The 'AI truth engine' fails procurement
The question no honest system answers yes to — and why the auditable oracle answers the real one.
The "AI truth engine" is the product that fails procurement the moment the examiner asks the one question no honest system can answer yes to: "do you certify this is correct?" The auditable oracle does not answer that question, because no honest system does. It answers the question the examiner actually asks — and that is why it passes.
There is a product shape the market keeps asking for and that no honest system can deliver: the "AI truth engine," which certifies that its outputs are correct. The product fails procurement the moment the examiner asks the one question that defines the category boundary: "do you certify this is correct?" No honest system answers yes, because no honest system can. The systems that say yes are the ones that fail the audit, lose the procurement, or both — because the claim is the liability.
The question that kills the truth engine
A procurement officer under the EU AI Act (binding law, not guidance) who buys an "AI truth engine" will eventually hand it to an examiner. The examiner asks: "do you certify this output is correct?" The vendor has two answers. "Yes" is a claim no honest system can make — and the examiner knows it, and the next question is "show me the basis for that certification," which does not exist. "No" is honest, but it collapses the product: an AI truth engine that does not certify truth is not an AI truth engine. The product is shaped around a question it cannot answer, and the procurement fails at the question.
The question the examiner actually asks
The question the examiner actually asks is not "is this correct?" It is "can you reproduce how you got this?" The first is a truth claim; the second is a provenance claim. The auditable oracle answers the second, by construction: a per-output hash-chained manifest (BR-842), a replay anchor, the model version, the input NodeVersions, the agent trace. The manifest attests provenance and reproducibility, never truth (Certified ≠ True). The oracle passes procurement because it answers the question the examiner asks, not the one the truth engine wished it asked.
The examiner does not ask "is this correct?" The examiner asks "can you reproduce how you got this?" The truth engine is shaped for the first; the auditable oracle is built for the second — and only one of them passes.
- The truth claim ("certifies correct") is the liability — no honest system can make it.
- The provenance claim ("reproducible by construction") is the property — and it is writable into an RFP.
- Certified ≠ True is the boundary: the refusal to certify truth is what makes certifying provenance credible.
Why the refusal is the moat
The refusal to certify truth is not a limitation of the auditable oracle; it is the thing that makes the rest of what it attests credible. A manifest that says "this is reproducible" and also says "this is correct" is a manifest that overclaims, and overclaiming is what an examiner catches. A manifest that says "this is reproducible" and refuses to say "this is correct" is a manifest that knows its boundary — and an examiner reads the boundary as credibility. The refusal is the moat: a competitor who adds "certifies correct" to their manifest has added the claim that fails the audit. We are pre-revenue, building the POC; the manifest is built, the calibration is gated (Gate 1b), and the refusal is the one feature we will not add.
The "AI truth engine" is the product that fails procurement at the question no honest system answers yes to. The auditable oracle is the product that passes procurement by answering the question the examiner actually asks. The difference is the boundary: Certified ≠ True.
— category note