Infographic · ThetaDriven · 2026-04-15

A System Cannot Prove A Property Of Itself

Why deterministic behaviour is not determinism. Self-reference, not randomness, is what Turing proved. The diagnostic test for every AI verification claim, on one page, for the people writing the checks.

01The Diagnostic Question

Does the mechanism that is supposed to catch the problem run on a substrate that can execute arbitrary programs? YES NO INHERITS THE REGRESS Turing-complete. Can exhibit any failure mode it was meant to catch. Cannot verify itself. SPECIFY THE CLASS Combinational logic. No branch, no loop, no state to mutate. Escapes by construction.

The whole test has two answers. Everything else is detail.

02Self-Reference, Not Randomness

In 1936, Alan Turing proved that no general machine can decide whether an arbitrary program halts. The proof uses diagonalization — constructing a machine that asks itself whether another machine, built from its own specification, would halt. Any answer contradicts the decision procedure.

The proof is not about randomness. It is about self-reference. Any sufficiently powerful system contains descriptions of itself. When it tries to decide a property of itself, the reference loops back into the diagonal. That is where decidability fails.

Gödel 1931 proved the same structural result in logic. Rice 1953 extended it: any non-trivial semantic property of a Turing-complete program is undecidable.

SYSTEM S does S have property P? asks itself → diagonal → undecidable Turing 1936 · Gödel 1931 · Rice 1953 89+ years settled · not contested The randomness is not the problem. The self-reference is.

03Deterministic Is Not The Same As Delegable

IN PRINCIPLE (CONTROLLED) Inputs A (pinned) Deterministic inference Output A ✓ Same inputs → same output. Reproducible. Useful for debugging. Not verification. IN DEPLOYMENT (RUNTIME) Inputs drifted context · tool-use retrieval · injection Deterministic inference Output X (wrong) Deterministic wrong output. Re-running gets the wrong answer twice. "Same inputs, same output" is a property about the machine. Delegability is a property about whether you can trust it across time. Determinism is the baseline of the halting-problem proof. Not the exception.

Deterministic inference on drifted inputs produces deterministically wrong outputs. Re-running gets the wrong answer twice.

04Seven Independent Paths, One Conclusion

When seven independent arguments converge on the same answer, the answer is structural — not rhetorical.

Runtime role continuity requires a lower class 01 · COMPUTABILITY Turing · Gödel · Rice 02 · INPUT-PLANE DRIFT context · tools · retrieval 03 · REGULATORY Article 14 04 · ACTUARIAL carriers price signals 05 · LEGAL PRECEDENT Notified Bodies · SOX · MiFID II 06 · ENGINEERING stabilizer needs stabilizer 07 · MARKET 2008 CDS pattern

Each path is independent. Each arrives at the same conclusion. That is what makes the conclusion load-bearing.

05Role Continuity Sits Below Identity

IDENTITY hash · signature · model card answers: what are the bits? ↓ sits above ↓ ROLE CONTINUITY the relationship between what the thing does and what it was trusted to do answers: is it still performing the function? Regulators, carriers, courts all ask the bottom question.

Identity is cheap. A hash, a signature, a model-card version string. It tells you what the bits are. It does not tell you whether the bits are still performing the function they were authorized to perform.

A system can have a valid hash and still drift behaviorally. Weights can accumulate fine-tuning updates while the top-level hash records the change as authorized. A policy document can be signed and structurally inadequate for what the deployment is now doing.

Nick Mabe on the record "Hashes only ever give you tamper-evidence, not continuity."
Tamper-evidence is not continuity-evidence.

Role continuity is what regulators, carriers, and courts will ask about. Not: is the code signed. But: is the function intact.

06Which Layers Escape The Regress?

Verification layer Turing-complete? Escapes regress? Why
Software governance dashboard YES NO Inherits
Cryptographic attestation chain YES NO Verifier software inherits
TEE (Intel TDX / AMD SEV-SNP / NVIDIA CC) YES (inside) NO Isolation ≠ class separation
Policy engine over symbolic state YES NO Rule evaluator inherits
Formal verification suite YES PARTIAL Only decidable properties
Human-in-the-loop review YES (tools) NO Review tools inherit
Legal commitment framework N/A NO Liability-allocation, not prevention
Combinational logic comparator NO YES No instruction set to drift into
The Category Error Most Of The Market Is Making A GPU is hardware. A GPU is Turing-complete. A GPU does not escape the regress. The boundary is not silicon-vs-software. The boundary is combinational logic vs instruction execution. Different chips, same class. Different class, possibly same chip.

07Before Today / After Today

BEFORE 2026-04-15 this post is published AFTER AUG 2, 2026 EU AI Act live DEFENSIBLE POSITION "I did not know." State of the art was unclear. ARGUMENT IS WEAKER "Could you have known?" Information public, referenced, traceable. Access to the argument shifts the standard of care. This is not a threat. It is the structure of fiduciary duty in the presence of accessible technical information.
The Fiduciary Question — Compressed When your deployment is questioned — by a regulator, a carrier, a board member, or a plaintiff — will you be able to point to a measurable runtime signal that the AI was performing its authorized functional role, or will you be able to point only to policy documents and attestation signatures? The first answer survives. The second answer survives in proportion to how available the first answer was at the time of signing.

08What Closes The Regress

EXPECTED rank-computed address ACTUAL where data resolves XOR = 0 → match · cache hit · role intact ≠ 0 → displacement · cache miss · role drifted NO PROGRAM COUNTER NO BRANCH NO LOOP NO STATE TO MUTATE One clock cycle. Deterministic output (0 or non-zero). Cannot hallucinate. Cannot drift. Not in the class the halting problem applies to. US 19/637,714 · 36 CLAIMS · TRACK ONE

Position encodes functional role. The fetch IS the verification. Detection and correction vector in one hardware cycle.

Deterministic inference on drifted inputs produces deterministically wrong outputs. Re-running gets the wrong answer twice.