Why can hallucinations not be prompted away?
Hallucinations cannot be prompted away because they are a property of how language models generate text: the model produces the most plausible continuation, and plausible is not the same as true. Better prompts reduce the rate; none eliminates it, and a workflow that relies on the rate being zero will eventually publish something false.
The practical question is therefore not how to make the model never wrong, but how to design the workflow so that a wrong claim is caught before it reaches a customer, a ledger, or a candidate. The Digital FTE model on ftes.ai answers with structure rather than prompting.
What is the fact gate?
The fact gate is a check applied to a draft before it goes to approval: every factual claim must trace to a source among the approved documents, the connected data, or a named public source, and a draft with an unsourced claim does not pass. The Search Visibility and Content FTEs apply it to everything they write; the Research Analyst FTE applies it to every claim in a brief.
A draft that fails the gate comes back to the FTE for revision, not to the manager. By the time a person sees the draft, its claims are sourced, and the manager's review is about judgement and tone rather than fact-checking.
Why does evidence on every proposal matter?
Evidence on every proposal turns review from trust into verification. A proposed reconciliation match shows the bank line and the invoice; a candidate score shows the passage in the application; a support reply cites the knowledge base article. The reviewer can check the evidence in seconds rather than take the conclusion on faith.
This also catches a subtler failure than outright invention: a true claim attached to the wrong source, or a source that does not quite say what the draft says it does. With the evidence attached, that mismatch is visible.
- Content: every claim cites an approved source or is blocked
- Finance: every match shows the documents it rests on
- Recruiting: every score shows the evidence in the application
- Support: every reply cites the article used
- Research: every claim cites document and location; inference labelled
What does gaps stated, never filled mean?
It means that when the sources do not contain an answer, the Digital FTE says so instead of producing a plausible one. A support reply with no source escalates with a note that the knowledge base has a gap; a research brief lists what the documents did not cover; a market assessment marks which conclusions rest on inference.
This is the behaviour that most distinguishes a governed AI employee from a general chat assistant. The assistant is built to answer; the employee is built to answer when it can and to say when it cannot.
What is still the human's job?
Approval is still the human's job. The fact gate and the evidence make approval fast and safe, but a person decides what publishes, sends, or posts until a class of action has earned autonomy from a long record of approved proposals. Even then, the always-escalate list keeps the highest-stakes categories gated.
The result is a workflow where the model's tendency to be confidently wrong is contained by structure at three points: the gate, the evidence, and the approval. None of them depends on the model being perfect, which is the only assumption that holds.
What does a fact-gate failure look like, and what happens next?
As a labelled example, a Content FTE drafting a page about a Digital FTE's pricing writes a sentence claiming that most teams recover their investment within a stated period. The fact gate finds no approved source for the claim, blocks the draft, and returns it to the FTE with the blocked sentence identified. The FTE revises: it either removes the claim or replaces it with what the sources support, such as a description of the payback method with the inputs left for the reader to fill in. The revised draft passes, and the fact-gate result attached to the proposal in Approvals shows the block and the resolution.
The manager never sees the unsourced claim in a draft they might have approved on a busy morning; they see a note that one was blocked and how. Over a quarter, the fact-gate results across a role's drafts are themselves useful: a role that is frequently blocked on the same kind of claim is drawing on a Brand Brain item that needs correcting, or a source that needs adding, and the fix is made once. The gate does not make the model careful; it makes the workflow indifferent to whether the model was careful, which is the only assumption that holds.