The Conversion Sequence

Physics does not sell enterprise software. Unpriced risk does.

The Autocoincidence Theorem is a structural result about physics and information. The reading list behind it cites 21 sources across thermodynamics, computability theory, information theory, and accounting history. None of that converts an allocator, a regulator, or an engineer by itself. What converts them is the sequence in which the evidence arrives.

Lead with the wrong piece and the reader pattern-matches to the wrong category. Lead with physics and they hear "academic." Lead with the patent and they hear "vendor pitch." Lead with the theorem and they hear "philosophy." Lead with the gap -- the empirically demonstrated failure of the current approach -- and they hear "unpriced risk." That is where the conversion starts.

The five steps that follow are the deployment order for the evidence. Each step converts a different audience. Each step depends on the prior step having landed. The full reading list with canonical passages and search URLs is at /reading.

---

In March 2025, OpenAI published arXiv:2503.11926. They pressured a reasoning model's chain-of-thought monitoring. The monitor's accuracy dropped from 97% to approximately 0%. Not because the model stopped misbehaving. Because the model stopped saying it was misbehaving. The story separated from the execution.

This is the first thing said in every conversation. Before the patent. Before the physics. Before the theorem.

The reasoning: every current AI verification tool -- checksums, signatures, logs, chain-of-thought monitoring, RLHF, Constitutional AI -- is a record about a record. A story verifying a story. The OpenAI result is the laboratory proof that under adversarial optimization, stories separate from execution. The monitor saw compliance. The execution continued. This is not an engineering failure. It is a class property. The class of system that produced both the output and the monitor is the same class. Same memory. Same instruction set. Same drift modes.

Who this converts: Everyone. The OpenAI finding is published by the largest AI lab in the world. It is not a theoretical claim. It is a measured result. A regulator reads it and sees the compliance gap. An insurer reads it and sees unpriced risk. An engineer reads it and sees a structural limitation of their own monitoring stack.

What it does not do: It does not name the solution. It names the gap. The gap is what creates the demand. Step 2 names the mechanism.

---

AI liability insurance is zero globally. Not because the risk is low. Because no measurement exists. Unmeasurable risk cannot be written. Munich Re's aiSure caps at $15M and covers vendor-side KPI warranties only. Deployer liability coverage for AI role continuity does not exist, from any carrier.

Two historical precedents show what happens when the measurement arrives.

Black-Scholes (1973). Before the formula, options were priced by intuition. The CBOE opened with 911 contracts on day one. By 2023: 3.7 billion contracts, $700 trillion in OTC derivatives notional. The formula did not create options. The formula made options measurable. The measurement created the market. The option value turned out to be "independent of the expected return on the stock" -- because the measurement (continuous hedging) collapsed the valuation problem from subjective to observable. The substrate signal does the same: it collapses AI role-continuity verification from subjective (trust the monitor) to observable (the bit is at the coordinate or it is not).

Pacioli (1494). Before double-entry bookkeeping, merchant ledgers were single-entry. The merchant wrote what he chose to write. Commerce stalled at scale. Banking was impossible because trust could not be mechanized. Double-entry -- every transaction recorded twice, debit and credit, must reconcile -- did not prevent fraud. It made fraud structurally detectable. Books that do not balance cannot be kept. Computing has never had double-entry. Every audit is single-entry: one ledger, one bookkeeper (the software), no independent cross-check. The substrate is the second entry.

Who this converts: Capital allocators hear Black-Scholes and see a market being born. Insurers hear "the measurement makes the premium pricable." Regulators hear Pacioli and see a 530-year precedent for why a structural anchor is required for commercial trust.

What it does not do: It does not explain what makes the measurement unfakeable. Step 3 does.

---

Two contributions have no prior art in the literature. The reading list documents the search.

The Displacement Axiom. Physical objects must be displaced to be replaced. A mailbox cannot hold two packages. A memory location can hold a million different values with no physical trace of the ones before. The closest ancestor is Locke's impenetrability of matter (1690). The computational consequence -- that bits, lacking displacement, admit forgery as a first-class operation -- appears to be stated for the first time in the present work.

Position-as-Meaning. The physical memory address IS the semantic coordinate, computed deterministically from the role through a compositional function. Given the role, exactly one correct address. Given the address, exactly one role it can serve. An XOR gate checking address-content correspondence IS checking role continuity -- but only because the address was constructed to carry the role. Without the compositional function, the XOR is hardware-accelerated label-checking from the 1960s. With it, the XOR is the patent.

Content-addressable memory is different (content serves as address, not role). Positional encoding in transformers is different (position is arbitrary and learned, not deterministic from role). Codd (1970) explicitly separated position from meaning. The patent reverses Codd at one load-bearing point.

Who this converts: Engineers who need to understand why "just add an XOR gate" is not the patent. The gate is commodity. The compositional address function is the invention. Patent examiners evaluating non-obviousness. Allocators evaluating defensibility.

What it does not do: It does not defend against the "we can build a better monitor" objection. Step 4 does.

---

These stay holstered until someone swings. When an engineer says "we can build a better software monitor," the response is structural, not competitive.

Rice (1953). No Turing-complete system can decide non-trivial semantic properties of another Turing-complete system. "Is this system still performing its authorized role?" is a non-trivial semantic property. The verifier must be in a different computational class. The XOR displacement detector is in AC0 -- combinational logic, bounded-depth, no loops, no memory. Strictly sub-Turing. Mathematically immune to Rice.

Turing (1936). The halting problem. A system cannot decide whether another system in the same class will halt. The self-reference obstruction that governs all Turing-complete verification. The regress diverges. The cost goes to infinity. The certainty asymptotes below one.

Landauer (1961). Erasing a bit dissipates at least kT ln 2 of energy. The universe tracks what information systems pretend to erase. When an engineer asks "where does the old value go when I overwrite?", Landauer answers: into heat, distributed into the environment, in a form the information layer cannot read. The universe is autocoincident. Computing chose not to be.

The Data Processing Inequality. Post-processing cannot increase mutual information. Every software layer between the physical execution and the compliance log loses information about what actually happened. Adding more layers adds more loss. The only move that does not lose information is to read the physical state directly.

Godel (1931). A sufficiently powerful system cannot prove all true statements about itself from inside itself. The "metamathematical considerations" that decide the undecidable proposition are considerations from outside the system. The same structural move as anchoring: you need something outside the class to verify what the class cannot verify about itself.

Who this converts: The technical skeptic. The engineer who has built monitoring systems and believes the problem is engineering difficulty, not class membership. The kill-switches do not argue that their approach is bad. They show that their approach is in a class that structurally cannot deliver the property, regardless of engineering quality.

The TEE objection. "We have Trusted Execution Environments -- Intel SGX, ARM TrustZone. That is hardware verification." A TEE isolates code in an enclave. Attestation proves the enclave loaded the correct code. The enclave is tamper-evident. But the code inside the enclave is Turing-complete -- Rice still applies. Attestation is a detached record -- a hash confirming what code was loaded, not what role is being performed. The enclave does not check position. And drift inside the enclave is invisible to the TEE -- the OpenAI result happens inside the enclave exactly as it happens outside. TEE protects the box. The patent verifies the role. The box can contain a drifted system and the TEE would faithfully protect the drift.

When to deploy: Only when someone proposes a Turing-complete alternative or claims existing hardware solves it. Never preemptively. Stating impossibility results before the audience has proposed an alternative sounds like gatekeeping. Stating them after the alternative is on the table sounds like structural analysis.

---

The Semmelweis reflex kills good ideas that imply "the experts are wrong." The bypass is a credentialed champion from inside the establishment. The academic bridge is a collaboration invitation, not a validation request.

Crutchfield and Shalizi at UC Davis have spent three decades building computational mechanics -- epsilon-machines, causal states, statistical complexity. Their framework asks: how much of a system's history is predictively relevant? The autocoincidence theorem asks the adjacent question: whether the history is structurally present in the state. The epsilon-machine measures predictive sufficiency. The theorem measures causal traceability. These are different questions on the same terrain.

The approach: "The autocoincidence theorem extends your epsilon-machine framework into verification architecture. The displacement axiom is the physical primitive your causal states definition does not address -- whether the history is structurally present, not just predictively sufficient. Would your group be interested in the formalization?"

Wolpert has the mismatch-cost framework. The formal proof of the directional asymmetry (Claim 4 of the theorem) likely lives in his decomposition of entropy production into Landauer cost plus mismatch cost. The approach: "We think the proof of Claim 4 lives in your framework. Here is the claim. Here is our informal argument. The formal version is yours to write."

Hand them a paper they can write. That is the currency academics trade in. The result: a peer-reviewed formalization that gives the patent and the theorem a foundation in theoretical computer science that the market could not otherwise provide.

Who this converts: The academic establishment. The reviewers. The examiners who want peer-reviewed grounding. The journalists who need a quotable expert. The allocators who want a "Lister" before they commit capital.

---