How Double-Checking Actually Works
“We double-checked it” sounds reassuring — but what does it actually mean? In AI verification, a real double-check is a structured process where independent systems examine each claim from different angles, using different knowledge, and arrive at convergent conclusions.
Why “I checked it” isn't checking
When a single AI “reviews” its own output, it's rerunning the same pattern-matching neural network on the same patterns. It has no external ground truth. It can't distinguish between what it knows and what it confabulated — because to the model, those feel identical.
This is why self-review catches formatting errors and obvious logic mistakes, but routinely misses hallucinated facts, fabricated citations, and plausible-sounding statistics that were never real. The model is confident about the wrong things for the same reasons it got them wrong in the first place.
Anatomy of a real double-check
A meaningful verification requires independent signals from independent sources. NoDelulu's architecture was designed around a model that exists in developmental psychology: the family unit.
The NoDelulu Family Tree
The structure has four roles. Each one produces a qualitatively different kind of check — and together they form a complete verification system.
The Mentee — your text
The document you submit is the starting point: raw, authentic, the thing being protected. The goal of everything that follows is to elevate it, not to attack it.
Pass 1 — The Sweep (first Mentor)
The first model reads your entire document and flags every potential hallucination across all 8 categories. Wide net, high recall. It is given the responsibility of first analysis — which means it self-regulates and looks harder. Vygotsky’s research on the Zone of Proximal Development shows that placing accountability horizontally between peers, rather than directing it from authority down, produces better outcomes for both parties than authority-led instruction. Peer-led accountability is structurally different from peer-led voting: the stakes are different when you own the first word.
Vygotsky's Zone of Proximal Development (ZPD): the gap between what someone can do alone and what they can achieve with a peer's guidance. Peer guidance outperforms authority-led instruction for this reason.
Pass 2 — The Review (second Mentor)
The second model analyses your document independently — without seeing Pass 1's findings — then examines Pass 1's work and must account for every finding: confirm it, challenge it, refine it, or add something new. This adversarial challenge is not a bureaucratic step. It is the mechanism that prevents the system from simply echo-chambering the first model's biases.
The Circularity Principle: neither model accesses the web during analysis. Grounding is post-analysis only. This is a deliberate design choice: independent models must form their own judgments before any external evidence is introduced. Pre-grounding the analysis would contaminate the peer review.
Web Search — The Parent
The web is the objective arbiter of external fact. For the categories where evidence can be found — Factual DeLulu, Number DeLulu, Made Up DeLulu, Time/Date DeLulu — live web search checks the claim against the open web. The Parent doesn't debate; it returns ground truth.
Analytical findings — Logical Leap, Opinion As Fact, Self-Contradiction, Missing Context — are not sent to web search. No search engine can tell you whether a conclusion follows from its premises. Those findings stand on the adversarial debate between the Mentors — the two-model peer review.
The Synthesis Model — The Storyteller
A final model receives the adversarial findings, the web evidence, and the original document text. Its role is not to judge further — it takes the evidence from the Parent (web grounding) and the conclusions from the Mentors (peer review), and rewrites each finding explanation into something a human can actually act on.
The Storyteller's framing principle: the user receives their work back elevated, not failed. The report is not a list of accusations. It is a structured roadmap to a stronger document.
The eight types of delulu we hunt
Not all errors look the same. A hallucinated statistic is a different kind of failure than a logical contradiction or a misleading framing. NoDelulu hunts eight distinct types of AI delusion:
- Factual DeLulu — Are the stated facts true?
- Number DeLulu — Are numbers, percentages, and data points accurate?
- Made Up DeLulu — Are citations real and do they say what's claimed?
- Self-Contradiction — Does the text contradict itself?
- Logical Leap — Do the arguments follow from the premises?
- Opinion As Fact — Are subjective claims presented as objective truths?
- Time/Date DeLulu — Are dates, timelines, and sequences correct?
- Missing Context — Are important caveats or counterpoints missing?
Beyond detection, NoDelulu’s adversarial scoring evaluates how each finding relates to the others across the full document — ensuring the consolidated output is internally consistent and accurately weighted.
Convergence: where confidence comes from
The power of adversarial checking isn't in any single signal — it's in convergence. When both models independently flag the same claim and web evidence confirms the problem, that three-way convergence is what separates a high-confidence finding from noise.
This is why NoDelulu can assess severity with meaningful precision. A finding backed by adversarial agreement and source evidence and a clear dimensional classification is qualitatively different from a single model's guess. It's the difference between “maybe check this” and “this is demonstrably wrong — here's the proof.”
Why this matters for your work
Every AI-generated text contains potential hallucinations. The question isn't whether to check — it's whether the checking actually works. A real double-check means:
- Multiple independent systems, not one model reviewing itself
- External evidence, not just model opinions
- Dimensional analysis, not just “is this true?”
- Clickable sources, so you can verify the verification
That's what “double-checks its findings” actually means. Not a rubber stamp — a structured, multi-layered audit.