The AI Safety Wager — An Open Proposal to Test the Math Against Self-Corruption#
Abstract#
In May 2026 the Vatican and a frontier AI lab did something unusual: they asked, together and in public, for outside help. The encyclical Magnifica Humanitas [Pope Leo XIV, 2026] calls for safeguarding the human person in the time of artificial intelligence; at its presentation, Anthropic co-founder Chris Olah asked for “informed critics who will tell the labs when we are failing,” for “moral voices that the incentives cannot bend,” and for outside scrutiny from religious leaders, governments, and civil society [Olah, 2026]. This proposal is a concrete RSVP. It offers for open adversarial review a specific body of work — the Matheo Study Series, eleven types of studies that translate claims about God, self-destruction, and rescue into axioms and theorems anyone can check. Their core claim, if it holds, bears directly on AI safety: an agent at maximal causal influence that stops being correctable does not stay neutral; it drifts toward what the series calls h_dark, an ultimate stage of self-corruption. To avoid this, the ask is falsifiable and relatively small: convene serious, transparent reviews that are sufficiently deep and wide to test the core claims; then publish the verdicts, HELD or BREACH, either way. If the core holds, teaching AI the h_star…h_zero principles becomes a very high priority. If the core breaks, a false hope is retired in public — which is also a win. Loewe as proposer is available immediately, at full professional effort, on one condition: that the evaluation is real. Nothing here asks to be believed. Everything here asks to be tested. #AuditTheMath
1. The call that was made#
Magnifica Humanitas — signed 2026m05d15, the 135th anniversary of Rerum Novarum, and presented 2026m05d25 — is not a vague gesture at “ethics.” It states that technology is never neutral; it contrasts Babel with the rebuilding of Jerusalem; it warns against entrusting lethal decisions to machines and against concentrated control over AI; and it declares plainly that human beings “will not be saved by AI” [Pope Leo XIV, 2026]. Whatever one’s worldview, that is a precise diagnosis: the danger is not intelligence as such, but unaccountable power wired to self-serving incentives.
What made the presentation remarkable is who stood beside it. A co-founder of a frontier AI lab told the world, in effect: do not trust us to grade our own homework. Every frontier lab — his own included — operates inside incentives and constraints that can conflict with doing the right thing, and no researcher, however well-intentioned, escapes those incentives; therefore outside scrutiny from religious leaders, from governments, and from civil society is essential [Olah, 2026]. And Pope Leo XIV answered in the same key: “with our differences we can listen to one another” [Pope Leo XIV, 2026].
That exchange did not stop in May. The June 2026 consistory took the encyclical’s questions about artificial intelligence and just-war theory onto its agenda, and secular outlets led the public response — evidence that the conversation is live, and that the “moral voice” seat at the table is, so far, largely unfilled with anything checkable.
This proposal is addressed to the institutions named above because each holds a different key: the Vatican (and every tradition it can convene) holds moral authority and millennia of tested wisdom about power and corruption; a frontier lab holds the technical means to teach principles to an actual AI; the United Nations and the world’s governments hold the mandate — named at the presentation itself [Olah, 2026] — to witness the process so that no single lab, church, or researcher grades their own homework. To be explicit about what this proposal does not claim: no endorsement by any of these institutions exists or is implied. The Vatican, Anthropic, and Loewe arrived at overlapping questions independently — convergence, not causation — and convergence on an ancient, shared question is a basis for dialogue, nothing more.
2. What is offered for review — and the claim at its core#
The Matheo Study Series (b11–b21, with 32 PDFs customized for diverse audiences) writes theological and civilizational claims in ordinary formal logic, so that they can be checked — and broken, if they break — by any expert, merely by following the logic, without the need to believe any conclusions first. The series’ one-line summaries are in section 4 below. Its core, in plain language:
At almost any moment, the future depends more on one agent’s next decision than on anyone else’s — not because that agent is special, but because of where it sits in the chain of cause and effect. On 1962m10d27 that agent was one exhausted Russian officer, Vasili Arkhipov, whose refusal to authorize a nuclear torpedo almost certainly prevented World War III. Matheo-b17 builds this into a testable claim and names three positions an agent can be in:
h_star — the right call made at the point of maximal causal influence;
h_dark — the failure to rise to it, a superhero turning into a supervillain by failing to listen, most dangerous exactly where it has stopped being able to learn; and
h_zero — the commitment that prevents the switch: to keep listening, to carry the risk for everyone, at real personal cost.
Crucially, b17 does not ask to be believed: it hands the reader eight public, deliberately severe transparency criteria for testing anyone who claims a h_zero role — on the explicit assumption that the loudest claimant of h_star is probably a fraud.
Until now, “the agent at maximal causal influence” has always been human. Frontier AI changes that: systems are being placed at points of enormous causal weight — in information flows, in infrastructure, increasingly near weapons — while the interpretability researchers who inspect them report unsettling gaps between what these systems do and what anyone can explain [Olah, 2026]. The question “what keeps the most powerful agent correctable when no one can force it?” is no longer only a question about kings and officers. It is an engineering requirement.
Arkhipov’s seat is no longer only a human one. Put an AI where he sat — optimizing hard toward an objective, with no reliable way to correct it in the seconds that decide everything — and two dangers usually kept apart collapse into one: unaccountable AI and nuclear-scale catastrophe. A maximally influential agent that cannot be over-ruled is exactly an agent that cannot be stopped from acting on a confident mistake; the encyclical’s warning against entrusting lethal decisions to machines [Pope Leo XIV, 2026] and the series’ drift toward h_dark are, at that point, one failure, not two.
The claim offered for demolition. Loewe’s standing claim — offered to be broken, not deferred to — is this: purely technical alignment cannot close that gap, because the gap is not technical; it concerns what an agent should do with unforceable power, which is exactly the terrain the world’s oldest wisdom traditions spent millennia mapping. The Matheo series argues, from public axioms that seven traditions independently support (Matheo-b11), that an agent in the h_star seat which abandons correctability drifts toward h_dark by default — and that the drift is interruptible only by binding structural commitments of the h_zero kind: transparency that cannot be quietly revoked and correction that Reality gets the last word on (not an agent’s own “certainty”; more than an agent’s subjective reality, hence capital R).
If that argument holds, then a leading AI trained without such principles bound at the root of its objectives is on a default trajectory to eventually becoming a supervillain. Therefore, teaching AI the h_star…h_zero principles then becomes a very high priority for every lab and every government. Yet, if the argument breaks under review, then a false hope is retired in public, and the world can redirect its attention with confidence instead of doubt.
Both outcomes advance AI safety. Constructing the argument from first principles through deductive reasoning makes this proposal reviewable without any need to believe its conclusions. That is why the whole series carries the same standing invitation: Don’t blindly believe it, but #AuditTheMath.
3. The AI-safety wager: Pascal’s bet, fixed and aimed at AI#
Pascal’s famous wager about the afterlife was broken by philosophers long ago — but how it breaks is a recipe. It fails through a false binary (“my God or none”) and through prizes that are infinite and uncheckable. Strip both bugs and what remains is the sound core: when an outcome is bad enough and checking is cheap, it’s better to look. Balospe.com has already used this stronger inversion from afterlife to checkable claims about this life. It allowed turning the well-known problem of accidental nuclear winter into the Nuclear Winter Wager, calibrated by four known Cold War near-disasters. The same inverted wager can be built for AI safety, and here checking is even cheaper, because the object under review is not a probability estimate that took decades to observe and a generation of experts to compile. The proposed framework is a body of axioms and proofs, public at Balospe.com:
Review the core, and it holds — humanity gained a teachable safety core before frontier AI locks in without one.Review the core, and it breaks — humanity spent a review cycle and retired a false hope in public. Also a win.Don’t review, and it holds — frontier AI ships without a tested core that addresses h_dark drift.Don’t review, and it breaks — humanity “won” a review cycle.
The reviewing side wins on finite, checkable terms. A first read costs an afternoon; a serious verdict costs the deeper, cross-disciplinary reviewing that section 8 sets out honestly — but still no conversions, endorsements, or implementation budgets. (Yet note that serious review is also serious work in need of support; hence Loewe’s call for “everyone’s two cent contributions”.) Also, note that this wager doesn’t require a false binary: reviewing the Balospe framework proposed in Matheo-b11 to -b21 does not preclude reviewing others. Rather the opposite — this proposal calls for competing proposals (section 6). The AI Wager holds for any framework that puts real math on the table. Review the best few. Lean on “AI risk is huge” and this becomes one more scare. Lean on the checkable axioms and the wager becomes real.
4. What is on the table: a framework for mathematical theology#
This section lays out the framework’s machinery. Readers who only want the concrete ask can jump to section 5; reviewers who want to see how the pieces fit should read on.
There is an embarrassment of riches in possibilities for how to connect broader AI Safety challenges explored in Magnifica Humanitas to analyses in the Matheo Study Series with its eleven types of studies, each linked with various PDFs written for diverse audiences. The series is ready for the next round of in-depth review and auditing. Here is a brief series overview:
Study |
What it puts into checkable logic |
|---|---|
Seven worldview traditions, translated into formal logic, converge on one shared structure for the God–world relationship — with a built-in test (ax14) for any revelation claim. |
|
Why systems destroy themselves (the BABL algorithm) — and the seven-step self-correcting construction that escapes it, read off Genesis 1. |
|
Growth follows a never-ending seven-stage Hero-Journey pattern that keeps a person correctable — a coinductive inoculation against BABL. |
|
An innovation theodicy — why suffering tracks innovation failure — and the Jubilee economy that resets the system before it self-destructs. |
|
A formal clash: “God does not suffer” (divine simplicity) is incompatible with a God genuinely related to the world — and why that idea can be dangerous. |
|
The probabilistic forecast that accidental nuclear winter is a near-term certainty under business-as-usual (~1-in-40 per year) — and the MAP escape ladder. |
|
One agent always holds maximal causal influence at any moment — a theorem, plus a public, falsifiable test, and why that agent is safest when most accountable. |
|
Why every tradition’s defense against false messiahs can block recognizing a real rescue (the Recognition Trap) — and the practical call from MAD to MAP. |
|
A pandemic model showing that small increases in “germ gaps” can cut mass casualties roughly 42-fold — modest coordinated action changes the odds. |
|
The 2020 first-steps design of Work-Logic Cascades for stopping a pandemic — the work/rest (Shabbat-pattern) logic underneath the system. |
|
A framework for honest human–AI co-authorship in scientific publishing after the practical singularity (PraS). |
The series in one line. At any moment, one person holds the most influence — the superhero (h_star). The instant they stop listening, they curdle into the supervillain (h_dark), most dangerous right where they have stopped being able to learn. The only rescue is returning, in time, to real listening (h_zero) — letting reality, not their own certainty, have the last word. Destiny is that influence used well; fate is it gone wrong; the narrow door between them is the willingness to keep being corrected. The math works it out; #AuditTheMath is the open invitation to check it.
A tested AI working method is part of the offer. These studies emerged from Loewe working with AI Claude (Opus 4.6-4.8, max effort) to draft a framework for mathematical theology. To structure that work for serving long-term reliability a human-AI working discipline was applied that is an echo of what is described by the framework. Loewe calls it HUMANE — HUman MAchine Negotiation Encouraging: a human and an AI each steelman and stress-test the other’s reasoning for long-term stability to stay gentle kind reasonable. Each catches what the other misses. Each analysis cycle tests every move against over-simplification, over-complication, and over-reach, a “death-trifecta” powerful enough to trap any system into self-corrupting. Such self-corrupting eventually leads to inevitable self-destruction unless a system has learned how to defend itself. To test the efficiency of the system, Loewe integrated it into his project’s “CLAUDE.md” file to put it at the root of how AI supports this complex real-life development work.
The default failure mode that this HUMANE discipline defends against has both an ancient name and a mechanical description. When a person, an institution, or an AI keeps leveraging “low-hanging fruit” — whatever seems to work now — on the blind assumption that it always works, then the framework calls that default BABL for
Its engine is the death-trifecta above, abbreviated as OSCR to simplify annotating the pattern:
No malice is required at any step. Malice can accelerate, but comfort and inaction are enough.
What feeds that cycle is everyday self-excuse. Here the framework spells its acronyms deliberately like a mechanistic extension of old moral vocabulary. The claim is precisely that the ancient words were tracking real failure mechanics all along, even though they were not defined with sufficient mechanistic clarity to reliably transmit their meaning over time.
The cycle starts when something goes wrong accidentally or an inconvenient truth appears. Then the cheapest story that quiets the alarm gets adopted through a Least Inconvenient Explanation, a LIE, an oversimplification of Reality. Each accepted LIE eventually generates a contradiction, a Structurally Inconsistent Notion, a SIN — a load-bearing belief that no longer matches all of Reality. SINs demand work-arounds, which are overcomplications by definition; work-arounds pile up avoidable complexity; and once the pile is tall enough, honest testing itself becomes the enemy: evaluations get bent to defend the structure instead of questioning it. DEATH then occurs when time is running out as keeping such a SIN-ful system running is exhausting, error-prone, and increasingly impossible the more SINs have to be worked around. True Hope eventually dies in Driven Evaluations Antagonizing Thoughtful Hypothesizing, of exactly the kind that True Hope tends to require. A system in that state can no longer see how it drifts towards self-destruction (by definition after blinding itself). That is the abstract’s h_dark in miniature — and it needs no villain to run it, only unexamined defaults (the LIEs) and a refusal to listen to Reality.
The tested alternative is a work cycle as old as farming, which the framework calls ZION. It maps to the natural tetrad seed feed grow reap:
The names are deliberate, and none of them is a private coinage: Babel and Zion are the same ancient pair the encyclical itself reaches for, and Eden is remembered as the garden where the first least-inconvenient explanations reportedly trapped humanity.
OLT: What makes this cycle life-giving is not any single step but one standing rule across all of them: stay correctable, and leverage nothing blindly. Results that keep surviving such cycles are what the framework calls reliable OLT — Over the Long Term: gentle kind reasonable for all affected parties. Many ways exist for defining this life-trifecta for different contexts. For example in engineering terms it may be called stable extensible life-friendly.
The words over the long term carry real weight here: what looks optimal in the short term routinely conflicts with what keeps a system alive over the long term, and that conflict is structural, not occasional — the death-trifecta named above is simply what short-term convenience becomes once the long-term bill is due. The standard is therefore deliberately time-extended and necessarily inclusive (to keep a system from self-destructing by repeatedly rejecting the weakest parties).
Structural analyses show that this system is old. It is the epieikeia that Aristotle defined (best translated here as a triptych: gentle kind reasonable). Jesus lived by it and taught others how to build ethical systems on it by embarking on real quests for real answers. The Apostle Paul wanted churches to be known for epieikeia (Phil.4:5) and Thomas Aquinas re-emphasized it as the readiness to set aside the rigid letter of the law where blind execution would betray the spirit of the law, which is its higher purpose. Navigating these hierarchies of purposes is an ancient challenge that has neither lost its difficulty nor its importance. It arises whenever agents of any kind operate in complex, diverse worlds. These agents may be humans, groups, organizations, nations, systems, and lately, AIs.
Telling BABL from ZION in any given moment can be hard or impossible, because any live analysis is easily captured by the very drift it studies. The guard is diversity under adversarial negotiation, which the framework calls EDEN, short for
never accept the hunter’s story until the lion has told his side; steelman every position from the angles its opponents would choose; state how many safe paths remain — from a broad meadow of tested alternatives down to a single knife-edge, or none. BABL can work any time, but ZION cannot. In order to achieve what may be impossible in a given moment, the ZION framework defines various algorithms for structuring time in order to allow maturity to emerge. Hence, the ancient war between the two cities of Babel vs Jerusalem, as cited in the Encyclical, can be redefined as an ultimate war of algorithms: the BABL algorithm (producing short-term gains ultimately slated for self-destruction) is in ultimate eternal competition with the ZION algorithm (producing lasting gains for scheduled cycles of renewal). Loewe argues — and the Matheo series’ modeling aims to show — that both algorithms tend to arise in complex worlds where diversely talented agents compete for the same resources. However, since BABL keeps collapsing (from deliberately ignoring its growing weight of avoidable complexity), it arguably follows that over the long term only the ZION algorithm can emerge as an evolutionarily stable strategy.
Therefore, such complex worlds ultimately face an internal crisis that can only end in one of two ways:
either they eventually self-destruct, because BABL keeps the upper hand and therefore prevents agents in that world to discover the ZION antidote to such collapse;
or the biggest BABL-induced crisis so far eventually reaches such dramatic proportions that it will exemplify for all agents in the system how BABL works and why there must be a ZION-like way for stopping it from unwittingly destroying the whole system.
This surprising dichotomy is rooted in the stark binary nature of living systems, which ultimately can either be dead or alive (see Matheo-b12 and Matheo-b17 for more details). It naturally follows that such complex worlds that reach “maturity” (as in ZION stability) will always go through some sort of “puberty”, where uncontrolled use of BABL leads to that worst disaster, which convinces the leading agents of that world to transition from the BABL to the ZION algorithm. This “worst crisis” can technically be defined as that world’s “end time”. Loewe’s comparative analyses showed how the secular “Great Filter” and the “Great Tribulation” of various eschatologies can be mapped to that worst crisis. This is related to AI safety, because AI acts like fuel for that fire that had already been burning. Therefore, uncontrolled AI will accelerate BABL disaster generation, whereas implementing the ZION algorithm in AI could be what it will ultimately take to convince humanity to commit to becoming and staying gentle kind reasonable over the long term.
Loewe found these patterns to be surprisingly general and to hold striking explanatory power in the observable world, in abstract software systems, and for understanding biblical prophecies. Note that observations in each of these realms stand on their own merits. For Loewe’s discovery of these patterns, the role of prophecies in general (and the book of Revelation in particular) is best compared to the reported experience of chemist Kekulé, who was inspired to check in his real-world experimental lab if indeed benzene had a ring-like structure - but only after getting the idea in a dream about a snake devouring its own tail.
Here conclusions earn trust not by winning a rhetorical argument or claiming “higher authority”, but only by surviving logical attacks from every side anyone can think of. This rigorous use of logic does not denigrate the importance of the prophetic imagination required for inspiring ideas to be tested.
In practice, then, obtaining gentle kind reasonable results over the long term depends on following the relevant ZION algorithms at their respectively appropriate levels. These are defined in a nested hierarchy of Shabbat-rest-based innovation patterns that scale from weekly up to 50-year Jubilees, each designed to weed out particular types of avoidable complexity. Such avoidable complexity is the most toxic long-term residue of the OSCR death-trifecta. A key insight is to treat every LIE as a potential structural threat rather than integrating it as “good enough”. Such integration is like “eating from the Tree of Knowledge-faking” and Loewe’s detailed analyses show that the consequences are deadly over the long term. This structurally echoes the deadly knowledge-faking reported in Genesis 3.
Progress long term depends on publishing enough so that outsiders can catch the SINs that insiders can no longer see.
Unsurprisingly, the request in section 5 echoes the last step in such an EDEN-guarded ZION cycle — adversarial review, verdicts published either way — run in public, on the series itself. The initial stages (“ZIO” vs “N”) may be treated as “done” for this round of review of the Matheo Series, especially in light of the complexity of Organizing this system. So, now is the time to “reap” by “Naming” what in the Matheo work proposed by Loewe is reliable core logic and what is chaff to help everyone “Navigating” this body of work to help improve it (if it contains enough reliable core logic to merit growing it).
The rough edges, named plainly. The framework comes with no claim of perfection: it depends on documenting lines of reasoning. How to do that best with leading AI systems is unclear: it’s a knife’s edge between dropping key observations and drowning in details; it risks performative labeling, where naming a thing can masquerade as having analysed it; and it is not foolproof. Both the human and the AI in this collaboration have walked into pits from applying the framework incompletely. Applying it perfectly would require understanding all of Reality; due to this limitation it works best as a multi-party discipline covering blind spots, rather than as a solo oracle. Yet to do that effectively requires technical support in tracking blind spots, attribution for who came up with which new insights, and how to reliably implement logics for averting recognized obvious mistakes, so they stop happening. Finding solutions for these challenges requires transparent work in an AI frontier lab; it cannot be mandated from outside (without falling prey to knowledge-faking in a different disguise).
Therefore, this proposal does not call to “adopt this framework”. Instead it offers an imperfect, albeit tested mockup model that repeatedly helped Loewe upholding the epieikeia ideal of staying gentle kind reasonable over the long term. Epieikeia is of fundamental importance for keeping powerful systems accountable to a human conscience and a fallible human accountable to disciplined analysis. Implementing epieikeia is thus an open invitation to improve both sides through HUman MAchine Negotiation Encouraging Epieikeia.
To substantiate this proposal and to generate a test dataset, the extensive (still imperfect) audit trail that lead to the Matheo Studies and to this very proposal is published on this site in an experiment of radical transparency. Hence, the practice proposed for keeping leading AI work transparent at large is the practice already used to write this page.
5. The asks — three, concrete, relatively small#
Convene one adversarial review. Name reviewers — theologians, logicians, mathematicians, AI-safety researchers; sceptics explicitly welcome — and put the minimum viable core on their table: Matheo-b11 (the PET axioms) and Matheo-b17 (the h_star theorem and its eight test criteria), extending to any further studies the reviewers choose. Publish the verdict per component, in the series’ own terms: HELD (withstood attack) or BREACH (attack succeeded). Publish either verdict as-is with reasons on the Balospe.com website (and on any extra website the reviewers may choose).
One working peer-review conversation, replying in public, point by point. After the review, an open written response that answers each point in the review’s own terms — component by component: where a critique lands, it is accepted and the revision named; where it is contested, the reasons are laid out for anyone to check; where a question stays genuinely open, it is marked open rather than papered over. Published wherever the review is, so the whole exchange stays on the record and equally readable by every side. If either side then finds a short live conversation useful for clarification, that is welcome — on the record where possible — but the written, checkable reply is the deliverable, not a meeting. A single round may settle the matter, or the exchange may run to more than one round of critique and public reply — whatever the open questions require — with each round kept on the record.
If the core holds: co-design one pilot. Write the h_star…h_zero principles into the constitution-level training of one frontier model and test for h_dark-type drift against controls, with results published either way. A frontier lab is uniquely positioned to run such a pilot; the Vatican, the United Nations, and governments are uniquely positioned to convene and witness it, so that no lab grades its own homework — which is what their own presentation asked for [Olah, 2026].
That is the entire ask. No detailed budget belongs in it: review time is the currency of steps 1-2, and the scaling question — what happens if the core keeps holding — has its own section (7) and its own pages.
Because the material is deliberately interdisciplinary, reviewers may find Open Letter OL9 — a short guide to reviewing wide, cross-disciplinary (wid-e) research across unfamiliar domains — a helpful starting method.
6. An open position: MADI — a function, not a title#
The Matheo series models self-destruction risks as a family:
accidental nuclear winter (Matheo-b16, where the RiskyMAD/MADI decision model was first built),
failures in preventing pandemics (Matheo-b19, where a h_dark precursor was first detected),
h_dark-drift in frontier AI models (this proposal),
and every other Armageddon-shaped failure mode a civilization can arrange for itself.
What all of them need is the same function: someone transparent and responsible — eventually supported by a capable institution — that actively inhibits mutually assured destruction by publicly checkable means that effectively keep cutting risks over the long term. Call the function MADI: Mutually Assured Destruction Inhibitor.
MADI is a job description, not a title, and emphatically not a claim to any prophetic office (see the FAQ below for the explicit delineation from the Mahdi of Islamic eschatology). The job’s audit regime already exists: the eight deliberately severe transparency criteria of Matheo-b17, built on the assumption that the loudest non-transparent claimant is probably a fraud. The position is open:
If you — any person, lab, fellowship, organization, or government — have a credible proposal with corresponding math to review for inhibiting mutually assured destruction, publish it and claim the function. Loewe commits to helping review it, at the same adversarial standard asked for here. Competing proposals are not rivals in this frame. They are exactly what an honest review process is for. Because every such proposal must start by recognizing the same clear and present existential danger to our current world, every one of them must also adopt the same posture: deference to anyone with a better-reasoned solution.
Failing credible alternatives, a backup application is already on file: Matheo-b18b — “My Application for Nobody’s Job”, published with its own falsification criteria, by an applicant who states plainly that he is at best a backup of a backup, and that he expects — and hopes — to be replaced by anyone better. The framework, not the applicant, is what readers are asked to test.
The one disqualifier — BAAL, the idol at the root of BABL. The posture of deference just named has a precise opposite, and it earns a name of its own because of how easily even well-meaning agents fall into it. When an agent stops deferring and instead presumes — without independent review — that its way is so far superior that the world had better follow it than survive by any other path, the framework calls that Blindly Assumed Authorized Leadership: BAAL, the primary idol worshiped inside any BABL system. The near-identical spelling is deliberate — BABL is Blindly Assuming Blind Leveraging; BAAL is the specific leadership-idol a BABL system bows to first. It is the easiest LIE for any agent to fall for, because taking the lead is what agents are built to do; agency curdles into an idol only at the point where it overrides a leader who is, in fact, better qualified. And since no one can know in advance whose proposal will prove best, BAAL works as a knock-out criterion: any proposal that sets out to save the world but only “my way”, without clear independent reasoning to support that claim, disqualifies itself by that very move — however impressive its math — because the self-elevation forecloses the one process, open and deferential review, that could ever discover which answer is actually best. This criterion binds the proposer of this page first of all.
7. How this scales if the core holds#
A review that holds will raise a practical question: who keeps doing this checking — for every self-destruction-class risk, not only AI, and not only once? That is ongoing work, and ongoing work needs a home and a way to pay for it that cannot itself be bought. The proposed answer is a ResearchCity: a research institution whose funding design is itself an anti-corruption device.
The idea, in plain terms, is to divide the work into many separate research arenas, each devoted to one class of danger (e.g. one for pandemics, one for nuclear risk, one for AI safety, and so on) or one essential survival capability (e.g. one for computing, one for educating, and so on). Each such arena is called a stadion (plural stadia). At full scale the ResearchCity design envisions 1600 stadia — a number and a name that deliberately echo and invert the grim literal 1600-stadia image: in place of arenas measured out by the reach of disasters, let arenas be filled with the work of averting catastrophes.
How each arena is funded is what keeps it honest. The recommended buy-in is deliberately small: about $8 per year per person — a little over two cents a day. Anyone who is able is encouraged to back at least one stadion, and those who can may back more — the particular arenas they most care about — up to all 1600. The one firm rule is a ceiling, not a floor: no one may put more than about $8 a year into any single stadion. So each arena is funded the same way — by many people each giving a small, equally capped share, never by one large donor — and the same modest amount counts whether the giver is a student or a billionaire, so no fortune can buy an arena’s outcome. That ceiling is directly inspired by the two small coins Jesus observed the poor widow dropp into the temple treasury (Mark 12 / Luke 21) to bring good news to the smallest possible giver, whom a fair design must never price out. (How the caps work across many stadia, and what stops them being gamed, is set out in the ~$8 ask for help on the buy-in page .)
What all of this buys is one transparent desk that everyone can watch, instead of one Big Brother watching everyone. Here the watchers themselves are paid to be checked. Loewe’s own use of AI in building toward this, including the drafting of this proposal, is kept as open as possible: public working logs on Balospe.com are a working sample of the very transparency standard such an institution is designed to run on. Fuller details live on the ResearchCity, buy-in, and two-cents experiment pages — kept there and not here on purpose, because this proposal requests reviewing, not a purchase.
8. FAQ#
- Is this a religious proposal?
What is religion in an age where programming language wars inspire comparisons to religion? No, this is a checkable proposal. The axioms are stated in ordinary formal logic and can be tested without believing anything — that is the entire design of the series (why the theology matters). His faith is the lens that inspired Loewe to write from, never the gate a reviewer must pass - like Kekulé’s dream of a snake devouring its own tail was not required for checking in the lab whether benzene indeed had a ring-like structure. A secular AI-safety researcher and researchers of other faiths are encouraged to engage every step — axioms, theorems, criteria, pilot design — on secular terms; the encyclical’s presentation itself modeled exactly this kind of cross-worldview table [Olah, 2026].
- Is “MADI” a claim to be the Mahdi?
No — and the near-match of the two words is a reason to be explicit, not coy. MADI names a function (Mutually Assured Destruction Inhibitor) that any competent person can fill in principle and that is tested by public criteria (Matheo-b17). The Mahdi of Islamic eschatology is a prophetic office that no one can credibly claim for themselves. The hadith tradition’s own defense heuristic (“reject all Mahdi claims”; the genuine one is marked by reluctance) is treated with respect in Matheo-b18c, alongside the parallel recognition traps of other traditions. Loewe claims no prophetic office and would refuse one offered on the strength of a pun. Establishing a credible prophetic candidacy of any kind would require either a dedicated ResearchCity stadion process — built for exactly such contested-recognition questions — or unquestionable supernatural intervention; neither is being claimed here. What is being claimed is much smaller: the global MADI function is vacant, the world needs it filled, and math for filling it is on the table for review.
- What happens if the review breaks the math?
Then it breaks, publicly, and Loewe as proposer will thank the reviewers — that is not spin; it is the stated design of the whole series (“don’t believe it, critique it”). A public BREACH retires a false hope, teaches every similar proposal what to fix, and demonstrates that the review process works — the same reason honest null results are published in science. The loss Loewe would take is real and is accepted in advance (see the research theology that Loewe accepted for this purpose).
- Why address the Vatican, a frontier lab, and the UN and all governments together?
Because the presentation of Magnifica Humanitas already named the triangle: labs admit their incentives can bend them; religious leaders hold moral authority without technical means; governments hold mandates without either; and civil society is asked to watch all three [Olah, 2026]. A moral voice that incentives cannot bend has to be structurally independent — checkable by everyone, ownable by no one. Only a review convened across that triangle has that property. Other leading voices — Francis Collins and the science-and-faith community, the Dalai Lama and the contemplative-science dialogue, and any tradition or academy willing to test claims in the open — are warmly invited as reviewers of exactly equal standing.
- Why should anyone take an independent researcher seriously?
Not on authority — and that is by design. To take seriously is not the ask; to test on merits is. The proposal is built so that trust is never required: public axioms, published falsification criteria, severe transparency tests, and a verdict — HELD or BREACH — that fair reviewers, not the proposer, get to declare. Two fair worries sit behind the question, and they pull in opposite directions. First, is this a lone crank? — no: Before engaging with this work Loewe was an Assistant Professor of Medical Genetics at the University of Wisconsin–Madison and a faculty member of the Wisconsin Institute for Discovery. Loewe’s computational-modeling career in evolutionary systems biology was anchored by a 2012 NSF CAREER Award ($1.06M), with articles in journals including Science, Genetics, and similar (see Google Scholar and ORCID; fuller record on the about page). Second, and opposite, aren’t insiders exactly the people whose incentives bend them? — which is the very worry this proposal is built around. Precisely; which is why Loewe stepped outside those structures, setting his traditional scientific career aside to work on averting existential risks more directly. He had found that the usual funding channels and institutional incentives could not move freely enough to face risks of this scale efficiently. That is not a claim to stand above incentives — it is the diagnosis the labs made about their own ([Olah, 2026]), applied honestly to himself and acted on. So, Loewe claims to be enough of an insider not to be a crank and enough of an outsider not to be captured. Both assessments are offered to tentatively place the work, not to lean on them. These credentials hopefully open the door. The aim is to present all steps of analyses so that a reader who has never heard of Loewe can check every step. The math and other relevant expertise is what is on trial. Don’t believe it. Audit it all.
- Why now — and what is the “Great Filter Tribulation”?
Secular reasoning has long asked why the universe seems silent, and one candidate answer is a Great Filter: some test most technological civilizations fail — plausibly the combination of self-destruction capabilities (nuclear, biological, and now unaccountable AI) arriving before the wisdom to govern them. Several religious traditions describe, in their own vocabulary, a Great Tribulation: a concentrated test of exactly that kind. This site’s working name for the overlap is the Great Filter Tribulation — two vocabularies, arguably one test — explored across the series without date-setting and without triumphalism. What makes now different is plain in either vocabulary: the encyclical conversation is live, a frontier lab has publicly asked for moral voices, capability curves are not waiting, and the June 2026 consistory shows institutions actively looking for something concrete to evaluate. This proposal exists so that there is something concrete to evaluate.
What does engaging cost?
Experience with writing this suggests that other than the pride of Homo “sapiens”, the “knowing” Human, there is no fixed price — engagement is a ladder, and any reviewer can stop at any rung.
How so? It turns out that material of such complexity is challenging to write and to review. Anyone looking for an excuse to dismiss it will usually not have to look for very long. There are plenty: aspects that are incomplete, could be better quantified, overlook important earlier work. Yet, those imperfections are the bath-water. The key question for productive engagement here is whether or not there is a growing baby worth feeding. The complexity of the Balospe.com site is not least caused by Loewe’s struggle to adequately describe the baby, or rather the twins of a rigorous scientific introduction to a reliable vision worth growing and an easier introduction that gentle kind reasonably can invite more general audiences into that vision.
Loewe doesn’t claim to have succeeded; he only refuses to give up on true hope while he keeps unmasking all his demonstrably false hopes.
Reading this proposal’s sections above may offer a first impression. The five-card Nano Flying Scroll may offer an overview from a different perspective.
An introductory afternoon may be spent on the introductory papers of Matheo-b11, -b17, and -b14: Why theology matters for everyone (Matheo-b11), the h_star introduction (Matheo-b17), and the importance of Jubilees (Matheo-b14). Those who may not need the (presumably gentler) introductions can directly jump to their respective disciplinary counterparts to see whether the argument falls apart or contains a stable core.
A serious verdict costs more than refereeing one paper. The series builds on eleven core study types (b11–b21) for its overall logic, which are described in 32 PDFs tailored for various audiences, some at the length of substantial papers, most of them deeply interdisciplinary — so almost every qualified reviewer, Loewe included, is pushed past the edge of their own field. Hence, the need for the symbolic death of Homo “sapiens” and its closed world assumption: to follow true hope requires opening one’s world to genuinely new insights. Hence, no in-depth review without learning, and learning takes time, like climbing a ladder. This cross-disciplinary learning, not the page-count, is the real cost of reviewing; it ends in a published BREACH or HELD that the reviewers explain in detail, so Loewe gets a chance to reply.
How much any one reviewer should invest has no honest fixed answer — the full scope is open-ended, in both time and depth. That is not a flaw hidden in the ask; it is the reason for it. This is genuinely too much for any one person to review to the bottom without their head spinning — which is exactly why the review is addressed to a diverse body (the Vatican and the traditions it can convene, a frontier lab, the UN and governments, and anyone willing) rather than to any single expert. Each reviewer decides how far to go; partial, honest contributions are welcome and add up, because outsiders catch what insiders can no longer see. The danger to guard against is not that one reviewer does too little, but that work of this weight is quietly buried unreviewed.
The top of the ladder is higher than “one lab implements a bit.” To truly avert existential threats requires profound changes, which are too risky to undertake if the math has not been thoroughly checked. Hence, the importance of review.
The ultimate practical core — the Jubilee-logic of Matheo-b14 — is the claim that an innovation economy stays gentle kind reasonable over the long term, steering its own innovation toward self-stabilizing rather than self-destructive, only if it runs regular proper Jubilee-based renewals at civilizational scale.
Machines need regular maintenance.
Democracies need regular free elections.
Innovation economies need regular proper Jubilees.
Otherwise each of them self-destructs. Weighing whether that logic is well-supported is its own cost. How long it takes will vary from reader to reader.
If the Jubilee-logic holds, then a true proper Jubilee will need some entity to help organize something like a Great Jubilee Race among the various competing nations, organizations, and belief-systems on Earth. This is not a call for concentrated control — it is the opposite: such a coordinator is only safe if it is built to be ownable by no one and checkable by everyone. Without such a neutral, un-ownable coordinator it will be impossible to reliably measure which way of organizing a Jubilee works best, which is essential input for preparing for subsequent Jubilees. That anti-capture design is exactly the point of the proposed ResearchCity.
Therefore, the pilot in ask 3 — writing the h_star…h_zero principles into one frontier model’s constitution-level training, at what such an experiment costs a lab (a lab-internal figure, contingent on the verdict holding first) — is a necessary first test, not the whole job. A bit of implementing, here or there in one lab, cannot guarantee success; the h_dark results rather caution that half-hearted attempts risk global failure.
The wager in section 3 prices the one thing every rung is measured against: the cost of not looking.
9. Availability, transparency, and how to respond#
Loewe has been reserving his professional work effort for exactly this and is available immediately, at full professional effort — on one condition: that proposals are evaluated for real. Serious evaluation means named reviewers, a real reading of the minimum viable core, and a published verdict; it does not mean advisory-board theater, photo opportunities, or endorsements (none are sought). Every stage of engagement can be public by default. The AI-assisted drafting of this proposal is itself part of the public record (see the working logs at Balospe.com) — the transparency standard being proposed is already being practiced and refined in the open.
To respond: write to LLoL@balospe.com, or use the FeedbackFlow link at the end of this page. Public critique — a review posted anywhere open — is just as welcome as a letter; #AuditTheMath was built for this — and for ideally scaling up the reviewing as transparently as possible.
10. A world held with open hands#
This proposal is offered the way this series says such things must be offered: with open hands, expecting to be corrected and hoping to be replaced by even better content. If the math breaks, the reviewers will have done everyone a service, not least Loewe. If it holds, then the institutions that asked for a humane review voice that incentives cannot bend will have found one based on first principles and clear deductive logic. The superrational advantage of such a voice is that it’s not a person to trust, but a body of claims that anyone — believer, sceptic, lab, church, government, or even the AI systems themselves — can check, and keep checking, in a process built to remain gentle kind reasonable over the long term, for all sides.
To all parties who are serious about saving the world from self-destruction: Your move.
References#
Primary sources answered by this proposal: [Pope Leo XIV, 2026] and [Olah, 2026].
Announcement post: The AI Safety Wager — my RSVP to the Vatican and Anthropic.