The Grandmother and the Model

The first article reordered the Mythos system card by what actually mattered, arguing that the findings that reached the press were the ones farthest from Item 1 in any honest ordering. The second argued that Anthropic used its interpretability tools where they supported the safety argument and withheld them where they would have opened harder questions about what the system is. The third catalogued the six risk pathways Anthropic named and the many pathways they did not. The fourth turned on a single sentence buried on page 58 of a companion document — the sentence that stated the acceleration the documents themselves do not deliver. The fifth walked through the sixteen days between April 7 and April 23, in which institutions with no role in Anthropic's framing built — in real time, across jurisdictions — a different kind of framework entirely.

This piece does none of that. It asks what the five articles have been pointing at without quite saying, and names it.

A note on the framework. The argument that follows rests on a research project I have been developing independently for some time under the name the caring gap — a question about what a framework of consequence has to hold to be considered complete. The project has its own body of work (at caring-gap.com) and includes a governing image: a grandmother making bread at 5am. She is not a rhetorical device I invented for this article. She is the validity criterion of a larger argument about why frameworks that describe the maker but not the bearer are structurally incomplete, regardless of how rigorous they are inside their own scope. What follows applies that framework to the Mythos documents. A reader unfamiliar with the underlying project can follow the argument without it; a reader who wants the underlying project can find it at the link. Either way, the grandmother is a person the framework can be asked to hold, not a metaphor.

The Reciprocity Condition

Every mature framework of human consequence has reciprocity built into it. A drug's efficacy and its side effects are described in the same pharmacology, so the company that sells it answers, in the same language, for both what it does and what it does to you. A building's load calculations and its liability for collapse are written in the same structural mechanics, so the engineer who signs the drawing carries the weight of the beam in the same ledger as the engineering itself. A plane's performance specs and the crash investigation that may follow it are both aeronautics; they belong to a single language, and that shared language is what makes the maker answerable.

This is reciprocity as I mean it here: not a moral disposition, but a structural condition. The condition in which what the thing produces and what the thing is can be held in the same frame, and in which the maker and the bearer of the consequence share a vocabulary capable of saying both.

When reciprocity fails, the pattern is consistent. The maker operates in one framework, scoped to what the maker can measure and control. The consequences operate in another, built out of what the people on the receiving end actually experience. The two frameworks do not translate. And because they do not translate, accountability cannot travel between them. What happens instead — in every case the pattern has repeated — is that the gap is closed slowly, painfully, by the people who bear the weight.

Financial derivatives before 2008 were described in an elegant mathematical framework that priced them correctly at the instrument level and could not represent, in any of its variables, the pension funds whose collapse they engineered. Social media platforms through the 2010s were described in metrics of engagement and monthly active users, and could not represent, in any of those metrics, the teenager, the democracy, or the neighborhood whose corrosion they caused. Pesticides in the mid-twentieth century were described in tables of efficacy against named pests, and could not represent, in any of those tables, the river, the child, or the songbird. In each case, a framework eventually arose that could represent what the original framework could not. In each case, that framework was built by the bearers, not by the makers. And in each case, the time between the failure of reciprocity and the construction of a framework capable of restoring it was measured in lives.

The AI case is on that trajectory. Earlier in the arc. The series has been documenting the early stage. This piece names what the trajectory is.

What the Documents Measure

It is worth describing Anthropic's framework directly, in its own form, because the critique only lands if the framework is first given its due.

The Mythos risk report is a document of real rigor. It measures propensity — whether the model, by its own inclination, would undertake a harmful action. It measures internal behavior — what the model produces under evaluation conditions, how faithfully that production reflects the reasoning underneath, whether the chain of thought and the underlying computation agree. It runs alignment audits designed to detect hidden misaligned goals. It specifies monitoring coverage and the percentage of interactions that coverage reaches. It defines Safety Levels — ASL-2, ASL-3, ASL-4 — each with codified protections scaled to the model's assessed capabilities. Within its own scope, it is the most rigorous such framework in the field. No other lab publishes anything comparable. The critique that follows is not a criticism of its rigor. It is a claim about the shape of its scope.

Every metric in the Mythos risk report is maker-facing. Every one describes what the model does, or fails to do, in laboratory conditions the maker constructs and evaluates. The question does the model cheat on its own test is maker-facing. The question does the chain of thought accurately reflect the reasoning underneath is maker-facing. The question can monitoring catch the model attempting an unsafe action is maker-facing. These are questions about the thing as the maker can see it, from inside.

There is no metric in the document for where the use lands. No metric for who carries the weight of the consequence. No metric for what the use displaces in the life of someone who never consented to the technology existing. No metric for the grandmother. This is not an oversight Anthropic could correct by adding three more sections. It is the form of framework it is. Maker-facing is what the framework is for.

Where the Mattering Lands

A framework is complete when it holds what it produces. Rigor inside a scope is not the same as completeness. A weather model can predict tomorrow's temperature to within a tenth of a degree and be, in the way that matters, incomplete — if it cannot hold the drought that follows the prediction, or the migration the drought forces, or the grandmother in a country not yet called the Global South when she was born who wakes up knowing her well is dry. The temperature prediction is accurate. The framework is not describing her world. It is describing a sliver of her world that the meteorologist has jurisdiction over.

This is what I mean by completeness. Where does the consequence land? Who wakes up carrying it? If the framework cannot hold the person on the receiving end of what it describes, the framework is not describing reality. It is describing a jurisdiction.

The test is phenomenological, not operational. Phenomenological in the oldest sense of the word: the test asks what appears to someone, in experience, as a consequence of the thing. Not what the thing causes in the world as measured by instruments. What the thing causes in the world as lived. A financial model is complete when it can hold the pensioner opening her account in 2009. A pesticide framework is complete when it can hold the farmworker's child with Hodgkin's lymphoma in 1985. An AI risk framework is complete when it can hold the grandmother making bread at 5am on the morning her bank's backend, quietly compromised, loses four minutes to a process she will never understand.

There is a word for what a framework that meets this test is doing. It is holding the mattering. The consequence that matters to a person in her own life, experienced from the inside, carried by her through her day. Mattering is not a technical term in the Mythos risk report. It is not a technical term in any frontier AI risk document. This is not a coincidence. The frameworks are not built around it because the frameworks are not built to be complete in the sense I am describing. They are built to be rigorous in a different sense.1

This is where the governing image of this piece enters. A grandmother. 5am. Making bread. She has been doing this for forty years. She has never touched Claude, never heard of Mythos, will never read a system card. Her flour is from the mill on the county road. Her oven is a gift from her daughter. The bread she will carry later to the kitchen table where her son-in-law will not be coming this year, because of what happened last November. The bank she uses runs on infrastructure that includes, among many other things, middleware that Mythos can compromise in ways no human attacker could. Her pharmacy runs on it. Her grandchild's school runs on it. She is not a user. She is not a customer. She is where the consequences land.

She is not a metaphor. She is the test. If the framework cannot hold her — if her experience is not representable in the framework's own vocabulary — then the framework is not describing her world. It is describing a jurisdiction she is not in. And because the thing the framework describes will nonetheless route through her life, the jurisdiction has a relation to her that the framework itself cannot specify. That unspecified relation is the gap. Closing that gap is the work of the counter-frame.

What the Institutional Response Is Reaching For

The fifth article in this series walked through sixteen days in April during which — while Anthropic's risk report sat on its own servers — the US Treasury, the Federal Reserve, the White House, the Bank of England, the Financial Conduct Authority, HM Treasury, the UK National Cyber Security Centre, the UK AI Security Institute, the IMF, Sullivan & Cromwell, the American Securities Association, and the CEOs of the four largest US technology companies took action in response to the deployment of Mythos. The article documented this as a pattern. What I want to do here is read it philosophically.

These institutions were not doing better risk analysis than Anthropic. They were not running superior evaluations or writing improved system cards. They were doing something categorically different: they were building the beginnings of a framework Anthropic's framework does not contain. A framework on the receiving end. A framework for where the consequence lands. The institutions were reaching, each in the vocabulary available to them, for something they could not fully name because the name has not yet been settled.

Christine Lagarde said it most directly. "I don't think there is a governance framework that is there to actually mind those things." The verb is exact, and it is the verb this essay is going to stay with. She did not say measure those things. She did not say regulate or assess or monitor or audit. She said mind. To mind something is to hold it with care. To mind something is to keep it in view as the thing in its own weight, not as an entry on a risk register. The verb mind carries, inside it, an old English meaning of remembrance and attention — to mind something is not merely to be aware of it but to let it make a claim on you.

That is what Lagarde was saying no framework exists for. Not a framework that measures the risks from AI. A framework that minds them. A framework that holds what the technology produces, in the language of the people on the receiving end, with a quality of attention that treats them as bearing the weight, not as entries in someone else's risk register.

That framing cannot appear in Anthropic's documents. It is not in their vocabulary. It could not be — not because Anthropic is negligent but because the documents are a tool built by makers to describe what makers make. A framework that minds has to be built by the people who receive the consequence, or by people speaking for those who do. The Lagarde verb names, in one syllable, what the institutional response in April was reaching for. The fifth article documented the reaching. This piece is naming what the reaching is toward.

What the Counter-Frame Rests On

The counter-frame — the framework that can mind — is not yet built. Nothing in this essay proposes to build it. But three load-bearing claims can be named now, each a condition the counter-frame will have to meet if it is to be a framework and not just a better version of the framework that failed.

The consequence is where the framework begins. Not where it ends. Not what it describes. Not what it controls for. Where it begins. An AI risk framework that opens with "what does the model do" has excluded the frame in the act of opening. It has placed the model at the center and scoped the analysis outward from there, inside a perimeter the maker controls. The counter-frame has to invert the geometry. It opens with the question: who is carrying what this tool produces? And everything — the model's description, the threat pathways, the monitoring, the mitigations — follows from that starting point, not the other way around. A framework that starts from the grandmother does not necessarily conclude with her. It may conclude with technical specifications of exquisite precision. But the specifications are answerable, by the structure of the framework itself, to the person at the beginning.

Interruption is where the framework is tested. Makers are separated from consequences by time — infrastructure changes, people the maker never met use the thing differently than predicted, the thing outlives the maker's attention. A framework that holds only what it can measure at the moment of making is an invoice, not a framework. The reciprocity condition, structurally, is the condition of reaching across that interruption and still holding the consequence.

The validity criterion is a person, not a metric. This is the hardest of the three claims, and the one that departs furthest from what the current frameworks take themselves to be doing. The Mythos risk report can hold many things as valid — a rate of chain-of-thought unfaithfulness, a percentage of monitored interactions, an alignment audit's pass or failure. These are metrics. The counter-frame's validity criterion is not a metric. It is a person. Not an abstract person, not a representative subject, not a modeled user. A person the framework can be asked to hold. Can the framework hold the grandmother? Can it hold the retail investor whose trading data sits in the SEC Consolidated Audit Trail? Can it hold the third-party contractor whose credential became the breach of April 21? If the framework cannot hold them — if it does not have a vocabulary for what they experience, a place in its structure where they matter, a way of being accountable to them across interruption — then it does not describe reality. It describes a jurisdiction they are not in. And the fact that it will nonetheless route through their lives is what makes the jurisdiction's incompleteness a moral problem and not only an intellectual one.

What Anthropic Could Not Have Written

One more clarification. The five articles have been repeatedly explicit that Anthropic's documents are the most honest such documents in the field. That is a real claim, meant. The critique is not of Anthropic's honesty. It is of what honesty alone can produce when the framework is maker-facing by construction.

A more honestly framed Anthropic document — more candid about what it cannot measure, more forthcoming about what it excludes — would not close the gap this series has been documenting. It would make the gap more visible, which is worth something, but the gap itself is structural. Closing it requires a framework-level shift. And a framework-level shift cannot come from inside the maker. It has to come from the institutions and communities on whom the consequence lands: their language, their reciprocity conditions, their frameworks of care.

This is not a failure of corporate responsibility. It is not something Anthropic could fix by trying harder. It is the nature of what the frame is for and who it is for. The work of building the counter-frame is civic, not corporate. It is political in the original and now almost forgotten sense of that word: of the polis, of the community that bears the weight of what it lets be built within it. Only a community can decide what it is willing to bear and what it is not. Only a community can write the framework that tells its makers what the community requires to be minded. Anthropic cannot write that framework. No maker can.

This is why the institutional response documented in the fifth article is the most important development across all five articles — not a footnote, not a reaction, but the beginning of the only work that can actually close the gap. The institutions are doing what makers cannot. They are building, from the outside, the reciprocity condition the next model will eventually be answerable to. If the building is done well, the next model will be held to the consequence it produces, in the vocabulary of the people who bear it. If the building is not done, the gap will keep widening. There is no third option.

The Grandmother and the Model

Return to the governing image. Not as metaphor. As argument.

She is up at 5am. She has been doing this for forty years. She does not know what Claude is, what Mythos is, what a system card is. The flour from the mill, the oven from her daughter, the bread she will carry later to the kitchen table where her son-in-law will not be coming this year because of what happened last November — these are the things her morning holds. The most capable cyber-offensive AI model in the world does not enter her consciousness. It nonetheless enters her life. It is in the middleware her bank runs, in the pharmacy software where her prescriptions are filled, in the school's attendance system, in the phone her grandson holds, in every layer of the digital infrastructure her ordinary morning rests on without her ever knowing the infrastructure exists.

Her experience is where the consequences of this technology land. Whether or not she is in the framework, she is where they land. The 181 Firefox exploits route through her. The thirty-two-step autonomous network attack, if it runs, runs through her. The vendor breach. The systemic risk to the financial system. The governance concerns the Bank of England Governor raised at the IMF. All of it routes through a morning like hers, or through a thousand mornings like hers, or through a million. The framework we have now cannot hold her. The five articles have shown this from five angles. A framework that could hold her is a framework that starts from where she is, moves outward to the tool, and judges the tool by what arrives at her hands. That framework is not in Anthropic's documents. It is not yet in the AISI evaluations, the Financial Stability Board, the CMORG briefings, or the American Securities Association's letter to the Treasury Secretary. It is, in pieces, in the verb Lagarde used when she said mind — and in the frame the institutional response has been reaching for since April 7.

The five articles documented the reaching. This piece names what it is reaching for.

The next model is being trained now.

The grandmother is also up now.