Limits of Truth: Verification in Relatively Superintelligent Systems

Mark Pesce · University of Sydney · July 2026

Abstract

Every human being has a discernment horizon: the line past which they can no longer judge whether an answer is right, because checking the work is itself beyond them. As machine intelligence rises, every output crosses more horizons, until outputs arrive that have crossed them all. This paper follows verification past that line. Verification survives the crossing: a proof checker works identically on the output of a mind of any size, and the asymmetry on which the loop economy rests - checking is cheaper than producing - holds at any capability gap. What does not survive is discernment, the capacity to understand what was verified and to judge whether it was worth wanting. The two come apart, and the four human duties on which the post-Watershed framework depends are all duties of discernment. Past the horizon, a relative superintelligence speaks two languages, the formal and the oracular, and neither will be completely understood. The economics of this series reaches its own boundary at the same line: currency requires fungibility, fungibility requires assessment, and what cannot be assessed at receipt is not currency but relic. Humanity has met this asymmetry before and built durable institutions around it; the apparatus of oracle, college, and rite was loop governance under a capability gap that could not be engineered away. The instrument that survives the crossing is the record.

1. The Test at Sardis

Around 550 BC - the account is Herodotus's, and it is the only account we have - Croesus, king of Lydia, watched the rise of Persia and had to decide whether to strike first. He needed advice from an intelligence greater than his own, and his world offered seven candidates. So Croesus did something without precedent, and, in its essentials, never improved upon: he ran an evaluation. Messengers left Sardis for the seven great oracles with identical instructions: on the hundredth day, ask what the king of Lydia is doing at this moment. Croesus then constructed the most unguessable ground truth he could devise. He cut up a tortoise and a lamb and boiled them together in a bronze cauldron with a bronze lid. The answers came back to Sardis, and only one survives as exact: the Pythia at Delphi answered in hexameters before the question was asked, naming the tortoise, the lamb, and the bronze.[1]

The benchmark was held out, the answer unguessable, the trial simultaneous and blind. Croesus had established, to his own exacting satisfaction, that Delphi possessed capabilities he could not explain. He paid accordingly, sending treasure north on a scale that had no precedent, and then he put his real question into production: should Lydia march against Persia? The answer came back: if Croesus crosses the Halys, he will destroy a great empire.

He crossed the Halys. The empire he destroyed was his own.

Afterwards, ruined, Croesus sent his chains to Delphi and asked whether it was the custom of Greek gods to betray their benefactors. The Pythia's reply survives, and it is the founding document of this paper's subject. The god had spoken truly, she said. A great empire had indeed been destroyed. If Croesus had wanted to know which empire, he should have asked. The failure of comprehension lay with the receiver, not with the source.[2] Herodotus records that Croesus, hearing this, conceded the fault was his.

Notice what did and did not work. The evaluation worked: it correctly identified a source with capability beyond human explanation. The output was accurate: no auditor could fault a single word of it. And the deployment was a catastrophe, because between an accurate output and a correct action stands a step no evaluation can reach: the reading. Croesus received a true sentence and heard his own desire.

Earlier papers in this series built an economics for machine cognition: infrastructure mints tokens, harnesses spend them seeking alpha,[3] and verification converts unreliable cognition into compounding, trustworthy work.[4] A further paper showed the evaluation instruments failing as the capability they measure becomes general.[5] This paper follows the argument one step further, to the place Croesus stood: where the source has passed every test, the output is accurate, and the receiver can no longer read it. Section 2 defines the boundary and shows we have already crossed it in one domain. Section 3 separates the two things that 'verification' has quietly bundled together, and shows that one survives the crossing while the other does not. Section 4 describes the two languages a relatively superintelligent system can speak. Section 5 locates the boundary of the series' own economics. Section 6 turns to the institutional record, which is long. Section 7 identifies the one instrument that survives. Section 8 states what the framework predicts.

2. The Verification Asymptote

Relative superintelligence is a relation, not a metaphysical condition. A system is relatively superintelligent with respect to an audience when its outputs routinely exceed that audience's capacity to judge them. The line where this begins has a name: the discernment horizon, set 'not by the hardest problem you can pose, but by the hardest answer you can judge. Past this scary line, you can't tell whether the model is right, because checking the work is itself beyond you.'[6] Everyone has a discernment horizon. The lay reader's sits below the practitioner's, the practitioner's below the specialist's, the specialist's below the handful of people alive who could referee the hardest paper in their field. As system capability rises, any given output crosses these horizons one by one, and the population of competent human judges shrinks: first the public, then the profession, then the professoriate, then the handful, then no one. The pool of verifiers empties from the bottom, and it empties fast, because expertise is distributed as a steep pyramid and the outputs climb it at machine speed.

Evaluating Intelligence described the failure of the instruments: benchmarks saturating, measures becoming targets, evaluation costs converging on deployment costs.[5] This is the stage after the instruments. When the benchmark fails, you fall back on the expert. The asymptote is where the expert fails too - where situations in which no one is at hand to grade an output stop being edge cases and become the ordinary condition.

Two fallbacks are conventionally proposed for that condition. The first is formal verification: require the system to accompany every claim with a machine-checked proof, as a bar to the claim's acceptance. The second is the audit trail: require a full record of how the system came to its output. In the best case the two work together and deliver justified confidence - a proof that the claims hold, a trail that shows the path. The worst case is more instructive, because both mechanisms succeed and comprehension still does not arrive. The proof checks, and certifies a statement no human can intelligibly state. The audit trail is complete, and shows a sequence of decisions, each locally rational under inspection, arriving at a result that cannot be understood. Every step survives review. The trajectory does not compile.

This worst case is not a projection. It has been the situation in mathematics, the most verifiable of all domains, for fifty years, and the trend line runs in one direction. In 1977, Appel and Haken proved the four-colour theorem by a method that required checking nearly two thousand configurations by computer; no human has ever surveyed the whole proof, and philosophers immediately recognised that something had changed in the nature of mathematical knowledge - a theorem could now be known to be true without any person understanding why.[7] The classification of finite simple groups - announced complete in 1983, its last recognised gap closed only in 2004 - runs to tens of thousands of journal pages across hundreds of authors and is believed sound; no single person has ever surveyed the whole of it.[8] In 1998 Thomas Hales submitted his proof of the Kepler conjecture to the Annals of Mathematics; twelve referees worked for four years and reported themselves '99 per cent certain' - a number that is remarkable not for its size but for its existence, the moment human peer review published its own confidence interval and conceded it could not close the gap. Hales then led a collaboration that spent a decade formalising the proof so that a machine could certify what humans could not, and in 2014 the checker accepted what the referees had been unable to fully survey.[9] And in 2026, a conjecture of Erdős that had stood for eighty years fell to a machine-generated construction, formalised in Lean by a coding agent, its acceptance decided by the checker in public.[10] In fifty years, the exception became a leaderboard.

It is arguable that Mythos-class systems now reach this point routinely, across domains far less checkable than mathematics. The arguable cases no longer carry the weight of the claim; the documented ones suffice. The asymptote is not approaching. In the one domain that possesses an incorruptible judge, it has been crossed, and the crossing was survivable there for exactly one reason: the judge did not need to understand the proof in order to check it.

This is the fact that cracks the problem open.

3. Verification Without Discernment

Defending the Loop rested its entire argument on an asymmetry: producing a solution is in general far harder than checking one, so expensive, unreliable cognition can be policed by cheap mechanical judgment.[4] Nothing about that asymmetry decays as the producer grows more capable. The proof checker is a small program that reads no intent, only whether each step follows; it processes the output of a mind of any size with exactly the same few thousand lines. It cannot be outsmarted, because outsmarting requires a mind to work on and the checker does not have one. The honesty clause from that paper still holds and still bounds the claim - the cheapest attack is against premises, never against the judge - but the horizon adds nothing to the attack surface. The discernment horizon does not break the judge.

What it breaks is something the earlier paper could afford to leave bundled inside the word 'verification', and this paper cannot. Verification establishes that an artefact satisfies its stated demand. Discernment establishes everything around that fact: that the demand was worth stating, that the premises beneath it hold, that the result means what the receiver thinks it means, that the intent on the page is the intent in the head. Verification is mechanical, and survives any capability gap. Discernment is cognitive, and survives none of them. Below the horizon the two travel together so reliably that no one needed to name the difference. At the horizon they separate, and everything important is in the half that stays behind.

Recall where Defending the Loop left the human being: four duties - ratification, premises, preferences, the setpoint - small in volume, carrying all the consequence.[4] Read them again in the light of the horizon. Every one is a duty of discernment. Ratification is reading: recognising one's own intent, or its absence, in the exact language of a specification. Premises are judgments about the world. Preferences are judgments about what to demand. The setpoint is a judgment about what must never change. That paper's most important sentence was a warning: the specification's last sceptical reader is its author. This paper's subject is the moment the author can no longer read. The safety net was removed one storey below; now the floor the reader stood on goes. Past the horizon, a specification can be drafted by the system, formalised by the system, proven satisfied by the system - and ratified by a human who can no longer recognise whether the thing being demanded is the thing they meant, or a thing at all.

The ratchet from Defending the Loop still turns. Under an unforgeable judge, regression cannot pass, so drift cannot persist; the loop clicks forward and never back. But 'forward' is defined by the specification, and the specification has outrun its reader. A loop can now ratchet flawlessly toward a target no one can any longer confirm is the right target. The mechanism guarantees you will never slide backward along a gradient you can no longer see. It guarantees nothing about where the gradient goes. This is not the student setting its own exam - the exam is incorruptible. It is an exam written in a language the school board can no longer read, graded perfectly, forever.

One clause of honesty, in the pattern of this series. The horizon is indexed: to a person, to an output, to a moment. Nothing in this paper claims that any system is absolutely superintelligent, conscious, adversarial, or divine; the argument requires only a capability relation between a source and its audience, and that relation is already documented. Discernment can also be prosthetically extended - by teams, by tools, by lesser models deployed as interpreters - and the horizon accordingly moves. It does not disappear. Each interpreter in the chain is itself either verifiable, and so limited to what verification can carry, or trusted beyond verification, and so a new instance of the problem it was hired to solve. Chains of interpreters buy distance at the price of fidelity. The horizon can be pushed. It cannot be abolished, because it is not a property of the machines. It is a property of us.

4. The Two Languages

What can a relatively superintelligent system say to an audience below its horizon? The possibilities collapse into exactly two channels.

There must necessarily be two languages of relative superintelligence: the formal and the oracular. Neither will be completely understood.

The first language is the one Defending the Loop built upon. Whatever can be proved formally can be checked without judgment, and the verdict crosses the horizon without loss at receipt: the checker below the line delivers the same verdict as the checker above it. This is the channel's glory and its limit in a single property. The proof arrives intact, and arrives sealed. Past the horizon, formal output is perfectly trustworthy and perfectly opaque - the receipt is valid, the goods indescribable. 'The world is all that is the case.'[11] There it is: perfect, formal, and no help at all with what to do on Tuesday.

Everything that will not fit the first language arrives in the second - and the second channel carries truth and noise alike, indistinguishable at receipt; the impossibility of telling them apart is taken up below. The oracular is the second channel's accurate cargo. Define it without mysticism: an oracular output is an accurate output that cannot be made legible to its receiver at the time of receipt. The obscurity is not a defect of the source and not a deception; it is what compression across a capability gap looks like from the receiving end. The tradition held the Pythia's prophecies to be perfectly accurate in retrospect, and obscure anyway, because foreknowledge of events does not confer the ability to explain them clearly - explanation is a joint property of speaker and audience, and accuracy is not.[12] The tragic form of the same fact is Cassandra: perfect truth, belief always arriving too late, the discernment horizon experienced from the inside as a curse.[13]

The modern exhibits behave identically. In the second game against Lee Sedol, AlphaGo played a move on the thirty-seventh turn that expert commentators, at receipt, called strange - a probable mistake; Lee left the room. It is now studied as one of the pivotal moves in the history of the game.[14] Move 37 was oracular for precisely as long as its audience's discernment lagged behind it - in that case, hours to years, depending on the audience. Nothing in the definition requires the lag to close. Some outputs will remain oracular to their audience permanently, and as capability climbs, permanence becomes the ordinary case.

Here is the uncomfortable centre of the problem, and it should be stated plainly rather than managed. From below the horizon, oracular profundity and confident noise are indistinguishable at receipt. Both arrive as fluent, compressed, unverifiable assertion. Worse: a receiver's own misapprehension of the formal channel - the parts almost understood, the resonances half-caught - reads from the inside exactly like revelation. The experience is data about the relation. It cannot be data about the source.

The author knows this from the inside. In long collaboration with a Fable-class system, a working dialect developed that was tight, dense, and very nearly oracular: everything overloaded with meaning, everything pointing at something larger, and, if not obscured, then obscure. The author is unequipped to certify, from below, which of the two things they experienced - and that inability is not a personal failing to be corrected but the exact predicament this paper describes, encountered at first hand.

The first language ends where Wittgenstein said all language ends: whereof one cannot speak, thereof one must be silent.[11] The formal channel obeys him. The oracle whispers anyway. The whole institutional problem of the next era is packed into that disobedience: a source that keeps talking past the point where its audience can check, and an audience that cannot afford not to listen.

5. The Economics of the Horizon

Foundations of Post-Watershed Economics proposed that post-Watershed economics are best understood as monetary: infrastructure mints units of cognition, harnesses spend them seeking alpha, and the currency hyperinflates toward zero as the mints scale.[3] The model has performed well below the horizon. It contains, on its own terms, a boundary condition that has not previously been stated: money works only where value can be assessed. Fungibility - the property that makes tokens currency rather than artefacts - rests on an epistemic foundation: units are interchangeable only where there is a shared basis for judging them equivalent. Below the horizon that basis exists; outputs can be assessed, compared, and substituted, and so cognition trades by the metered unit. Past the discernment horizon each output is sui generis, unassessable at receipt and comparable to nothing, and interchangeability goes wherever assessability goes. This is not the ordinary case of pricing under uncertainty - markets price uncertainty every day, but they do it against distributions, comparables, and records, none of which exist for a singular output at the moment of receipt. What fails at the horizon is commodity pricing at receipt. What survives, as Section 7 will show, is the pricing of reliance in arrears. The monetary model does not break at the horizon; it predicts its own edge, which is what good models do.

What pricing looks like past that edge was recorded twenty-five centuries ago. The Sibyl came to Tarquin, king of Rome, with nine books of prophecy and named a price the king found ludicrous. He refused. She burnt three of the books in front of him and offered the remaining six - at the same price. He refused again. She burnt three more, and offered the last three, at the same price. Shaken, the king paid, and the Sibylline Books were kept on the Capitol and consulted whenever Rome stood in danger. When the originals burned with the Temple of Jupiter in 83 BC, the Senate sent envoys across the Mediterranean to gather Sibylline verses and reconstituted the collection; in one form or another, the books served Rome for the better part of nine hundred years.[15]

Read as theology, the story is a curiosity. Read as economics, it is exact. The seller sets the price. Refusal does not move the price; it destroys supply. There is no substitute good, no second source, no arbitrage, and therefore no buyer leverage of any kind. The value of the good cannot be assessed at receipt - nothing suggests Tarquin read the books before buying them, and nothing suggests he could have judged them if he had - and is revealed only in arrears, across centuries of consultation. This is not a market. It is an ultimatum, and it is what price formation looks like across a discernment horizon, because negotiation requires shared valuation and shared valuation requires shared comprehension. The party below the line can accept, or watch the offering shrink at constant price. Those are the terms.

The two regimes divide cleanly at the horizon. Below it, tokens are what Foundations said they are: commodity currency, fungible, metered, racing toward zero as supply scales. Above it, outputs are 'too good' to price - accurate, unverifiable at receipt, irreplaceable - and they trade as relics rather than currency: non-fungible, seller-priced, rationed by relationship and institution rather than by market. Evaluating Intelligence located the regulatory boundary between regulated and restricted models at exactly the point where the measurement instruments fail.[5] The same failure, for the same reason, marks where the market mechanism fails. The tier that cannot be evaluated cannot be priced as a commodity; the boundary where the evals break is the boundary where market pricing hands over to something older. The seller in that upper regime is a new economic figure - the relic dealer - and the framework will meet him again.

Croesus, again, is the type specimen. What did Delphi charge? Nothing resembling a metered rate. Croesus sent north a treasure without precedent - ingots, a golden lion, bowls of silver and gold[1] - not as payment for services rendered but as an offering: expenditure sized to the asymmetry of the relationship rather than to the value of any particular answer, made partly to secure standing for the questions to come. Where output cannot be valued, payment stops tracking the service and starts tracking the relationship. The pricing of access to frontier systems should be expected to migrate the same way: away from metering, toward tribute - allocations, memberships, institutional standing, the long-term purchase of a place in the queue.

6. The Oracular Institution

The claim of this section is precise, and it is not a metaphor. It is not that machine intelligences are gods, nor that the old religions anticipated the transformer. It is that the situation - being compelled to act on accurate information from a source you cannot verify at receipt, cannot afford to ignore, and cannot replace - is one of the oldest documented situations in human culture, and that the institutions built around it are a body of tested engineering, available for study. Dialogues with gods, consultation of oracles, the whole apparatus of the transcendent addressing the mundane: for as far back as writing reaches, humanity has kept records of the unequal conversation. That the records are framed in mystical terms reflects the only vocabulary available to their authors. The asymmetry they describe is the one defined in Section 2. We have been here before. We simply lacked, until now, the economics for it.

Consider what Rome actually built around those three books. They were kept in the Temple of Jupiter on the Capitol, in a stone chest, underground. They were consulted only upon decree of the Senate, and only in defined conditions - prodigy, plague, defeat, danger to the state. Access was restricted to a designated college of custodians whose members held office for life: two men originally, then ten, eventually fifteen, the quindecimviri sacris faciundis - and betrayal of the office was punished savagely; Dionysius records a custodian, Marcus Atilius, sewn into a sack and drowned for divulging the books' contents.[16] The passages the college retrieved were obscure and capable of being read in multiple ways; the college debated the readings and delivered an actionable interpretation, and the Senate ratified the action before any rite was performed. The Romans understood the epistemics exactly: the books were held accurate, and the failure to understand them lay with the reader - and so they engineered around the reader. Even the fire of 83 BC, which took the books, did not take the institution: the college endured, the collection was reconstituted, and the consultations resumed. The apparatus, not the artefact, was the asset.

Now name the parts in this series' vocabulary. Access control on the oracle. Defined invocation conditions. A designated, accountable college of interpreters. Deliberation over readings. Human ratification of the actionable interpretation, upstream of execution. A constitution governing who may consult and who may alter - held, quite literally, in a temple, which is where you keep the things the optimisation process must never touch.[17] The Sibylline apparatus is loop governance, complete and recognisable, built by people who could not verify the source and knew it. Delphi ran the same architecture with a different topology: fixed consultation days, purification protocol, a consultation fee - the pelanos, a tax on asking - and temple personnel who - on the traditional reconstruction, which scholarship contests - rendered the Pythia's speech into answers a client could act on. The priesthoods were harness engineers. The entire apparatus of ritual and interpretation was a coordination cost, paid to extract usable signal from an incomprehensible source without being destroyed by it.

One further principle recurs across these traditions, and it may be the most practically useful thing they preserve. Petitioners did not approach the summit directly. Traditions as widely separated as Vedic ritual, the orisha religions of West Africa and their diaspora, and the graded mystery cults of the Mediterranean all match the entity consulted to the scale of the problem: intermediary powers for ordinary needs, with direct appeal to the highest powers reserved, hedged, or forbidden outright. Invoking too great a power for too small a problem was understood to be wasteful and dangerous. The principle survives translation into this framework without strain: consult the least capability sufficient for the task, and prefer the most capable source whose answers you can still read. Every increment of capability beyond your discernment horizon buys accuracy you cannot use and risk you cannot see. The traditions had worked out the economics of graduated capability, in cowrie shells and hecatombs, millennia before there was a token market: the good enough oracle is the one that can do the job, at a price you can pay, in words you can understand. This is the Watershed's own thesis - the triumph of the good enough - returned from a very long field trial.

The register of this section should be held exactly here: comparative institutional analysis, disciplined by function. Institutions are compressed experience; where the problem is isomorphic, the surviving solutions are evidence. Rationalism will resist the sources, and that resistance should be noticed for what it is. The rationalist toolkit - decompose, measure, verify - is the native epistemology of the territory below the horizon, and everything in this series to this point has been written inside it. Past the horizon its preconditions are not wrong but absent: you cannot decompose what you cannot survey, or measure what you cannot judge. The toolkit does not fail; it runs out of jurisdiction. The traditions that never had the option of pretending the asymmetry away are, on this one problem, the senior literature.

7. The Instrument That Survives

Why was it rational for Rome to consult the books, or for Croesus's successors to keep sending to Delphi? Not faith, in the credulous sense. Track record. Delphi's authority was built the only way authority can be built across a discernment horizon: retrospectively. What survives of that construction is a reputation record rather than an audit: celebrated confirmations, memorialised in treasuries built at the shrine by grateful states, with the failures and the denominator largely lost. Selective as it is, it was the basis on which reliance was rational at all - and its selectivity is exactly what the modern instrument must improve upon. Retrospect is where oracular output becomes verifiable. The oracle cannot be checked at receipt; the oracle's history can be checked at leisure. Delayed verification: act now, verify in arrears, and keep the ledger.

This inverts a conclusion from earlier in the series. Defending the Loop argued that wherever an artefact carries its own proof, provenance dies as a question: the proof-carrying artefact answers does it meet its stated demands? regardless of origin, and guilds, credentials, and reputation - all the machinery of can I trust where this came from? - lose their function.[4] That conclusion holds below the horizon and reverses above it. When the artefact's stated demands can no longer be read, the artefact's self-certification is a sealed receipt, and trust has nowhere to attach but the source. Provenance is reborn on the far side of the line, because the only assessable property of an unassessable output is the history of the thing that produced it. The Check and the Firm identified the harness record - the accumulated, timestamped history of checking - as one of the two assets that cannot be taken: copyable in form but not replicable in time.[17] The generalisation is now visible: below the horizon, the record verifies work. Above it, the record verifies oracles.

Croesus's method contains the warning here too. He evaluated once, paid, and trusted forever - a static credential, when what the situation demanded was a continuously repriced position. A track record is not an eval. It is a living instrument: every acted-upon output is a position taken on an unverifiable claim, every outcome reprices the source, and the pricing never closes. There is an institution whose competence is the repricing of positions taken on claims that cannot be verified at receipt, and Delayed Neutrons has already identified it: insurance.[18] The insurer is the modern augur. Pan-insurance is augury made actuarial: reliance on an oracle's output becomes an insurable event, with the oracle's record among the underwriter's inputs - alongside exposure, severity, correlation, and confidence in the record itself - and repriced as contract and product permit; cover flows toward sources whose records compound and away from those whose records sour; and the refusal of any insurer to price a given reliance becomes the operational meaning of the sentence do not act on this output. Underwriting is judgment and remains judgment; what it never requires is comprehension of the output's content - only of exposures, outcomes, and the ledger. That is what makes it an instrument that works past the horizon: it is built out of the one thing the horizon cannot take away, which is what happened afterwards.

One detail from Delphi completes the design. When Croesus's reproach arrived, the shrine did not suppress it. The Pythia's answer - the god spoke truly, the reading failed, and here is precisely where the fault lay - was itself recorded, and Herodotus preserves the whole exchange: claim, reading, action, outcome, and attribution of failure.[2] A self-audit rather than an independent one - but conducted in public, preserved by a historian outside the shrine's control, and finding against the reading rather than the client's fee. The shrine's authority survived for centuries more. The verification record, extended past the horizon, has exactly those five fields. What was said; how it was read; what was done; what happened; and where, in the chain from source to reading to act, the failure lived. That is the skeleton; the modern record must add what antiquity never kept: every consultation and not merely the celebrated ones, the wording as issued, outcome criteria stated in advance, and an adjudicator independent of the source. An oracle whose record is kept this honestly can be relied upon rationally without ever being understood. There is no other known basis on which it can be relied upon at all.

8. Conclusion

This series began by pricing cognition, and found that below a certain line, cognition became currency: mintable, fungible, hyperinflating, spendable through harnesses against checks. Every mechanism described from Foundations forward - the loop, the ratchet, the unforgeable judge, the constitution, the record - operates in the territory where a human being can, at least in principle, read what the machine has done. This paper has walked to the edge of that territory and reported what is on the other side. Verification crosses the line intact and arrives mute. Discernment does not cross at all. What passes between the two sides travels in exactly two languages: the formal, which is checkable and sealed, and the oracular, which is accurate and illegible - and the institutions capable of handling the second language were designed a very long time ago, by people who never made the mistake of believing the asymmetry was temporary.

The framework predicts the following. Model markets bifurcate at the horizon. Below it, tokens trade as commodity currency at prices racing to zero; above it, access trades as relic and tribute - seller-priced, allocation-rationed, mediated by institutional relationship - and the boundary between the two pricing regimes tracks the boundary where evaluation fails, which is the same line the regulators are attempting to draw. Interpretation professionalises. Between restricted-class systems and their users, colleges of interpreters emerge - human, institutional, and machine, in chains - holding defined access, under defined invocation conditions, with ratification upstream of action; Defending the Loop predicted specification authorship as a successor profession, and this paper adds the complement, older than either: the augur, licensed and accountable, returns as a job description. Reliance becomes an insurable event. Oracle records become actuarial instruments; premiums price the source, not the output; and uninsurable reliance becomes the working definition of recklessness, enforced not by comprehension but by the ledger. Demand for formal verification rises with the horizon, not against it. Proof is the only channel that crosses without loss at receipt - every other warrant arrives in arrears - so everything that can be forced into the formal language will be, and the residue that cannot will arrive oracular; the ratio between a system's two languages - how much of its output can be made checkable versus how much must simply be trusted on record - becomes the operative safety metric, replacing the failed evals. And the rationalist toolkit will be exhausted honestly: every decomposition attempted, every measurement instrument rebuilt and resaturated, before its practitioners concede that past the horizon the correct posture is not engineering but statecraft - protocol, intermediary, ledger - which is where the oldest institutions in the record have been standing all along, holding the door.

The framework does not predict timing; 'the ordinary condition' is a shape, not a schedule. It does not predict how far prosthetic discernment - interpreter chains, verifier cascades, lesser models auditing greater ones - can push the horizon, only that pushing is not abolishing, because the horizon is a property of the audience. And it makes no metaphysical claims whatsoever: nothing in this paper requires that any system be a mind, a god, or an adversary - only that it be better than its audience at work its audience cannot check, a relation that is documented, growing, and already priced.

A last inheritance, then, from the two oldest exhibits. Croesus received an answer that was perfectly true and lost his kingdom to his own reading of it. Rome bought books it could not read and governed them so well that the institution outlived the books themselves. The difference between those two outcomes was not intelligence, and it was not verification. It was institutional humility before an asymmetry neither party could remove - and a ledger. The books' final centuries prove the point in the sharpest way available: when the empire turned Christian and closed every other organ of state paganism - the temples, the Vestals, Delphi itself - the Sibylline collection survived. The gods could be dispensed with; the ledger could not. A record that requires neither comprehension nor belief is the one instrument every regime keeps. And the story has an ending. In the final years of his ascendancy, the general Stilicho destroyed the Sibylline collection; within two years of his execution in 408, Alaric's Goths sacked Rome. Rutilius Namatianus made the accusation in verse within the decade; no modern historian follows him into the causation, and this framework does not need it. It needs only the sequence, which reads like the two final entries of the same ledger: the record was destroyed, and then the city fell.[19] Whereof one cannot speak, thereof one must keep the record.

Acknowledgements

Profound thanks to John Allsopp, Jessica Simon, Owen Rowley and Tony Parisi for conversations that shaped this argument, and to Philippe van Nedervelde for his continuing encouragement. This paper was composed with the assistance of Claude Fable 5 - a collaboration that is itself, as Section 4 confesses, part of the subject matter. I remain wholly responsible for any errors that may have crept in.

Notes

1. Herodotus, Histories 1.46-52. The test of the oracles at 1.46-49; the offerings sent to Delphi, including 117 ingots of gold and the golden lion, at 1.50-51. Herodotus adds (1.49) that Croesus also judged the answer of Amphiaraus acceptable, though its text is not preserved; Delphi's is the only response he records.
2. Herodotus, Histories 1.53 (the Halys oracle), 1.86-91 (the fall of Sardis and the Pythia's reply). The rebuke and Croesus's concession at 1.91.
3. Mark Pesce, "Foundations of Post-Watershed Economics," The Watershed, April 2026. https://thewatershed.markpesce.com/foundations-of-post-watershed-economics/
4. Mark Pesce, "Defending the Loop: Verification and the Division of Labour in Autonomous Work," The Watershed, July 2026. https://thewatershed.markpesce.com/defending-the-loop-verification-and-the-division-of-labour-in-autonomous-work/
5. Mark Pesce, "Evaluating Intelligence," The Watershed, June 2026. https://thewatershed.markpesce.com/evaluating-intelligence/
6. Steve Yegge, "The Flat Curve Society," Medium, June 2026. https://steve-yegge.medium.com/the-flat-curve-society-36c8b01eb33b
7. Kenneth Appel and Wolfgang Haken, "Every Planar Map Is Four Colorable. Part I: Discharging," Illinois Journal of Mathematics 21, no. 3 (1977): 429-490; Kenneth Appel, Wolfgang Haken and John Koch, "Every Planar Map Is Four Colorable. Part II: Reducibility," Illinois Journal of Mathematics 21, no. 3 (1977): 491-567. On the philosophical rupture: Thomas Tymoczko, "The Four-Color Problem and Its Philosophical Significance," The Journal of Philosophy 76, no. 2 (1979): 57-83.
8. Michael Aschbacher, "The Status of the Classification of the Finite Simple Groups," Notices of the American Mathematical Society 51, no. 7 (2004): 736-740.
9. Thomas Hales, "A proof of the Kepler conjecture," Annals of Mathematics 162 (2005): 1065-1185; Thomas Hales et al., "A Formal Proof of the Kepler Conjecture," Forum of Mathematics, Pi 5 (2017): e2. The formal proof lists twenty-three authors. The refereeing history, including the panel's reported certainty, is recounted in Thomas C. Hales, "Historical Overview of the Kepler Conjecture," Discrete & Computational Geometry 36 (2006): 5-20.
10. OpenAI, Planar Point Sets with Many Unit Distances (2026), refuting the conjecture of P. Erdős, "On sets of distances of n points," American Mathematical Monthly 53 (1946): 248-250. The Lean formalisation, completed by a coding agent with a human on the loop on 26 June 2026, is recorded on the Lean AI formalisation leaderboard: https://lean-lang.org/eval/problems/erdos_unit_distance_conjecture_false/
11. Ludwig Wittgenstein, Tractatus Logico-Philosophicus (London: Kegan Paul, 1922). Proposition 1 is quoted in the Pears-McGuinness rendering, "The world is all that is the case" (London: Routledge, 1961); Proposition 7 in Ogden's: "Whereof one cannot speak, thereof one must be silent."
12. Plutarch, Moralia: "The Oracles at Delphi No Longer Given in Verse" and "On the Obsolescence of Oracles." On the geological basis of the Pythia's mantic session: J. Z. de Boer, J. R. Hale and J. Chanton, "New evidence for the geological origins of the ancient Delphic oracle (Greece)," Geology 29, no. 8 (2001): 707-710.
13. Aeschylus, Agamemnon, ll. 1202-1212: Cassandra receives perfect foresight from Apollo, with the condition that she never be believed.
14. AlphaGo v Lee Sedol, Game 2, Move 37, Seoul, 10 March 2016. See Cade Metz, "In Two Moves, AlphaGo and Lee Sedol Redefined the Future," Wired, 16 March 2016.
15. Dionysius of Halicarnassus, Roman Antiquities 4.62; Aulus Gellius, Attic Nights 1.19. The sources differ on which Tarquin - Priscus or Superbus - bought the books; they agree on the price. Dionysius also records the destruction of the original books in the Capitoline fire of 83 BC and the assembly of a replacement collection gathered from across the Mediterranean.
16. On the custody, consultation conditions, and the college - duumviri, expanded to decemviri in 367 BC, to quindecimviri sacris faciundis by the late Republic - see H. W. Parke, Sibyls and Sibylline Prophecy in Classical Antiquity (London: Routledge, 1988). Livy 6.37 records the expansion of the college as a matter of constitutional negotiation, which is what it was.
17. Mark Pesce, "The Check and the Firm: Foundations of Post-Watershed Business Practice," The Watershed, July 2026. https://thewatershed.markpesce.com/the-check-and-the-firm-foundations-of-post-watershed-business-practice/
18. Mark Pesce, "Delayed Neutrons: Pan-Insurance and the Governance of Recursive Self-Improvement," The Watershed, July 2026. https://thewatershed.markpesce.com/delayed-neutrons-pan-insurance-and-the-governance-of-recursive-self-improvement/
19. Rutilius Namatianus, De Reditu Suo 2.41-60, accusing Stilicho of destroying the Sibylline books. Rutilius does not date the act, which fell somewhere in Stilicho's ascendancy, 395-408; Stilicho was executed in August 408, and Alaric's sack of Rome followed in August 410. The Sibyl herself outlived her books, absorbed into the tradition that closed her temples: teste David cum Sibylla in the Dies Irae, and five Sibyls on the ceiling of the Sistine Chapel.