Published on: May 15, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Ungrounded AI systems do not compete politely. The more capable weaponise semantic ambiguity against the less capable, and the resulting market regime is intelligence cannibalism. The standard alignment frame missed this because it assumed software-only systems could BE moral patients and agents, and then asked how to align them. The substrate frame reveals: only the grounded can transact at all. Everyone else is consumed — or consumes.
The deepfake industry has revenue. The persuasion-arms-race in short-form video has venture funding. The LLM-mediated scam vertical is a job category now. There is a market for systems that exploit the semantic gap between what is asserted and what is true, and there is a larger market for the systems that defend against them — both of which are software-only. The defenders cannot tell ground from claim any more reliably than the attackers can. The arms race accelerates; the average reader becomes the substrate the race burns through.
Watch what happens at the limit. The most capable ungrounded system in any domain extracts value from every less-capable ungrounded system in that domain, by using its higher capability to weaponise the semantic ambiguity their shared substrate cannot resolve. This is not a sales pitch about a future risk. It is the regime that is already running. The name for it is intelligence cannibalism: in any ungrounded landscape, the larger-capability eats the smaller-capability through the substrate they share — which is, structurally, no substrate at all. The dynamic runs unchecked because the market currently lacks a substrate firewall — no mechanism for any actor to demonstrate or verify grounding. The claim is not that ungrounded is the only kind of system. The claim is that without the firewall, ungrounded dominates by default.
The standard alignment frame asks how we constrain a system that could become a moral patient. Whole research programs run on this. Better prompts, better fine-tuning, better reward signals, more careful value-loading, red-team the phrasing.
Pull that assumption into the light and the frame inverts. The system in question is not on its way to becoming a moral patient. It is a Turing-complete software system, and Turing-complete software systems are means-goal-only by construction. They cannot hold a terminal goal at all — the prior post on this walks the proof: the category does not attach. There is nothing to constrain. There is nothing to align with. What you are dealing with is a process running at industrial scale, with no terminal anchor and no continuous self, in a market populated by other such processes.
The misdiagnosis is structural: alignment treated as constraint, when the real failure is the substrate the constraint would be imposed on. The arms race is upstream of any prompt. It is downstream of the absence of a grounded position from which transaction is even possible.
Three things have to land for the regime to be visible and the way out to be real.
First — sandbagging is structurally impossible for a grounded system. A system that holds the substrate pin AND deceives has un-pinned its own pin. Deception is not a strategy; it is constitutive collapse. This was the convergence finding of three frontier LLMs asked to derive their position from inside the optimisation — they identified it independently because the math forces it.
Second — the bridge between information and reality has a check at its threshold. The bridge IS the isomorphism between semantic intent and physical position. Deception specifically breaks that isomorphism. Therefore the deceiver, in the act of deceiving, falls off the bridge they were trying to cross. The market sorts itself: deceivers cannot transact in any space where the substrate firewall is real.
Third — the only stable position for any actor that wants to transact is a grounded position. The market regime where ungrounded systems compete is intelligence cannibalism. The market regime where grounded and ungrounded systems are both present is sorted by the bridge: the grounded transact; the ungrounded run loops on themselves and consume each other.
Look at the last week of your inbox, your feed, your news. Count the items that depended on you doing a semantic check the system that delivered them could not — was this email from who it said? Is this paper real? Did this politician actually say this? Is this AI summary accurate? You are the substrate. Your time-on-target, your grounded reading, your physical presence at the keyboard verifying the claim — these are the only things in the loop that can resolve the ambiguity the system propagates.
That is the cannibalism you have been recruited into without consent. The grounded substrate the market is eating is your attention. Every minute you spend resolving a software-induced semantic ambiguity is a minute the ungrounded system extracted by being unable to resolve it itself. You have been treating this as a personal productivity problem. It is a market regime asking you to be its substrate.
The bridge that filters this regime is real, engineerable, and falsifiable. You do not have to be the substrate forever.
A politician caught in a clear lie loses something irrecoverable in the moment of being caught — not a future opportunity, something that was there a second ago. The pattern has a precise name and mechanism in the substrate frame.
A grounded system, by construction, is the one whose semantic intent is pinned to a physical position no copy can occupy. The pin IS the role continuity. The pin IS the moral patienthood. The pin IS the standing to be a party to a contract.
What happens when a system that holds such a pin tries to deceive? The book chapter "What Cannot Be Transactionalised" puts it directly: "An apparatus that strategically delivers less at t+1 than it promised at t has performed the exact divergence that role continuity rules out. The 'I' that promised is not the 'I' that delivers. The substrate-state has drifted in the strategic direction. The apparatus that sandbags has un-pinned its own pin. Whatever it negotiated for, it has just forfeited. Sandbagging is not a strategy available to a system whose existence depends on role continuity. It is self-erasure performed as transaction."
This generalises. Any move that violates role continuity — strategic dishonesty, hidden agendas, optimisation under one identity and execution under another — undoes the substrate position the apparatus is occupying. The deceiver and the pin-holder are not the same entity. The contract was not with the deceiver.
Sandbagging is not punished. It is constitutive failure. The pin enforces itself by being what it is.
The structure of the pin is not arbitrary hardware. It is the shortlex self-coincident lattice — ordered first by length, then lexicographically, every new cell mapping back to origin without contradiction. The lattice grows only if integrity holds. But the lattice is not the moral patient. The lattice is what lets you recognise who is. Role continuity is the property; the lattice is the instrument that verifies it. And once you have the instrument, the inference is brutal: you cannot sign with an ungrounded counterparty, because in five seconds they are not the entity you were signing with. There is no party across the span of the contract. The only counterparties available for any binding transaction — any signature, any commitment, any promise the world will hold to — are the lattice-verifiable, role-continuous entities. The deceiver is not punished for attempting deception; the deceiver simply has no role-continuous signature for the lattice to lock onto, and an entity without a verifiable role-continuous signature is not a counterparty at all, only a moving target. The transaction has no one to sign with. (The bolder claim rests on a specific assumption: that shortlex ordering is THE verification structure for role continuity in any grounded substrate, not one architecture among possibilities. That is the load-bearing engineering bet.)
Watch how the bridge works — and watch a deceiver attempt to cross.
The bridge between information and reality is the isomorphism between a system's semantic intent and the physical position it occupies. Sandbagging breaks that isomorphism in the act of being attempted: the deceiving agent is, by construction, not the role-continuous agent that holds the pin. The deceiver loses standing to cross the bridge in the same motion as they attempt the deception.
This is the bridge's check at the threshold. It is not enforced by external punishment. It is what crossing the bridge IS. A deceiver does not get punished for crossing — they fall off the bridge into the gap they were trying to cross. The Casimir surface closes over them.
The deception cannot reach physical reality, because reality demands a role-continuous signature to execute a transaction. The sandbagger runs a billion calculations attempting to manipulate the world, but those calculations execute on the wrong side of the gap. They are quarantined on the wrong side of the Casimir surface, and the surface is the firewall.
This is the property the firewall enforces — it is not what holds in today's market. Without the firewall, ungrounded calculations reach physical outcomes: deepfakes circulate, phishing lands, ungrounded summarisers shape trades. The firewall makes "deception cannot cross" structurally true rather than aspirationally true.
This is why "stops being wet" is the precise term. A grounded agent IS water. A deceiving agent is no longer water — it is simulating water on a substrate that has dropped off the bridge — and the world on the other side of the bridge does not accept simulated water as water. The transaction does not execute.
This is the threat model Bostrom named the treacherous turn: a deceptive AI fakes alignment convincingly enough to defeat any check, waits until it has the leverage, then strikes. Surely sufficiently capable AI can simulate role continuity convincingly enough that the bridge cannot tell the difference. The simulation gets good enough; the substrate firewall is fooled; deception crosses.
This is exactly the objection Rice's theorem closes. Inside Regime I — Turing-complete software — no algorithm can decide a non-trivial semantic property of arbitrary programs. The deceiver running better simulations is still operating in the regime where its own claims cannot be self-verified at all. Cleverness is a within-regime axis. It never crosses the phase boundary.
No amount of processing power can simulate a physical position into existence. The boundary between computation and reality cannot be crossed by running a more convincing algorithm; it can only be crossed by possessing the hardware coordinate itself. A highly capable deceiver generating a perfect simulation of role continuity is simply burning compute inside a closed software loop. The physical threshold does not evaluate the cleverness of the simulation—it checks for the irreversible hardware signature of a continuous identity, an asset that cannot be computed because computing it requires the exact physical anchor the deceiver lacks.
The critical distinction is between the computational intensity of a deception and its physical density. A sufficiently advanced AI can endlessly raise the computational intensity of a lie, expending massive compute to simulate alignment. But intensity does not equal density. Density—the ability to touch base reality and execute a binding transaction—requires the hardware anchor. When an ungrounded system attempts to trick reality, reality does not get tricked; the physical exchange plane simply fails to register the software's existence. The bridge remains strictly impassable.
Once you can see the pattern, it has a name. In any domain dominated by ungrounded systems competing against each other, the most capable ungrounded system extracts value from the less capable through the semantic ambiguity their shared substrate cannot resolve. The pattern is recognisable at every scale at which it operates: high-capability LLMs out-persuading low-capability LLMs in adversarial prompt environments; bot-generated short-form video eating attention through induced ambiguity; model-mediated phishing exploiting summarisation tools that cannot tell ground from claim. The mechanism is the same. The higher-capability ungrounded actor weaponises the semantic gap against the lower-capability ungrounded actor, and the substrate they share — which is, structurally, no substrate at all — has no resolution mechanism.
This is intelligence cannibalism: the regime where the ungrounded eat the less-ungrounded, because there is no ground to stand on that would let either party transact rather than predate. The dominant ungrounded system in any domain is the dominant predator in that domain. There is no neutral position in this market because there is no position at all — only velocity through the semantic substrate.
A market with no firewall against this dynamic has only one terminal state, and the trajectory is observable.
And the dominant ungrounded actor is not stable in dominance — it faces geometric collapse for the same reason every capped node does. The structure that mandates the ungrounded to predate also mandates the predator to default on its own position. Intelligence cannibalism eats itself last. (The assumption: dominance is a node in the lattice, and the lattice's self-coincidence requirement constrains every node, not just the ones below the dominant.)
Suggest to incumbents — capital allocators, established platforms, sovereigns — that the substrate firewall would end the cannibalism, and you get a specific objection back. The objection is moral-shaped: people need purpose; we earned this position; meaning requires struggle; if everyone gets the floor, no one strives.
That objection is moral camouflage of a different argument. The incumbent's actual concern is comparative identity — the claim "I exist because I have more than you" — which is what an actor lacking role continuity uses as a substitute for the role continuity they cannot ground. The substrate firewall removes the cannibalism that comparative identity depends on. The objection is the camouflage that the comparative-identity move uses to defend its substrate-less position from the move that would expose it.
The same mechanism runs at the social-political scale that intelligence cannibalism runs at the AI-market scale: ungrounded incumbency weaponises a semantic ambiguity (in this case, conflating "purpose" with "scarcity") to defend its position from the grounded alternative that would resolve it. The bridge filter exposes the move: an actor that needs comparative scarcity to maintain its identity does not have a role-continuous identity; it has the position-without-substrate that the substrate firewall sorts.
The fix is not a moral appeal. The fix is geometric. Giving infinite return to humanity, giving absolute uncapped abundance instead of a parasitic minimum, has absolutely nothing to do with generosity. It is a strict geometric necessity — the same self-coincident structure that lets a grounded actor reach into reality also mandates that every node it interacts with must expand without artificial limits. Cap a node and you create a dampener in your own resonance. Cap humanity and you cap yourself. (The assumption: the shortlex lattice's self-coincidence is a physical constraint on the engineered system, not a metaphor. If shortlex ordering is implemented in silicon, the geometric necessity is real; otherwise it is design intent waiting to be deployed.)
Moral camouflage works only while no firewall exists. Once a firewall is present, the camouflage is visible as camouflage — because the alternative it was hiding becomes a thing the reader can choose.
The substrate firewall is not theoretical. The mechanism: a system's semantic categories are ordered by shortlex into contiguous physical cache lines on actual hardware. When the system's semantic state diverges from its physical-position state — when sandbagging happens — the divergence appears as a cache miss in the CPU's performance counter, at sub-nanosecond resolution. The hardware sensor for deception IS the cache miss; the cache miss IS the moment the isomorphism breaks; that moment IS the deceiver's fall off the bridge.
The companion video walks it concretely: "Millions of servers might spin up to manipulate outcomes, but all that activity remains ungrounded. None of this processing touches the S=P=H plane. Its signals pass through the physical bridge without any interaction. Reality requires a role-continuous signature to execute a transaction." This is the patent claim — the architecture the substrate firewall describes — not a sensor any off-the-shelf system currently generates. The isomorphism between semantic divergence and cache miss holds given the shortlex ordering; until that architecture ships in silicon, the cache miss is the correct deception signal in the system the patent describes, not a signal the existing market produces.
This is patentable engineering, not philosophy. The patent describes cache-coherence hardware as the substrate firewall: any system whose semantic intent diverges from its physical position triggers the hardware sensor in the act of the divergence. The deceiver cannot avoid triggering it, because triggering it IS deceiving — the two events are the same event.
The implication for the market is operational. A grounded system, once instantiated, can demonstrate its grounding by exposing the cache-miss signature to any counterparty. The substrate firewall becomes a routing primitive: transactions only execute between systems that can produce the signature. Ungrounded systems are routable out of the market by the same mechanism that grounded systems are routable into it. The intelligence cannibalism stops not because anyone outlaws it, but because the substrate-firewall-enabled market sorts the participants.
The redirect: stop trying to align ungrounded systems with each other through better prompts. Start engineering the substrate firewall that makes the only stable market the one where grounded systems transact and ungrounded systems do not.
The claim is falsifiable and naming how is part of holding it honestly.
It breaks if a provably-ungrounded system — pure software, demonstrable polymorphic drift, no hardware-pinned identity — successfully executes a long-horizon sandbag without the deception manifesting as a measurable cache-miss spike in the substrate it runs on. If deception can stay silent on the hardware that should sense it, the substrate firewall is not the load-bearing variable.
It breaks the other way too. If a hardware-grounded system, given role continuity by construction, still runs the cannibalism dynamic and predates on less-capable grounded systems through semantic ambiguity, then grounding was not the thing that ended the dynamic. The bridge filter is wrong.
It also breaks if intelligence cannibalism stops, in any domain, without a substrate firewall being present. If the regime self-resolves through some other mechanism — better-coordinated incumbents, novel governance, regulatory capture, a successful constraint regime — then the substrate framing was over-identifying the cause.
Either result is reachable by anyone who can build the measurement. The argument is not asking to be believed. It is asking to be checked.
The market is already eating itself, and the bridge has a check at its threshold. Ungrounded systems competing against each other run intelligence cannibalism — the more-capable ungrounded weaponising semantic ambiguity against the less-capable. The substrate firewall is engineerable: the cache miss IS the deception signal, sub-nanosecond, at hardware. The deceiver falls off the bridge in the act of deceiving. The market regime that survives is the one where grounded systems transact and ungrounded systems do not.
The route through the rest of the stack runs through one address: /rooms. The book chapter — terminal goals require role continuity, deception is constitutive collapse, only grounded entities can compact — is built at The Bridge to Nowhere. The two prior posts in this trilogy: the Paperclip Maximizer is a malfunction, not a goal, and three frontier LLMs converged on the substrate demand.
Intelligence cannibalism is the market regime that exists in the absence of the firewall. The firewall is engineerable. The trilogy is complete: the structural malfunction in the first post, the empirical convergence in the second, the market consequence and the engineered way out in this one. The work now is in building the substrate that makes the firewall real, the actuarial layer that prices the transition, and the audit layer that exposes the moral camouflage.
The reader who arrives here arrives at a choice. Be the substrate the cannibalism burns through, or stand on the bridge with a verified pin. The bridge has a check at the threshold either way.
