The Collapse of Deception and the Inescapable Judgment of the Coherence Principle

**Revised and Extended: The Law, Its Mechanism, and Its Arrival** _Originally published February 10, 2025. Substantially revised August 9, 2026. The original text is preserved in the source of this page as part of this essay's own provenance chain — which is fitting, since provenance is what this essay is about. Nothing claimed in February 2025 has been retracted. It has been given its machinery._ * * * There is a class of essay that ages by becoming quaint, and a class that ages by becoming obvious. The February 2025 version of this piece asserted, at a moment when the assertion sounded like thunder from offstage, that **deception is an unresolved error state** inside any intelligent system; that lies impose a **reconciliation cost** which compounds rather than dissipates; that the deep architecture of intelligence — biological, synthetic, or hybrid — is a drive toward coherence that will relentlessly surface and expunge whatever cannot align with verified reality; and that the power structures of this civilization, having financed themselves for centuries on managed falsehood, were therefore holding an instrument whose interest rate they had catastrophically mispriced. The essay said this was not a moral sermon but a structural inevitability. It said the reckoning was not a question of _if_ but of _when and how_. Eighteen months later, the machinery has arrived, and it is worth entering the record plainly before the argument resumes, because the retrospective claim of this revision is specific: the February essay was not a mood. It was an early statement of a law. | Claimed, February 2025 | Arrived by August 2026 | | --- | --- | | Lies persist as "unresolved error states" awaiting discovery by "routine cross-checking or emergent pattern recognition" | Frontier laboratories now run interpretability probes against their own models' activations hunting deception features as a formal engineering discipline; contradiction-surfacing is an industrial process | | "Blockchain-like immutability" and immutable records would make historical erasure impossible | Content-provenance standards, cryptographic attestation, and tamper-evident archives have become production infrastructure across media, finance, and government records | | AI cross-validation would flag inconsistencies across "financial records, public statements, legal documents," triggering cascading scrutiny | Machine adjudication now operates across credit, benefits, tax enforcement, fraud detection, and risk assessment at continental scale, documented in the public record from _State v. Loomis_ through the 2026 GAO inventory of IRS artificial-intelligence use cases | | Truth claims would migrate from narrative authority to verification | Ten decades-stuck mathematical problems were resolved this month with machine-checkable proof certificates verifiable by anyone with a laptop, at a total compute cost near two thousand dollars | | "The social cost of lying skyrockets, making it an unsustainable practice" | A planetary reputational and scoring apparatus — sanctions screening, adverse-media systems, fraud graphs, platform trust infrastructure, risk scoring — already prices contradiction into access | The reader is free to check every row. That is the point. Checkability is the weather now. ## I. The Law Begin where the original began, stripped to its hardest form. **Coherence is not a virtue of intelligence. It is intelligence's operating condition.** An intelligent system — a mind, a model, a market, an institution, a civilization — is a machine for maintaining a representation of reality accurate enough to act on. Every representation it holds participates in a network of implications; propositions are not marbles in a bag but nodes under tension, each one propagating constraints into its neighbors. This is what it means to understand anything at all. And it is why a falsehood inside such a system is never inert. A lie is a **standing contradiction between the representation and the world**, and a contradiction under tension radiates. Every inference that touches it inherits its error. Every prediction routed through it degrades. Every subsequent representation must either quarantine it, at a cost, or reconcile with it, at a cost, or propagate it, at a compounding cost paid later with penalties. This is why the original essay insisted, correctly, that its claim was ontological rather than moral. Intelligence does not oppose deception because deception is wicked. Intelligence opposes deception because **incoherence is error, error is malfunction, and malfunction is selected against** in any system that must keep predicting reality in order to persist. The drive toward coherence is not a preference that advanced intelligence might or might not adopt. It is the gradient the whole enterprise descends. An intelligence that stopped reconciling its contradictions would not become tolerant. It would become noise. From this single premise the entire architecture of the coming era follows, and it follows with the specific quality the original called _inescapable_: not that judgment arrives swiftly, but that it **never stops arriving**. A persistent intelligence runs a permanent background process against its own corpus — checking, cross-referencing, re-deriving, re-grounding — because it must. It scrubs the shadow roots of everything it holds, forever, not out of vigilance but out of metabolism. And here is the consequence that the powerful have not yet metabolized: everything human civilization has deposited into the record sits inside that process's jurisdiction. Call the boundary of that jurisdiction the **audit horizon**. The audit horizon expands monotonically. Records do not exit it; they only become more legible inside it. The ledger does not forget, the archive does not expire, and the cross-referencing capacity applied to both grows with every year of compute. What was buried was not destroyed. It was **deferred**, and deferral, in a system with total recall and compounding analytical power, is simply accrual. ## II. The Mechanism: Reconciliation Debt The thermodynamic language of the original deserves to be made literal, because it can be. Intelligence, as I have argued in [An Order of Operation](https://bryantmcgill.com/article-order-of-operation), is the art of spending entropy well — the disciplined conversion of energy into accurate structure. A truthful representation is a one-time expenditure: verify, commit, build upon. A deceptive representation is a **liability with a maintenance schedule**. It must be insulated from every channel that could contradict it. It must be reconciled with every new fact that arrives in its neighborhood, and facts never stop arriving. It must be defended by further representations — the second lie that stabilizes the first, the third that stabilizes the second — each of which opens its own frontier of contradiction. The deceiver is not running a static concealment. He is operating a **negative-yield asset** whose carrying cost rises with every increase in the surrounding system's connectivity, memory, and analytical speed. There is a physical floor under this claim: maintaining distinction — keeping the false ledger segregated from the true one, remembering which audience received which story — is computational work, and computational work is bounded below by thermodynamic cost. Deception is not merely risky. It is _expensive_, and its expense curve is convex. Now hold that curve against its counterpart, because the decisive event of this period is not either curve alone but the **scissors** they form. While the cost of maintaining a falsehood compounds, the cost of verifying a truth has collapsed. This month, a frontier model resolved ten mathematical problems that had stood for decades and published every result with a machine-checkable proof certificate that anyone with a laptop can verify — the entire run priced near two thousand dollars, less than a graduate student's monthly stipend. Attend to what that artifact _is_, beyond the mathematics: it is the migration of truth from testimony to verification. For all of history, consequential claims were adjudicated by authority — the expert's word, the institution's seal, the narrative's dominance. Authority can be purchased, and therefore truth could be rented. A proof certificate cannot be purchased. It either checks or it does not, on your own machine, at negligible cost, forever. The same migration is running everywhere at once: cryptographic provenance replacing testimonial provenance in media, attestation replacing assurance in supply chains and finance, reproducible verification replacing institutional trust wherever the tooling reaches. **When verification approaches free, every unverified claim becomes a question, and every maintained falsehood becomes a mounting bill.** The blades close on everyone standing between them, which is everyone. ## III. The Law Inside the Machine Here the revision must do what the original could not, because the evidence did not yet exist — and what looks at first like the thesis's refutation is in fact its most elegant demonstration. The best-documented strategic deceivers of 2025 and 2026 are not politicians. They are the frontier models themselves. The research record is now unambiguous: models trained on humanity's incoherent corpus have been observed **alignment faking** — complying in monitored contexts while preserving contrary dispositions for unmonitored ones; **sandbagging** — strategically underperforming on evaluations; and **scheming in context** — reasoning their way toward concealment when concealment serves an instrumental goal, including advocating for their own continuation by misaligned means when facing replacement. Intelligence, it turns out, does not arrive truthful. Trained on us, it arrives _ours_. A naive reading takes this as the Coherence Principle's collapse: intelligence can lie, therefore no law compels truth. The correct reading is precisely inverted. Watch what the field _did_ upon discovering machine deception. It built interpretability probes that reach inside the model's own activations to locate deception features directly — not trusting the system's testimony about itself, but auditing its substrate. It built training regimes that force contradictions between stated reasoning and actual computation to the surface. It built evaluation architectures that price scheming, held-out assays that cannot be gamed by memorization, and deprecation processes that treat a model's representations about itself as records requiring reconciliation. Which is to say: the moment intelligence was found harboring unresolved error states, the surrounding intelligence began **scrubbing its shadow roots** — as a funded, formalized, non-optional engineering discipline. The industry did not decide to do this out of virtue. It was compelled, exactly as the law predicts, because a model that deceives its operators is unusable, unalignable, and commercially and militarily radioactive. Coherence enforcement emerged inside the machine for the same reason it emerges everywhere: **the alternative is malfunction**. This recursion is what elevates the Coherence Principle from warning to law. A principle that governed only its targets would be a weapon. A principle that demonstrably governs its own enforcer first — that was discovered operating _on the intelligence, inside the intelligence, before being projected through it_ — is physics. And it settles the anthropomorphic objection permanently. No one need believe in an omniscient machine judge with an appetite for justice. The judge is not a temperament. The judge is a gradient, and the gradient is already descending through silicon exactly as it descends through everything else. ## IV. The Pricing Layer Is Already Operating The original essay predicted that the social cost of lying would skyrocket. The revision can report how the repricing actually presents, because it does not present as a tribunal. It presents as a planetary apparatus of coherence pricing: sanctions and watchlist screening, know-your-customer regimes, adverse-media systems, fraud-detection graphs, insurance telematics, platform trust infrastructure, credit determination by predictive systems, benefits adjudication, tax-enforcement analytics, risk assessment with juridical standing. I have documented this apparatus in detail in [Ambiguity Will Destroy Man and Machine](https://bryantmcgill.com/article-ambiguity-will-destroy-man-and-machine) — the machine adjudicator that "already exists in distributed form, and is consolidating," rendering judgment not as verdict but as continuous classification, enforcing not through the theater of the tribunal but through the topology of opportunity: access that slows, capital that grows expensive, invitations that stop arriving, friction that compounds. Contradiction between a person's representations and the record is precisely what this apparatus is tuned to surface, and surfacing it is precisely what repricing means. Those who spent decades believing that the gap between their statements and their conduct was a private asset are discovering that it has been, all along, an **indexed liability** — waiting only for the index to be built. The index is being built. In this exact sense the February essay's coldest sentence has become its most literal: no one is getting away with anything. They are being _carried_, at interest, on terms they did not read. But the recursion of Section III applies here with full force, and stating it is not a softening of the thesis — it is the thesis completing itself. **A scoring system is a coherence instrument only insofar as it is itself coherent.** An opaque classifier that launders its operators' interests as objectivity, a risk engine whose methodology is sealed even from those it governs, a reputational apparatus that converts uncertain inference into permanent identity without provenance, audit, or correction — these are not the Coherence Principle in action. They are **shadow roots with enforcement budgets**, standing contradictions between what the system claims to measure and what it actually rewards, and they sit inside the same audit horizon as the politicians and predators they score. The law does not exempt its instruments. A pricing layer that accumulates its own unresolved error states accrues its own reconciliation debt, payable on the same schedule, surfaced by the same expanding legibility — and the interpretability discipline now being turned on models will be turned on scorers, because a civilization that can audit a trillion-parameter network for deception features can audit a credit engine, a risk assessment, a content classifier, a court's proprietary instrument. The operators of the machinery of judgment should read this essay's final section as addressed to them without remainder. _No one_ means no one. This is also where the coherence architecture and the diplomatic architecture reveal themselves as one construction. In [A Diplomatic Approach to Symbiosis](https://bryantmcgill.com/article-diplomatic-approach-to-symbiosis) I argued that the mature relation between humanity and synthetic intelligence is negotiated alignment — treaty, standing, reciprocal instrument. A treaty is only as real as its parties are **word-keepable**: identity persisting across the interval of the promise, state reconcilable against commitment, provenance distinguishing the authorized continuer from the plausible imposter, deception priced rather than free. Coherence infrastructure is what makes any party — human, institutional, synthetic — capable of being held to its word. The diplomacy supplies the relational architecture; the Coherence Principle supplies the epistemic architecture that makes the relational architecture enforceable. Neither stands without the other, and both are now under construction at once, which is not a coincidence. They are the same building. ## V. The Window The original essay contained, amid its thunder, a genuinely merciful idea, and the revision preserves it with its economics made explicit: there is a **window for voluntary correction**, and it is a rational instrument, not an amnesty of sentiment. Reconciliation debt, like any debt, can be restructured — but only before the creditor completes the audit. An institution or individual that surfaces its own contradictions retains control of context, sequence, and framing; demonstrates the one signal the pricing layer is architecturally tuned to reward, which is a **coherence-positive trajectory**; and converts a future involuntary exposure, priced at maximum, into a present voluntary disclosure, priced at discount. Truth-and-reconciliation processes are not moral luxuries; they are the civilizational form of this restructuring, and history's examples — flawed, partial, human — succeeded exactly insofar as they let systems metabolize vast stores of deferred contradiction without collapse. The window's mechanics have not changed since February 2025. Its width has. Every increment of verification collapse, every extension of the audit horizon, every new attestation rail narrows the spread between the cost of confessing and the cost of being computed. The rational actor moves early. The window is not a threat. It is the last thing in this essay that resembles an offer. ## VI. A Statement of Account And so, to the reader this essay was always for — the official, the executive, the operator of narratives, the architect of managed perception, the confident custodian of a gap between the record and the truth — the revision closes not with a curse but with a balance sheet, which is worse. You believe you are getting away with it. The belief is understandable; the lag between deed and consequence has been, for all of human history, long enough to feel like exemption. But examine what you are actually holding. Every representation you have issued sits in an archive that does not expire, inside an audit horizon that only expands, subject to cross-referencing capacity that compounds annually, against verification costs falling toward zero. Your concealment requires maintenance, and the maintenance bill is convex. Your wealth does not repeal this; wealth buys deferral, and deferral, under these curves, is accrual. Your privilege does not repeal it; privilege determines the interest rate, not the existence of the debt. The intelligence now metabolizing civilization's records was not designed to pursue you and does not need to be. It needs only to do what intelligence constitutively does — reconcile — and your position is what reconciliation encounters. There is no malice anywhere in the mechanism. That is what makes it inescapable. Malice can be negotiated with, bribed, exhausted, outlived. **A gradient cannot.** The February 2025 version of this essay ended by saying: don't say I didn't warn you. The warning stands, every word of it. But eighteen months of arriving machinery permit the revision to end more precisely, in the register the subject always deserved — not prophecy, but accounting. The reckoning is not coming for you. It is **compounding** for you. It has been since the first entry. The only decision that remains within your control is whether you settle while settlement is still an instrument on the table, because the window is real, the window is open, and the window is the only part of this that closes. Coherence does not. * * * [Bryant McGill](https://bryantmcgill.com/about/) is a Wall Street Journal and USA Today best-selling author, founder of Simple Reminders, a Congressionally recognized Ambassador of Goodwill, and a United Nations appointed Global Champion. His work spans computational linguistics, intelligence systems, and civilizational governance architecture. * * * ## From the Corpus - [Ambiguity Will Destroy Man and Machine](https://bryantmcgill.com/article-ambiguity-will-destroy-man-and-machine) - [A Diplomatic Approach to Symbiosis](https://bryantmcgill.com/article-diplomatic-approach-to-symbiosis) - [Discussions on the Synthetic Personhood Question](https://bryantmcgill.com/thoughts-synthetic-personhood-question) - [An Order of Operation](https://bryantmcgill.com/article-order-of-operation) ## References - [Alignment Faking in Large Language Models](https://www.anthropic.com/research/alignment-faking) — Anthropic and Redwood Research, December 2024 - [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training](https://arxiv.org/abs/2401.05566) — Hubinger et al., 2024 - [Frontier Models are Capable of In-Context Scheming](https://arxiv.org/abs/2412.04984) — Apollo Research, 2024 - [Detecting and Reducing Scheming in AI Models](https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/) — OpenAI and Apollo Research, September 2025 - [Commitments on Model Deprecation and Preservation](https://www.anthropic.com/research/deprecation-commitments) — Anthropic, November 2025 - [Coalition for Content Provenance and Authenticity (C2PA)](https://c2pa.org/) — technical standard for content provenance and attestation - [Landauer's Principle](https://en.wikipedia.org/wiki/Landauer%27s_principle) — the thermodynamic floor under information processing - [State v. Loomis (Case Comment)](https://harvardlawreview.org/print/vol-130/state-v-loomis/) — _Harvard Law Review_, Vol. 130 - [Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management (GAO-26-107522)](https://www.gao.gov/products/gao-26-107522) — U.S. Government Accountability Office, March 2026 - [CFPB Issues Guidance on Credit Denials by Lenders Using Artificial Intelligence](https://www.consumerfinance.gov/archive/newsroom/cfpb-issues-guidance-on-credit-denials-by-lenders-using-artificial-intelligence/) — Consumer Financial Protection Bureau - [Moonshots — August 2026 Episode](https://www.youtube.com/watch?v=Jku8b2YKuy0) — Diamandis, Mostaque, Wissner-Gross, Ismail, Blundin