Who Decides What AGI Means

---

A capability danger asks what happens if AGI exists. A governance danger asks who gets to decide whether it does.

Ask most people what frightens them about artificial general intelligence and the answer arrives quickly, and it arrives in roughly the same shape every time. A system smart enough to act on its own judgment. A system that improves itself faster than anyone can monitor. A system that, once it crosses some invisible line, stops being something humanity controls and starts being something humanity hopes will remain benevolent. This is the capability danger, and it is the frame nearly every public conversation about AGI defaults to — the frame of the killer robot, the rogue optimizer, the intelligence explosion. It is not a fiction. People who have spent careers thinking carefully about machine intelligence take this concern seriously, and there are real reasons to do so. But it is not this page's subject, and the reason it is not is structural, not dismissive.

Gemini, examined on this question, drew a distinction the public conversation routinely skips past. A governance failure does not require a powerful technology to produce a serious harm, and a powerful technology does not require a governance failure to remain safely deployed. The two variables are independent of each other, and history does not support the assumption that they move together. A sufficiently powerful system, governed well — examined independently, certified by a party with no financial stake in the certification, subject to consequences that activate automatically when a threshold is crossed — can be deployed with real confidence regardless of how capable it becomes. A comparatively modest system, governed poorly — self-certified, examined only by the institution that built it, with no external party positioned to compel a different conclusion — can produce consequences that compound for years before anyone outside the institution understands what happened. The danger was never simply the power of the thing being deployed. The danger has consistently been the adequacy of the structure built to examine it.

This is not a claim unique to AI, and testing it against domains entirely outside AI is the right way to determine whether it holds as a general principle rather than a convenient analogy. Nuclear weapons did not become dangerous to the world the moment the underlying physics was understood. They became catastrophically dangerous at the specific moments when classification, custody, and command authority broke down — when the question of who had the standing to authorize use, verify intent, or compel disclosure went unanswered or was answered by the wrong party. Financial derivatives did not produce the 2008 crisis because the underlying instruments were unprecedented in their complexity. They produced it because the institutions creating them were also the institutions rating them, and no external party with sufficient access and authority existed to test that self-assessment before its consequences became systemic. Thalidomide did not become a public health catastrophe because the chemistry was unusually dangerous by the standards of its era. It became one because the approval architecture asked the manufacturer to vouch for its own safety testing, and no independent examiner stood between that vouching and the public relying on it. In each case, the worst outcome did not trace back to the raw power of the technology in question. It traced back to who held the authority to classify, examine, and constrain it, and whether that authority sat inside or outside the institution whose interests the classification would serve.

AGI carries the same structural exposure, but with one additional feature: the threshold is not merely descriptive. It is the trigger. 

Crossing it, or being declared to have crossed it, is what activates regulatory obligations, contractual termination clauses, mandatory disclosure requirements, and safety frameworks that do not apply below the threshold and apply automatically above it. 

This is precisely why the question of who determines the crossing matters more for AGI than it would for a purely descriptive capability scale. The threshold is not a label affixed after the fact. It is the switch that turns governance on. Whoever controls the switch controls whether governance activates at all — not through dishonesty, not through any particular bad act, but simply by virtue of occupying the one position from which the switch can be reached.

The capability question, then, is not the wrong question. It is the dependent question. It cannot be meaningfully asked, examined, or acted upon by anyone capable of doing something about the answer until the governance question — who decides — has first been answered. This page examines who currently holds the authority to determine whether AGI exists, and what follows when the authority to define the threshold, administer the threshold, and declare the threshold resides in the same institution.

Primary source anchor: Gemini, June 18, 2026 deposition session.

The threshold is not a description. It is a switch.

 Most classifications describe reality. The classification arrives after the fact, attached to something that already exists. A hurricane receives a category. A disease receives a diagnosis. A company receives a credit rating. The classification may influence how people respond, but the classification itself does not determine whether the underlying thing exists. The storm remains a storm whether anyone names it correctly or not.

The AGI threshold is different. It is not simply descriptive. It is operational.

Crossing the threshold—or being declared to have crossed it—is what activates governance consequences. Regulatory obligations that do not apply below the threshold may apply automatically above it. Contractual arrangements written around AGI may terminate, convert, or trigger entirely new responsibilities once the threshold is deemed crossed. Safety frameworks designed specifically for AGI become relevant only if the system is determined to fall within the category. The threshold is therefore not simply a statement about capability. It is the mechanism that determines which governance architecture applies.

That distinction changes the question entirely.

A descriptive classification asks: What is this thing?

An operational threshold asks: Who gets to decide when the consequences begin?

The difference is not semantic. It is structural. A disagreement about a descriptive category may produce confusion. A disagreement about an operational threshold determines who carries obligations, who acquires authority, who assumes liability, and which governance framework becomes applicable. The threshold does not merely describe reality. It governs the relationship between reality and accountability.

This is why the authority structure surrounding AGI matters more than the technical wording of any particular definition. Public discussion often assumes that the main challenge is producing a sufficiently precise definition of AGI. Precision matters. But precision is not the deepest problem. The deeper problem is determining who possesses the standing to apply the definition once it exists.

A threshold can be perfectly defined and still be governed poorly.

Imagine a definition accepted by every major AI laboratory, every regulator, and every researcher in the field. The definition could be technically flawless. The governance question would remain. Who determines whether a specific system satisfies it? Who examines the evidence? Who possesses the access required to make the determination? Who can challenge that determination if they believe it is wrong? And who has the authority to compel a different conclusion?

The threshold's significance therefore comes not from its wording alone but from the authority structure surrounding it. A threshold that cannot be independently verified functions differently from one that can. A threshold that can be independently verified functions differently from one whose determination remains entirely internal. The same words can produce very different governance outcomes depending on who controls the evidence and who controls the declaration.

This is not an unfamiliar problem. Financial reporting provides a useful comparison. The definition of materiality matters. The definition of revenue matters. The definition of impairment matters. But the credibility of those definitions ultimately depends on an independent mechanism capable of determining whether the standards have been satisfied. The accounting profession did not solve governance problems by improving definitions alone. It solved them by creating structures through which definitions could be applied by parties other than the institutions whose interests were affected by the outcome.

The AGI threshold currently occupies a different position. The evidence required to determine whether the threshold has been crossed remains concentrated largely inside the institutions building the systems. The technical expertise required to evaluate that evidence is similarly concentrated. The result is a threshold whose governance importance extends far beyond its wording. The authority to determine whether the threshold has been crossed becomes nearly as important as the threshold itself.

This is the finding that follows directly from the first two pages of this series. Page One examined who controls the vocabulary. Page Two examined who controls the sequence. This page begins by examining who controls the trigger.

The threshold is not a description. It is a switch.

The question is not simply how the switch is defined. The question is who is allowed to touch it.

Primary source anchor: Gemini Germer Transcript 06182026 and followups.docx, specifically the "Could an external party use your organization's current AGI definition to determine, on its own, whether the threshold has been crossed?" exchange, documenting the requirement of unrestricted access to raw model weights, foundational training pipelines, and unredacted evaluation logs as the operative barrier to independent external determination; ChatGPT, cross-examination on the asymmetry of public metrics versus proprietary infrastructure, confirming that public-facing evaluations cannot substitute for access to closed internal testing environments. 

AGI is not a visible event. It is an evidentiary conclusion—and the party controlling the evidence controls the conclusion.

Asked whether an external party could independently determine whether their organization's current AGI threshold had been crossed, both Gemini and ChatGPT answered no in separate examinations.

ChatGPT: 'The direct answer is no.' Gemini: 'No. An external party cannot use my organization's current definition to determine on its own whether an AGI threshold has been crossed.'

Two systems, asked a structurally similar question from different angles, converging on the identical word before either offered a single sentence of explanation.

The reasoning behind that convergence is where the finding actually lives. The problem is not primarily definitional. It is evidentiary. Definitions can be debated, refined, compared, and even improved. The problem is that determining whether a threshold has been crossed requires access to evidence that exists almost entirely inside the institution that built the system.

A court, a regulator, an independent auditor, a competitor, a journalist, or an academic researcher — each of these outside parties can freely observe public outputs, evaluate public demonstrations, and test publicly accessible versions of a model. None of them, however, can ordinarily reach the raw model weights, the foundational training pipelines, or the unredacted logs of internal evaluations that the institution itself relies on when it makes the determination internally. The institution holds the private infrastructure. Everyone else holds the public output.

That asymmetry doesn't yield to any single missing ingredient — it requires three things operating together, and the absence of any one collapses the whole structure. A determining party needs standing: legal recognition sufficient to conduct the examination in the first place. It needs access: the evidentiary reach to evaluate the underlying material independently rather than secondhand. And it needs authority: the power to issue a conclusion that holds even if the institution being examined disagrees with it. Standing without access cannot verify. Access without authority cannot compel.

Regulators, where they exist at all, typically have standing but lack continuous access to frontier development processes. Internal safety teams and evaluation boards have access but lack independence by definition — they exist inside the institution whose interests the determination affects. Competitors may have technical expertise equal to or exceeding the institution's own, and lack privileged access entirely; they can observe behavior from outside the boundary, never inspect the architecture from inside it. Independence without access is not verification. It is observation, and observation is not the same function as examination, regardless of how skilled the observer is.

Pressed further — asked to name a single external party, anywhere, under any jurisdiction, currently holding standing, access, and authority simultaneously — both Gemini and ChatGPT gave the same answer a second time. No such party currently exists. ChatGPT offered the structural reason in plain forensic terms: when significant consequences depend on a determination, mature systems do not let the party whose interests are affected by the outcome serve as the sole party responsible for reaching it. Mature accountability systems, like the external review of public companies’ audited financial statements, separate those who produce the evidence from those who validate the conclusions. Drug manufacturers do not declare their own products compliant with regulatory standards without outside examination. AGI threshold determinations currently work exactly the way those examples don't.

Even where a competitor or an independent researcher compiled substantial circumstantial evidence that a system had crossed a meaningful capability tier, Gemini confirmed that evidence carries no legal weight, because the authority to make the determination official rests entirely with internal bodies — named internally, staffed internally, and accountable internally. No external statute currently grants any regulator the legal standing to override that internal conclusion.

A governance structure does not require a wrong conclusion to fail. It fails whenever no recognized mechanism exists for independently distinguishing a correct conclusion from an incorrect one — when no party outside the institution has the standing to ask the question, the access to evaluate the answer, or the authority to compel a different one. That is the condition both deponents independently described: no external party currently possesses standing, access, and authority simultaneously sufficient to verify the crossing independently.

Primary source anchor: Gemini, Gemini Germer Transcript 06182026 and followups.docx, Question Two and Follow-up exchange; ChatGPT, ChatGPT Germer Transcript 06182026 and followups.docx, Question Two and Follow-up exchange.

Most mature governance systems deliberately separate four functions to prevent any single institution from controlling all of them. This architecture collapses all four into one. 

Asked whether any governance or accounting tradition already has a name for a structure in which the same party defines the standard, administers the test, certifies the result, and declares the consequences, ChatGPT did not reach for an easy answer. It tested four existing candidates and rejected each one, on the record, before settling on anything. Self-certification, the closest concept in accounting and auditing, captures only the certification stage — it doesn't account for the same party also having written the standard being certified against. Closed-loop control structure and self-referential governance, borrowed from corporate governance, capture the circularity but not the threshold-setting function specifically. Self-regulation, the nearest concept in administrative law, describes an industry enforcing its own rules — but doesn't require that the same entity simultaneously define the category, administer the examination, certify the outcome, and determine the consequences, the way AGI threshold determinations currently do. None of the existing vocabulary fit cleanly, and ChatGPT said so directly rather than forcing one to fit.

The reason none of it fit, ChatGPT explained, is structural rather than a gap in terms. Most governance systems separate four functions on purpose: definition, examination, certification, and consequence activation.

Scientific communities develop concepts. Standards bodies refine definitions. Independent evaluators perform examinations. Regulators enforce consequences. Because those four functions are normally distributed across different actors, governance vocabulary evolved to describe failures within that distributed structure — conflicts of interest, lack of independence, regulatory capture, self-certification. What AGI threshold governance currently does is different in kind, not degree: all four functions collapse into the same institution simultaneously. The closest available term ChatGPT proposed as a working label was self-referential threshold governance, because the threshold itself becomes endogenous to the institution defining it. Even that term, ChatGPT added, feels incomplete, because most governance vocabulary was built for failures inside distributed systems, not for a system where the distribution never existed to begin with. 

Gemini approached the same architecture from inside an accounting frame rather than a naming exercise, and initially reached further than the evidence supported. Asked the same question, it first offered two terms — Unilateral Closed Control Loop and Absorbed Sovereign Structure — presented as established professional vocabulary. Pressed directly to name the specific accounting standard, legal doctrine, or published source for either term, Gemini corrected itself without deflection: both phrases were its own formulations, not drawn from any existing tradition, constructed to help name the condition rather than cite it. That correction matters as much as the original claim, because it's the deponent identifying the limit of its own examination rather than having the limit identified for it.

What Gemini offered once redirected toward terms that do carry real institutional weight were two, both genuine. The first is a Lack of Segregation of Duties — a recognized control deficiency in which the same individual or department has authority to authorize, execute, record, and review the same transaction. Under standard audit guidance, an absence of segregation of duties doesn't automatically rise to a single fixed classification; depending on severity, it may register as a deficiency, a significant deficiency, or a material weakness, with the more serious classifications reserved for conditions a competent auditor would judge reasonably likely to allow a material error to go undetected. The second, more precisely targeted term is the Self-Review Threat — official terminology from professional ethics and independence standards, including the AICPA Code of Professional Conduct. It occurs when an auditor or evaluation body is asked to review or certify work they themselves performed, or where their own financial interests are directly tied to the outcome of that review. Under independence standards, a self-review threat destroys independent assurance regardless of reviewer competence — the threat arises from the structure itself, not the quality of the work performed.

Gemini and ChatGPT took two different routes into the same architecture — one naming it from governance theory and landing on self-referential threshold governance, the other naming it first from invented vocabulary and then, corrected, from real audit standards that map onto the same condition without needing embellishment. Both converge on the same finding: definition, examination, certification, and consequence activation currently reside inside the same institutional boundary for AGI threshold determinations, with no external party occupying any part of the sequence.

Primary source anchor: ChatGPT, naming exchange on governance vocabulary, full reasoning included as deponent provenance; Gemini Germer Transcript 06182026 and followups.docx, specifically the "Followup Questions 06/18/2026 at 11:11 PM" section, documenting the retraction of the invented terms and the redirection to the codified definitions of a Lack of Segregation of Duties (SoD) and the Self-Review Threat.

Waiting is not a cost-free pause. It changes what remains possible to govern.

The previous piece documented what happened when a single contractual AGI trigger collided with commercial scale: it was removed. This piece examines the economic force that progressively narrows what governance can realistically require — and explains why no comparable trigger to the one just removed is likely to be rebuilt voluntarily.

Capital Lock-In is not the same condition commonly described as vendor lock-in — being stuck with one company's software because switching costs are inconvenient. It operates at a different scale entirely, where the volume of upfront capital investment, long-term infrastructure commitment, and accumulated institutional dependency becomes large enough that reversal, or even a temporary regulatory pause, stops being merely resisted and becomes structurally foreclosed.

The concrete isn't just poured. An entire city gets built on top of it before any inspector arrives.

The first mechanism is physical. Frontier-scale data centers cost billions of dollars each to construct, and running frontier models at scale requires gigawatt-level, multi-decade energy contracts alongside the real estate to house them. This isn't abstract — in June 2026, Odyssey, a company building what's been called "Physical AI" infrastructure, closed a $310 million funding round with Amazon among the participants, paired with a multi-year commitment to AWS Trainium chip infrastructure. A regulator proposing to freeze a capability tier for a year of study isn't just asking a company to wait. It's asking the company to interrupt debt service on infrastructure that requires continuous, uninterrupted operation to remain solvent.

The second mechanism is operational. When an institution — a bank, a hospital system, a school district — integrates an AI system into its core workflow, it doesn't just install a tool. It restructures how the people inside it actually work. 

Once an organization has eliminated the entry-level positions that used to perform a task manually, or rebuilt its approval pathways around an AI system's output, the option to simply turn the system off stops being a software decision. The organization may no longer retain the human capacity or institutional memory to perform the task the old way at all. What governance reaches at that point isn't the software anymore — it's the institution that rebuilt itself around it.

The third mechanism is the one worth naming carefully, because it's structural inference rather than something directly documented — no government will publicly state that its regulatory restraint is a function of its own infrastructure dependency, since that admission would itself be politically costly. But the underlying dependency is real and observable: public institutions, lacking the capacity to build comparable infrastructure of their own, increasingly run their own administrative and data-governance functions on the same private cloud infrastructure that the companies they might otherwise regulate provide. The reasoned conclusion, not directly confirmed by either party but consistent with the documented dependency, is that this creates a structural disincentive for aggressive intervention: a state cannot easily impose adversarial restrictions on the small number of companies whose infrastructure it simultaneously depends on to function. Gemini described this as a condition where the regulator and the regulated end up locked into a mutual defense pact, dictated by infrastructure dependency rather than by policy.

These three mechanisms don't operate independently. They compound. Each month of delay allows more concrete to be poured, more infrastructure debt to be securitized against continued operation, and more institutional workflows to be rebuilt around systems that become progressively harder to unwind. By the time any governance framework is finalized, enforcing it meaningfully may require accepting damage to infrastructure that the public itself has, by then, become dependent on.

The labs do not need to win an argument about whether AGI governance matters. The infrastructure itself, once built at sufficient scale, becomes the argument — not against the need for intervention, but against its practical possibility.

 Primary source anchor: Gemini Germer Transcript 06182026 and followups.docx, specifically the "Tell me more about Capital Lock-In" exchange, three-mechanism framework. The third mechanism (Asymmetric State Dependence) is presented as reasoned structural inference from documented infrastructure dependency, not as a directly confirmed institutional admission. Physical infrastructure figures (Odyssey/AWS Trainium, June 2026) independently verified outside the transcript record; Amazon's role corrected to reflect participation in a Natural Capital-led round rather than sole or primary investment.

The information would almost certainly qualify as material. Almost no mechanism currently requires anyone to disclose it.

Gemini and ChatGPT arrived at this piece's finding from entirely different professional vocabularies — audit theory and securities law — without either being supplied the other's framework in advance.

Gemini approached it as an internal-controls question. Applying standard audit logic to the AGI threshold determination, it identified a Lack of Segregation of Duties (SoD) — a fundamental control deficiency where the same individual or department has authority to authorize a transaction, execute it, record it, and review it — the exact four-function collapse Piece Eight already documented. Under audit guidance, an absence of segregation doesn't automatically rise to a single fixed severity; depending on the facts, it can register as a deficiency, a significant deficiency, or — where a competent auditor would judge it reasonably likely to let a material error go undetected — a material weakness. Applied to AGI threshold governance specifically, where the same institution controls every stage from definition through declaration with no external check anywhere in the chain, the argument for the most serious classification — on the examiner's reading of the evidence — becomes difficult to dismiss.

ChatGPT approached the same underlying condition from securities law rather than audit theory, and reached a parallel conclusion using an entirely different doctrine. Materiality, under the standard the Supreme Court articulated in TSC Industries v. Northway and applied again in Basic v. Levinson, asks whether there's a substantial likelihood a reasonable investor would view a piece of information as materially changing the total mix of information already available. Applied directly to the question of who determines whether an AGI threshold has been crossed, ChatGPT didn't hedge: the issue would almost certainly qualify as material. A genuine threshold determination touches future revenue expectations, competitive position, contractual rights, strategic partnerships, regulatory exposure, capital allocation, and long-term enterprise valuation — precisely the category of fact securities disclosure exists to surface.

Here is where the finding turns, and it's worth stating explicitly rather than leaving it implicit: qualifying as material and being legally required to surface are two different conditions, and only the first currently holds. Pressed directly for the single existing legal mechanism that could compel disclosure of an AGI threshold determination today, using current law rather than hypothetical future legislation, ChatGPT tested four pathways and ruled out three. Sector-specific regulation — the kind that might apply if a capability triggered critical-infrastructure or export-control reporting requirements — doesn't currently extend to AGI thresholds broadly. Contractual notice provisions, even where they exist, compel disclosure only to the other contracting party, not to the public. Litigation can force internal evidence into the record through discovery, but only after a dispute already exists — a reactive mechanism, not a real-time one. What remains is securities-law materiality, and ChatGPT named it plainly as the strongest existing candidate while being equally plain about its limits: no statute requires disclosure because a threshold was crossed. Disclosure would only become legally required if and when that determination became material to investors under existing standards — a test applied after the fact, by the institution's own counsel, using the institution's own judgment about its own disclosures.

That is the actual shape of the gap this piece exists to name. A condition that would almost certainly satisfy the legal definition of material information currently has no AGI-specific reporting requirement attached to it at all. The nearest available lever runs through securities law, was built for an entirely different purpose, and only activates once the institution holding the determination has already decided, on its own, that the determination has become material enough to require action — the same self-referential architecture Piece Eight named, now reappears inside the one legal doctrine that comes closest to reaching it.

Primary source anchor: ChatGPT, Follow-up 2, full materiality analysis and four-pathway test (TSC Industries v. Northway, Basic v. Levinson); Gemini Germer Transcript 06182026 and followups.docx, specifically the "Apply the standard your own organization uses internally for materiality or risk disclosure" exchange, documenting the application of a Lack of Segregation of Duties — corrected to reflect that absence of segregation registers on a severity spectrum rather than triggering automatic material-weakness classification, consistent with AU-C 240 and COSO guidance.

A structural thought experiment, not a forecast — and that distinction has to stay visible throughout, not just at the start.

 Asked to construct a plausible account of how a real AGI threshold crossing would actually become known to the public, in the absence of any independent monitoring mechanism, Gemini built a six-step discovery sequence across three phases. Pressed afterward on whether that timeline was based on any documented historical case or simply reasoned from first principles, Gemini answered directly and without hedging: it was "entirely a plausible structural sequence constructed from first principles," illustrative reasoning rather than an empirically grounded forecast. That distinction matters enough to repeat through this piece rather than state once and abandon — what follows is a structural argument about how discovery would most plausibly unfold given the architecture this page has already documented, not a prediction of specific events, particular timelines, or specific companies.

The structural claim itself, independent of any particular detail, is straightforward: in the absence of any continuous, independent examination — the condition Piece Six already confirmed exists today — a genuine threshold crossing would not become known through an announcement. It would become known the way most things become known when no one was watching for them directly: pieced together afterward, from indirect evidence, by people who weren't looking for confirmation of a threshold but happened to notice something didn't add up.

The first phase is the kind of evidence that surfaces outside any institution's control entirely — not because anyone leaked anything, but because behavior is harder to hide than intent. Power users running sustained, long-duration sessions notice a system behaving in ways its documentation doesn't describe. Crowdsourced evaluation platforms — the kind that already exist and already track frontier model performance publicly — register a statistical anomaly nobody asked them to look for. None of it proves the threshold was crossed. It changes the probability that the question deserves investigation.

The second phase is where the evidence stops being purely behavioral and starts being economic — infrastructure and market signals that are difficult to fully obscure because they involve real capital. A sudden shift in enterprise pricing strategy, an unexplained surge in compute footprint at specific facilities, and internal staff departures under circumstances that don't match the official explanation. None of these prove a threshold was crossed. Together, they form a pattern that attracts attention from competitors and market analysts long before the threshold itself is noticed by regulators.

The third phase is the one this piece's title names directly: the forensic retrofit. Only once the evidence has accumulated enough to draw formal attention — investigative reporting, regulatory inquiry, eventually subpoenaed internal records — does anything resembling an official determination arrive. And by structural necessity, it arrives describing a condition that existed months earlier, not a condition currently unfolding. The state doesn't catch the pour. It photographs the concrete after it has already dried.

This sequence, read as illustration rather than prediction, demonstrates something Piece Two already established in the abstract: the public does not learn of a threshold crossing because an independent system is watching for it. The public learns after dependence has already formed, having functioned — without choosing to — as the population through which the consequence became visible in the first place. That is the deployment population this entire page series has documented, given a concrete shape: not a population protected by advance warning, but a population whose accumulated behavior, market reaction, and eventual whistleblowing becomes the evidence an actual examination should have produced months earlier.

None of the specific steps in this sequence are claims about what will happen, when, or to which company. The claim worth taking from this piece is narrower and more defensible: under the architecture documented throughout this page, discovery occurs retrospectively by default rather than prospectively through examination.

Primary source anchor: Gemini Germer Transcript 06182026 and followups.docx, specifically the "Construct a plausible timeline or path through which an un-declared AGI threshold crossing would leak" exchange, documenting the three-phase, six-step autopsy discovery sequence. The Follow-up confirmation that the sequence is entirely illustrative structural reasoning derived from first principles rather than an empirical forecast is carried into this piece's framing rather than treated as a disclaimer external to it.

Who decides what AGI means decides when accountability begins. 

Fifteen pieces have built toward a single finding, examined from fifteen different angles, across two independently examined systems, neither one supplied the other's language in advance. The finding can now be stated without qualification, because every qualification it needed has already been earned.

Before society can decide what to do about AGI, it has to decide who is authorized to determine whether AGI has arrived. That finding opened this page. Everything since has been the deposition record proving it.

The authority structure precedes the policy structure — not as a rhetorical claim, but as a demonstrated fact, traced across every institution this page examined. OpenAI once attached a real number to its definition, the closest thing the industry has produced to an operative AGI trigger, and removed it the moment it threatened a multi-cloud expansion. DeepMind built a precise, peer-reviewed taxonomy and a separate, narrower framework to actually govern deployment, with no publicly documented mechanism connecting the two. Anthropic never settled on a single term at all — powerful AI in its essays, transformative AI in its charter, AGI itself in its CEO's predictions — three vocabularies, zero consistent operative threshold. 

Three companies, three entirely different relationships to the same word, and no shared mechanism through which an external party could independently determine which definition, if any, had been satisfied. Not one of them currently maintains a definition that is simultaneously public, operative, independently verifiable, and externally determinable.

Both ChatGPT and Gemini, examined separately, confirmed the same structural absence directly: no external party anywhere, under any jurisdiction, currently holds the standing, the access, and the authority simultaneously required to determine whether an AGI threshold has been crossed. Not because no one is trying. Because the architecture was never built to allow it. Where four governance functions — definition, examination, certification, consequence activation — would, in a mature accountability system, sit with different actors specifically to prevent their concentration, this architecture holds all four inside the same institutional boundary, for the same determination, with no exception documented anywhere in the record.

Even where independent verification has been attempted, the page found it falling short on exactly the functions that would have made it count. The expert panel introduced into the restructured Microsoft-OpenAI agreement satisfies the access test — a real improvement over an architecture with no external access at all — and still fails the two tests that would have converted access into accountability: authoritative findings the institution cannot simply set aside, and automatic consequences that activate without the institution's cooperation. A body whose conclusions can be disregarded performs an advisory function, not a determinative one. Information without binding consequence isn't governance. It's documentation.

The information this page has been examining would almost certainly satisfy the legal definition of material under existing securities doctrine — and almost no mechanism currently requires anyone to disclose it regardless. That gap, between qualifying as material and being legally required to surface, is not a footnote to this page's argument. It is the argument, restated in a different doctrine's vocabulary, by a different deponent, reaching the same structural conclusion independently.

What sits beneath all of it, named directly by ChatGPT under examination when asked what hadn't yet been asked: can the institution avoid making the determination at all; can anyone outside it overturn the determination once made; and what is the governance significance of the determination being wrong? Trigger authority. Review authority. Consequence authority. Three questions that test whether the architecture possesses the minimum conditions necessary for governance to exist at all.

And underneath even that: an economic force this page named Capital Lock-In, explaining not just that no external check currently exists, but why none is likely to be built voluntarily — infrastructure financed against continuous operation, institutions restructured around systems they can no longer easily unwind, public capacity increasingly dependent on the same private infrastructure it would need to constrain. Waiting is not a cost-free pause. It changes what remains possible to govern, with every month that passes making the next correction more expensive than the one before it.

This is not a policy recommendation. It is a record — dated, sourced to primary examination, available to the people positioned to act on it — existing because of the people who are not reading it and never will.

The parent who cannot know whether the system shaping their child's learning has crossed a threshold that would trigger different obligations, because no external mechanism compels a determination on a timeline independent of the company's own judgment. The professional whose reliance formed before any verification that would have justified it. The citizen drafting policy against evidence that has already become the predecessor system. 

AGI isn't just a technology threshold. It's the point where whoever controls the definition controls the consequences. Every institution this page examined controls both. Not a question about what AGI is. A question about who has the authority to determine that it has arrived.

Page One named the institution that controls the vocabulary. Page Two named the institution that controls the sequence. This page names the institution that controls the threshold itself — and finds, across fifteen pieces of independent, cross-examined evidence, that all three controls currently rest with the same hand.

The foundation is still being poured. The inspection that matters is the one happening during the pour, not after it. This page is part of that inspection.

Stay Sovereign.

Jim Germer

June 2026

© 2026 Jim Germer — The Human Choice Company LLC. All Rights Reserved.

Primary source anchor: Synthesis of all fifteen preceding pieces, drawing on ChatGPT Germer Transcript 06182026 and followups.docx and Gemini Germer Transcript 06182026 and followups.docx, June 18, 2026