounding, and independent review.
02
How do you stop the model from inventing facts?
We ground the answer in verified sources. The system first retrieves relevant documents from your knowledge base, then builds its response from those materials, with a trace back to the source of each factual claim. The model does not rely on a vague recollection; it works with retrieved evidence.
- Answers come from sources, not from memory. The model reasons over retrieved fragments of your verified base rather than its pre-training memory.
- Every factual claim has a trace. A statement is paired with a source that can be opened and checked.
- No evidence means no fabricated answer. If the base provides no support, the system says that evidence is insufficient instead of filling the gap. “I do not know” is a valid result.
- The task is defined in verifiable terms. Before work starts, the system specifies what a correct result must satisfy, then evaluates against those criteria.
Result: the model answers a grounded question with evidence rather than generating smooth prose in a vacuum.
03
Does AI verify its own answer?
No—and that distinction is essential. The agent that produces work and the agent that reviews it are different agents. An answer’s author does not score its own output.
In the orchestrator, roles are separated: an executor drafts an answer or performs an action; an independent reviewer collects evidence and judges against facts, not against confidence in the author. Self-confirmation is a major source of hallucinations, so the architecture removes it.
- Executor is not reviewer. A separate agent, with no incentive to declare the work finished, performs the check.
- Review requires evidence. The reviewer asks for a source or confirmation; it does not accept an assertion on trust.
- Completion is criteria-based. The result is compared against a pre-defined specification rather than judged by appearance.
Why it matters: one model reviewing itself tends to endorse its own errors. Independent review is designed to catch them.
04
What if an error sounds plausible?
The most dangerous errors are persuasive ones. They are addressed with adversarial review: several independent reviewers do not merely confirm an answer; they actively try to disprove it.
- A mandate to falsify. The reviewer deliberately looks for where a claim breaks rather than where it seems acceptable.
- Different perspectives. Correctness, fidelity to the source, and reproducibility are checked from distinct angles instead of duplicating the same review.
- Only resilient claims pass. A statement is accepted only after it withstands attempts to refute it. Otherwise it is revised or escalated to a human.
The key idea: a claim becomes accepted not because it reads fluently, but because it survived attempts to disprove it.
05
Why does the system not rush to answer, and how does it know when work is complete?
The main risk in long tasks is premature completion: the system produces the first plausible result and stops. The orchestrator is designed not to trust that first answer.
- Iterative gap search. After each step, the system asks what remains unverified or missing. That starts another cycle until the work has no known gaps.
- Completion is evidence-based. Final acceptance begins only when the assembled evidence shows that the criteria have been met. Even a request to close work is subject to review.
- Plan and boundaries come first. The task is decomposed into a plan with explicit, checkable criteria. The outcome is compared with something concrete rather than an impression.
In short: “done” does not mean “a plausible answer appeared.” It means the criteria are met and independently supported.
06
Every protection layer: the verification pipeline
Hallucination control is not one wall. It is a series of checks. An answer passes through them in order, and the verification boundary separates the production of an answer from independent control.
Task definition and plan
The task is broken into steps with explicit, testable success criteria and control points. This creates the specification against which the result will later be checked.
Source grounding
Retrieval runs over your verified base; the response is built from retrieved fragments with citations. No support means “insufficient evidence,” not a guess.
Executor draft
A working agent produces an answer or action. It is not yet a result; it is a candidate that must pass review.
— — — verification boundary — — —Independent review
Another agent, or several agents, gathers evidence and adversarially tries to refute each claim:
- the reviewer optimises for facts, not for the author’s convenience;
- contested claims are checked from different angles: correctness, source support, and reproducibility;
- the decision relies on independent agreement rather than one voice.
Gap search
“What remains unverified or missing?” triggers another work cycle until no known gap remains. Premature completion does not pass this stage.
Delivery, stop, and audit
Final delivery occurs only when criteria are met and confirmed. Disputed points are labelled and escalated to a human, not claimed as truth; facts are traceable to sources and outgoing requests can be logged.