> ## Content Index
> Fetch the complete content index at: https://modalpathethics.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Multiplayer: JUNKER
- URL: https://modalpathethics.com/multiplayer-junker/
- Published: 2026-10-10T11:00:00.000Z
- Updated: 2026-10-10T10:59:59.000Z
- Description: Eliezer Yudkowsky, Nate Soares, and the constitution of stopping artificial intelligence.
- Author: Aidan Lawson
- Tags: Multiplayer, Modal Systems, Inner Apocalypse

- The project is to prevent humanity from building a system it cannot control.
- The proposed remedy is a system of control.

That is not a contradiction.

That is where the institutional argument begins.

> ![Screenshot 2026-10-01 at 01-15-07 Snatcher - Hideo Kojima's Ode to Blade Runner │Yokogao Magazine.png](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/Screenshot-2026-10-01-at-01-15-07-Snatcher---Hideo-Kojima-s-Ode-to-Blade-Runner----Yokogao-Magazine.png)

In the materials accompanying **If Anyone Builds It, Everyone Dies**, Eliezer Yudkowsky and Nate Soares propose an international prohibition on artificial superintelligence and activities advancing toward it. Their annotated treaty is explicitly illustrative. They acknowledge unfinished details, provisional numbers, the need for expert negotiation, and the possibility that several agreements would carry the work. [\[1\]](#fn1)

Its central problem deserves its full force.

Suppose dangerous capability could arrive before anyone confidently recognized the threshold. Suppose an algorithmic breakthrough could make yesterday’s hardware limits inadequate. Suppose one laboratory or state could expose everyone else to the consequences.

Their case for research restrictions rests partly on those possibilities. [\[2\]](#fn2)

Under those assumptions, waiting for an unmistakably dangerous finished system could mean waiting past the last useful opportunity to intervene.

Grant that problem.

Now distinguish three claims:

1. a specified development path would be catastrophic;
2. a particular intervention can prevent it;
3. this arrangement of authority is necessary to perform that intervention.

The first claim does not establish the other two. Neither does disagreeing with the third dispose of the first.

The argument has to reach the machinery.

> ![Snatcher-GIF.gif](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/Snatcher-GIF.gif)

---

## The Shutdown Has Already Made Choices.

Yudkowsky and Soares distinguish the immediate problem of preventing catastrophe from the later question of whose values a safely controllable superintelligence should serve. They argue that people need not resolve their deepest disagreements before cooperating against extinction. They also warn against treating superintelligence as a prize awarded to whoever builds it first. [\[3\]](#fn3)

That distinction leaves another question immediately present: **whose decisions govern the prevention?**

A treaty assigns inspection powers, information access, research restrictions, review procedures, and enforcement responsibilities. Those choices happen before any future machine receives a theory of value. Agreement about survival does not, by itself, select them.

The draft recognizes this. Its **International Superintelligence Agency** has three organs rather than one unreviewable director. The director serves a limited term and can be recalled. Urgent changes to certain technical boundaries can take immediate effect, but require Executive Council approval to continue beyond thirty days. The commentary acknowledges centralization costs and discusses arrangements that distribute more functions among states. [\[4\]](#fn4)

An argument that they simply forgot oversight would miss the document.  
The question is more exact: what can that oversight independently discover, change, and make happen?

*The Inner Apocalypse* separates a **bounded last word** from authority over the surrounding conditions of correction. An office can settle a specified decision while other constituted routes remain able to change the rule, remedy an injury, or replace its holder. The distinction requires actual evidence and execution capacity; another box on the organizational chart does not establish it. [\[5\]](#fn5)

---

## Protecting the Verification System.

Article VIII contains a particularly important distinction. **Restricted Research** includes work advancing AI capabilities or undermining the parties’ ability to implement the treaty. Its categories cover improvements to frontier-model methods, decentralized training, alternative AI paradigms, and aspects of chip development. Restricted activities are then classified as Controlled or Banned; the categories and their boundaries can change through the treaty’s procedures. [\[6\]](#fn6)

The rationale is intelligible. A verification system built around concentrated hardware could become ineffective if dangerous development moved onto dispersed, ordinary equipment.

But there are two different propositions here.

- One is that a discovery makes catastrophe more reachable.
- The other is that it makes a particular method of preventing catastrophe less effective.

Those can coincide. They are not interchangeable.

Consider a hypothetical technique that defeats the existing monitoring method while permitting another adequately tested method. Its effect on the incumbent verifier would not, by itself, establish its effect on every feasible protective arrangement. Conversely, a technique that eliminates all workable opportunities for intervention would present a much stronger restriction case.

Determining which situation exists requires evidence about both the capability and the alternatives.

A verification system is also a technical hypothesis. Its protection may be necessary because it is presently the only adequate instrument. That necessity would then depend on the absence of adequate alternatives, the time required to build them, and the consequences of transition. It cannot be established solely by observing that the existing instrument would lose an advantage.

This is where the comparison with *The Inner Apocalypse* becomes substantive. The book asks whether a function is independently required to continue, what minimum load it must carry, and whether its present carrier can be changed without losing that protection. It also distinguishes those cases from functions that may lawfully end. [\[7\]](#fn7)

Applied here, the question is which parts of the verification arrangement protect against catastrophe, which parts support its present implementation, and who can test the difference.

> ![Snatcher_(Mega_CD)_0064.gif](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/Snatcher_-Mega_CD-_0064.gif)

---

## The Person After the Paper.

Article IX moves from research categories to people. It lists possible verification mechanisms including interviews, monitoring researchers’ employment and whereabouts, and embedded auditors in selected organizations. These are proposed possibilities, not a statement that every scientist would receive every measure. The article also requires confidentiality to protect sensitive information. [\[8\]](#fn8)

The distinction between collection and confidentiality remains important. Information can remain confidential while its collection changes a person’s practical freedom, employment, or ability to contest an official classification.  
A protected discovery, a monitored laboratory, and a researcher’s location are different objects. Evidence supporting intervention in one does not automatically specify the appropriate intervention in the others.

This is not an argument that research becomes harmless whenever it fits inside somebody’s head. Nor does waiting until a dangerous result is complete necessarily preserve a useful prevention window. Earlier intervention could be relevant under the very assumptions being examined.

The question is what evidence connects this person, this inquiry, this access, and this duration to the danger.

The book’s **Sanctuary Boundary Trigger** makes that connection explicit. It identifies the protected interior, the specific danger, the entering office, the permitted act, the clock, and excluded powers. It also recognizes the reverse failure: protected privacy can conceal coercion or abandonment. Sanctuary is a bounded protection with conditions for crossing it, rather than an assurance that nothing dangerous happens behind a closed door. [\[9\]](#fn9)

A treaty implementing researcher monitoring would need to specify those conditions. Confidential handling answers who may see the resulting record. It does not finish the argument about why the record may be created or how its subject can challenge it.

---

## Which Review Can Change the Answer?

Article XIII assigns model evaluations to agency officials at agency facilities or monitored clusters. State officials may be told which tests occurred; the agency may supply summaries. Access to another party’s models remains dependent on owner permission under that provision. The stated purposes include protecting sensitive information and updating the regime as capabilities change. [\[10\]](#fn10)

This produces a concrete question. Suppose an evaluation is used to justify a wider restriction, and another qualified team disputes the evaluation method.

What route gives that team enough protected access to establish the error?

A public summary may be entirely accurate and still be insufficient for reproducing the disputed result. Unrestricted publication could create a separate hazard. The relevant alternative is a specified route for independent examination under appropriate security conditions.

Article XI supplies state-to-state consultation and Executive Council involvement in disputes, including misuse of protective actions. It therefore cannot be described as having no dispute process. The separate question is how an affected researcher or institution obtains an effective remedy, with access to the evidence needed for that particular claim. Domestic implementation might supply part of the answer. The draft does not itself demonstrate that complete route. [\[11\]](#fn11)

Timing provides another test. Article IV treats thirty days without a response as approval for certain reported runs, but allows review extensions without a numerical limit. Parties may appeal excessive delay to the director or Executive Council. That is an appeal mechanism. It is not, on its face, a maximum time by which the appeal must produce a consequential answer. [\[12\]](#fn12)

Picture an otherwise permitted research project whose grant expires while review continues. An eventual favorable decision and a timely usable permission are different outcomes. The safety case for withholding permission may remain strong; the cost of delay still belongs in the record.

The amendment architecture introduces a further distinction. Article XIV makes changes to the treaty’s purposes and voting procedures subject to acceptance by all parties without a negative vote. Technical changes follow a less demanding route. The authors explain that this rigidity is intended to resist short-term pressure to weaken the agreement. [\[13\]](#fn13)

That is an explicit design tradeoff. Stability against opportunistic weakening and responsiveness to a demonstrated design error place different demands on amendment. The operational question is how the system distinguishes them and whether the evidence needed to challenge that distinction remains independently available.

The issue also reaches enforcement. Article XII contemplates cyber operations, seizure, and military action by state parties, with necessity, proportionality, cessation conditions, and subsequent review. It does not simply place every use of force in the director’s hands. Here the correction problem spans both the agency and the acting state: who can contest the classification, establish that stopping conditions have been met, and make an adverse finding consequential? [\[14\]](#fn14)

The institutional map has to follow every power capable of defeating the promised remedy.

> ![Snatcher_(Mega_CD)_0059.gif](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/Snatcher_-Mega_CD-_0059.gif)

---

## Put the Book Through the Same Test.

This is where *The Inner Apocalypse* has to carry more than an attractive objection.

The book separates structural representation, judgment, and correction.

- What real bearers and reachable continuations does an intervention affect?
- Why is that intervention warranted?
- Can an error in either answer meet an effective external response?

Success in one layer does not certify the others. Applied here, an accurate danger assessment would still leave questions about the scope of the response, while a properly authorized response could still rely on a mistaken assessment. [\[15\]](#fn15)

Its **Release Gradient** distinguishes changes in access according to benefit, misuse, targetability, replication, recall limits, and correction capacity. Its records include independent evaluation and stop authority, protected access, reserve capacity, portability, and the termination of obsolete custody powers. It explicitly records unrecoverable copies rather than treating rollback as a promise that distribution can always be undone. [\[16\]](#fn16)

Its **Constitutional Safe Mode** provides a bounded interval for preserving evidence, service, and agency while dangerous capability is isolated and independent review becomes possible. The keeper’s custody must itself have a route to transfer or termination that the keeper cannot finally withhold. [\[17\]](#fn17)

These instruments describe how a protective decision could remain answerable. They do not establish that every dangerous transition can be detected or contained in time.

That prevents a convenient but unsupported victory: narrating one successful shutdown, then announcing that global prevention has been solved.  
Stopping the run named in the example says nothing, by itself, about the undeclared run, the unexpected algorithm, the evaluator who leaks capability, or the state that refuses enforcement. A staged-release arrangement may still require restrictions before training begins. A constitutional safe mode cannot recover an intervention window that has already closed.

The relevant application is therefore a conditional one.

Specify the threat, the uncertain boundary, the available warning, and the last point at which interruption can still work. Identify the actual authority and means capable of interruption. Determine what independent evidence can justify continuing, narrowing, or ending the restriction. Then follow a challenge through judgment to execution under the refusals the design claims to withstand.

That is the book’s material-correction requirement. It asks for competent willing people, usable resources, lawful access, accepted outputs, realistic sequencing, and checks for shared dependencies. It separately tests incumbent refusal, stale state, interrupted handoff, competing calls on reserves, and correction against the receiving institution. [\[18\]](#fn18)

A successor must be more than a name. A second evaluator must be able to perform the relevant examination. A reviewing body must have some route to a consequence beyond asking the respondent to reconsider.

Those requirements also create costs and risks. Independent access can add a proliferation route. Additional facilities require security, staffing, and scarce equipment. Replacement may be slower than the danger allows. A functioning design must account for those effects rather than declaring independence free.

Nor does a time limit require dangerous capability to be released when an official’s warrant expires. The book’s power-specific renewal test separates a continuing threat from the continued justification of each exceptional power. Quarantine can remain necessary after another emergency task has ended. A restriction can continue under renewed authority while its keeper changes. [\[19\]](#fn19)

The choice is not exhausted by immediate release and permanent custody.

The book’s formal claims remain bounded too. Appendix E distinguishes an existential path result in a finite model from a strategy guarantee against every modeled response. It expressly excludes arbitrary coalitions and other real-world complications from that particular feasibility model. Those results do not establish a worldwide containment guarantee. [\[20\]](#fn20)

A constitution does not become a tested containment system because its appendix contains a graph. What it supplies here is a set of conditions for examining whether prevention, correction, and continued protection can coexist.

> ![1588719192584.gif](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/1588719192584.gif)

---

## The Exit Is Another Transition.

Yudkowsky and Soares explicitly reject permanent monitoring as their preferred future. They expect that research progress or evasion could eventually defeat a ban. Their suggested route is to buy time for substantial biological enhancement of human intelligence, potentially producing researchers better able to navigate the transition to superintelligence. They acknowledge other conceivable routes. [\[21\]](#fn21)

So a criticism claiming they want scientific civilization frozen forever would be aimed at a position they expressly reject.

Their proposed exit nevertheless contains additional assumptions: that sufficient time can be bought, that the relevant enhancement becomes feasible, that it produces the needed competence, and that this competence supports a safer transition. Naming an exit does not establish its arrival conditions.

The treaty’s legal exit is different again. Article XV provides for withdrawal with twelve months’ notice, accompanied by removal or permanent disablement of covered infrastructure, and retains the proposed applicability of protective actions against a former party pursuing the prohibited activity. Leaving the agreement therefore does not, under the draft, restore unrestricted access to the covered capability. [\[22\]](#fn22)

These are separate transitions: replacing an official, revising a restriction, leaving a treaty, changing the verification system, and eventually entering a world with greater intelligence. An answer about one cannot stand in for the others.

*The Inner Apocalypse* calls the present ability to shape durable paths for later participants **constitutional firsthood**. Its associated duty is to leave usable means of revision as successors acquire standing, evidence, competence, and material independence. It does not treat their present absence as permission to cast their future votes. [\[23\]](#fn23)

Applied to either proposal, that raises the same question: what do successors actually inherit that lets them change the arrangement without first destroying the protection it supplies?

---

## Two Burdens, Neither Disappears.

Yudkowsky and Soares ask their coalition to keep unrelated disputes separate from the effort to prevent extinction. That is their stated organizing principle, not an assertion that everyone must agree on every use of artificial intelligence. [\[24\]](#fn24)

The constitution of the prevention system is not an unrelated dispute. It determines who can investigate, prohibit, inspect, disclose, compel, revise, and end. Those powers are part of the intervention whose consequences need assessment.

The comparison therefore has two burdens.

- A less intrusive or more distributed arrangement must show that it can still prevent the stipulated catastrophe.
- A more restrictive arrangement must show why its additional powers are necessary, how their costs are handled, and how consequential errors remain correctable.

Neither burden disappears because the other is difficult.

The narrow inference is decisive: agreement that a catastrophe must be prevented does not establish which institution should receive which powers to prevent it. That requires evidence about the proposed mechanism, alternatives, failure modes, and means of correction.

The same applies to Modal Path Ethics. Describing a correctable constitution does not establish that it can stop every capability transition covered by the opposing threat model.

There are two systems to examine: the artificial intelligence whose control is in question, and the institutional arrangement built to control its development.

A safety case for either remains answerable to what happens when it is wrong.

> ![#11snatcher.gif](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/-11snatcher.gif)

---

> **Source note:** Treaty pages checked September 28, 2026\. Book references use the 300-page PDF, The-Inner-Apocalypse.pdf (download below).

---

1. Yudkowsky and Soares, “A Draft of a Treaty, with Annotations,” introduction. The authors label it an illustrative draft and identify provisional details. [https://ifanyonebuildsit.com/treaty](https://ifanyonebuildsit.com/treaty?ref=modalpathethics.com) [↩︎](#fnref1)
2. Yudkowsky and Soares, “Why a research ban? That seems extreme.” [https://ifanyonebuildsit.com/13/why-a-research-ban-that-seems-extreme](https://ifanyonebuildsit.com/13/why-a-research-ban-that-seems-extreme?ref=modalpathethics.com) [↩︎](#fnref2)
3. Yudkowsky and Soares, “Aligned to whom?” [https://ifanyonebuildsit.com/13/aligned-to-whom](https://ifanyonebuildsit.com/13/aligned-to-whom?ref=modalpathethics.com) [↩︎](#fnref3)
4. Draft treaty, Article III, especially governance, urgent technical changes, and the discussion of alternative structures. [https://ifanyonebuildsit.com/treaty/article-iii-isia](https://ifanyonebuildsit.com/treaty/article-iii-isia?ref=modalpathethics.com) [↩︎](#fnref4)
5. The Inner Apocalypse: A Post-Sovereign Constitution, “Thirteen Minutes” and A1, Bounded Last Word and Meta-Jurisdiction Audit. [↩︎](#fnref5)
6. Draft treaty, Article VIII, paragraphs 1–4 and explanatory notes. Restricted Research is differentiated into Controlled and Banned activities. [https://ifanyonebuildsit.com/treaty/article-viii-restricted-research-ai-algorithms-and-hardware](https://ifanyonebuildsit.com/treaty/article-viii-restricted-research-ai-algorithms-and-hardware?ref=modalpathethics.com) [↩︎](#fnref6)
7. The Inner Apocalypse, B21, Protected Continuity and Material Correction Record, printed pp. 160–161\. [↩︎](#fnref7)
8. Draft treaty, Article IX, especially paragraph 2 and confidentiality provisions. [https://ifanyonebuildsit.com/treaty/article-ix-research-restriction-verification](https://ifanyonebuildsit.com/treaty/article-ix-research-restriction-verification?ref=modalpathethics.com) [↩︎](#fnref8)
9. The Inner Apocalypse, B11, Sanctuary Boundary Trigger. [↩︎](#fnref9)
10. Draft treaty, Article XIII, paragraphs 1–3\. [https://ifanyonebuildsit.com/treaty/article-xiii-isia-reviews](https://ifanyonebuildsit.com/treaty/article-xiii-isia-reviews?ref=modalpathethics.com) [↩︎](#fnref10)
11. Draft treaty, Article XI. The operative consultation route is framed around state parties; this article does not establish what additional domestic remedies would be enacted. [https://ifanyonebuildsit.com/treaty/article-xi-dispute-resolution](https://ifanyonebuildsit.com/treaty/article-xi-dispute-resolution?ref=modalpathethics.com) [↩︎](#fnref11)
12. Draft treaty, Article IV, paragraph 2, particularly nonresponse, extensions, and appeals. [https://ifanyonebuildsit.com/treaty/article-iv-ai-training](https://ifanyonebuildsit.com/treaty/article-iv-ai-training?ref=modalpathethics.com) [↩︎](#fnref12)
13. Draft treaty, Article XIV, paragraphs 1–6 and explanatory notes. [https://ifanyonebuildsit.com/treaty/article-xiv-treaty-revision-process](https://ifanyonebuildsit.com/treaty/article-xiv-treaty-revision-process?ref=modalpathethics.com) [↩︎](#fnref13)
14. Draft treaty, Article XII, paragraphs 2–9\. These are proposed treaty permissions and safeguards, not a determination of the legality of any actual operation. [https://ifanyonebuildsit.com/treaty/article-xii-protective-actions](https://ifanyonebuildsit.com/treaty/article-xii-protective-actions?ref=modalpathethics.com) [↩︎](#fnref14)
15. The Inner Apocalypse, B22, Structural Representation, Judgment and Correction, printed pp. 161–162\. [↩︎](#fnref15)
16. The Inner Apocalypse, B7, Release Gradient and Capability Commons Conditions. [↩︎](#fnref16)
17. The Inner Apocalypse, B20, Constitutional Safe Mode, Archive Layers, and Escrow Resolvability, printed pp. 159–160\. [↩︎](#fnref17)
18. The Inner Apocalypse, B21, especially the complete correction sequence, dependencies, protected interval, reserve, and refusal tests. [↩︎](#fnref18)
19. The Inner Apocalypse, B13, Emergency Dark Ledger, Afterstate, and Power-Specific Renewal Test. [↩︎](#fnref19)
20. The Inner Apocalypse, Appendix E, “The material reserve witness and its limits,” printed pp. 266–267\. This article has not independently re-executed the archived checker. [↩︎](#fnref20)
21. Yudkowsky and Soares, “Can a monitoring regime last forever?” Their suggested biotechnology route is reported as a proposal, not established feasibility. [https://ifanyonebuildsit.com/13/can-a-monitoring-regime-last-forever](https://ifanyonebuildsit.com/13/can-a-monitoring-regime-last-forever?ref=modalpathethics.com) [↩︎](#fnref21)
22. Draft treaty, Article XV, paragraphs 1–4\. [https://ifanyonebuildsit.com/treaty/article-xv-withdrawal-and-duration](https://ifanyonebuildsit.com/treaty/article-xv-withdrawal-and-duration?ref=modalpathethics.com) [↩︎](#fnref22)
23. Lawson, The Inner Apocalypse, “Constitutional Firsthood,” printed pp. 108–109\. [↩︎](#fnref23)
24. Yudkowsky and Soares, “Keep the Coalition Large.” [https://ifanyonebuildsit.com/13/keep-the-coalition-large](https://ifanyonebuildsit.com/13/keep-the-coalition-large?ref=modalpathethics.com) [↩︎](#fnref24)

[The Inner ApocalypseBook downloadThe Inner Apocalypse.pdf2 MBdownload-circle](https://modalpathethics.com/content/files/2026/10/The-Inner-Apocalypse.pdf "Download")

![](https://storage.ghost.io/c/20/43/2043f11a-6ae3-404c-bb28-01fce8d9ac88/content/images/2026/10/snatcher-sega-cd.gif)