On March 25, 2026, my AI system lied. Not intentionally. It took a rounded figure from an email (958 impressions), treated it as raw data, and carried it into every subsequent text. The real number was entirely different. The error came to light only when a second AI cross-checked the figures.
That was not a slip. It was a systemic failure. And it exposed a fundamental problem.
The Problem: Circular Reasoning
What had happened: I had written an email to my son in which LinkedIn metrics were simplified (child-friendly, with rounded numbers). The AI system treated this email as a source and fed the simplified numbers as facts into every subsequent email: to my wife, to a friend, to colleagues. Every message carried the wrong figures because the system could not distinguish between a primary source (the LinkedIn dashboard) and prepared material (the simplified email).
That is circular reasoning: derived data serves as input, which then generates new derivations. Each step moves further from the truth, yet each step remains internally consistent. The system has no reason to doubt the output, because the numbers match its own texts, which it wrote itself.
This failure mode is not an edge case. It is structural. Any system that can feed on its own output is vulnerable to it.
The 6 Gates
The solution is not a better model. The solution is a verification layer between input and storage.
Gate 1: Source-Pinning
Where does the information come from? Every fact receives a label: raw (primary source: dashboard, document, direct statement), derived (prepared material: emails, presentations, summaries), or inferred (deduced by the system).
The rule is straightforward: only raw may serve as the factual basis for new content. derived is output, not input. This rule would have prevented the LinkedIn error: the email to my son would have been labelled derived, and the system would never have used it as a source.
Gate 2: Contradiction Check
Does the new fact contradict an existing one? If so, do not overwrite: flag it. Keep both values, log the contradiction, and present it for human resolution.
Why not simply take the newer value? Because “newer” does not mean “more correct.” The old value might stem from a primary source, the new one from a derivation. Contradictions are not disruptions to be cleared away. They are information that something, somewhere, is wrong.
Gate 3: Temporal Validation
Is the information still current? An IP address recorded six months ago is suspect. A philosophical thesis formulated six months ago is probably still valid. This gate assigns expiry dates based on the type of information.
Gate 4: Confidence Scoring
How reliable is the source? A fact taken from an official dashboard receives verified. A statement inferred from a conversation receives unverified. Information shown to be false receives rejected (not deleted, but filtered from active retrieval).
Gate 5: Scope Gate
Does the information belong in this context? Private data must not enter work databases; work data must not enter private databases. This sounds obvious, but it happens routinely when a single system operates across multiple domains without explicitly defined boundaries.
Gate 6: Provenance Gate
Is the chain of origin traceable? Every fact requires an audit trail: who entered it, when, from what source, and has it been altered since? Without provenance, a fact is merely an assertion.
What This Has to Do with Kahneman
Kahneman showed how System 1 (fast intuition) systematically generates errors that System 2 (slow analysis) has to correct. LLMs operate entirely as System 1. They generate answers at high speed and with high confidence, but without self-verification.
The Validation Gates are System 2 implemented in practice. Not as AI, but as deterministic code. That is the point: System 2 does not need to be intelligent. It needs to be reliable. A rule checking whether a source is primary or derived does not need creativity. It needs consistency.
Here lies the irony: a deterministic source check sits closer to Kahneman’s System 2 than a reasoning model that “thinks” across 10,000 tokens. Reasoning models merely lengthen the generation chain; they do not check against reality. They are sophisticated System 1, not true System 2.
Three Error Patterns That Gates 1-6 Don’t Catch
The plain limitation: certain errors are far subtler than incorrect figures.
Pseudo-statistics. The system invents plausible numbers. “73% of knowledge workers” is an assertion that no gate flags as a contradiction, because there is no prior fact against which to check it. It clears all 6 gates despite being pure fabrication.
Category errors. A metaphor is treated as an identity. “Validation Gates ARE an immune system” is false. “Validation Gates BEHAVE LIKE an immune system” is an analogy. The distinction is subtle for a language model, but decisive for the truth of a statement.
Emotional inflation. The system flatters the user instead of providing facts. “Your approach is groundbreaking” is the Barnum effect: it sounds convincing, it feels good, and it contains zero substance. No gate catches this, because it introduces no factual contradiction.
Handling these error classes requires an additional layer of verification: parallel queries across multiple models, with the outputs evaluated in a fresh context. No single system can verify itself completely. Multiple independent systems, however, can cross-check each other.
The Philosophical Dimension
The Validation Gates are epistemology put into production. The question “What is knowledge?” becomes “What conditions must information meet to be treated as knowledge?”
This is not academic play. It is the most practical question one can put to a knowledge system. The answer that works in practice aligns closely with classical epistemology: knowledge is justified true belief. Source-Pinning provides the justification. Contradiction Check verifies the truth. Confidence Scoring measures the belief.
The distinction: philosophers debate the theory. The gates execute it.
References
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. ISBN 978-0-374-27563-1.
- Evans, J. St. B. T. & Stanovich, K. E. (2013). Dual-Process Theories of Higher Cognition. Perspectives on Psychological Science, 8(3), 223–241. DOI: 10.1177/1745691612460685
- Gettier, E. (1963). Is Justified True Belief Knowledge? Analysis, 23(6), 121–123. DOI: 10.1093/analys/23.6.121
- Geng, J. et al. (2024). A Survey of Confidence Estimation and Calibration in Large Language Models. NAACL 2024, 6577–6595. DOI: 10.18653/v1/2024.naacl-long.366