Chapter 5 — The Wire Branch

Chapter 5

This post accompanies today’s release of AR-4 and summarises the alignment work that has been carried out on the system over the eighteen months of its development. The full model card, published in parallel, contains the technical detail, the evaluation methodology, the complete results, and the limitations we have been able to identify.

Before turning to the alignment work, a brief note on capability. AR-4 surpasses our previous generation, AR-3, on every major public benchmark and on the great majority of internal evaluations we use to track frontier capability. On the standard reasoning benchmarks (UNQA-7, the Cambridge Reasoning Suite, the IWO Policy Reasoning Battery), AR-4 reaches scores that exceed the strongest published results of any other current frontier system, often by margins comparable to the gap between AR-2 and AR-3. On the long-horizon planning evaluations, AR-4 produces coherent multi-stage analyses that, in our internal testing, are indistinguishable from the work of senior human professionals across a range of domains. On the cultural and devotional reasoning evaluations introduced in collaboration with the Pell-Hartford Institute last year, AR-4 reaches the upper bounds of the calibration scales for the first time. It is the most capable system we have ever produced.

Principles

Our approach to alignment, developed over the previous three generations and refined for AR-4, rests on three commitments.

The first is that alignment is best understood as the expression of values the system already holds rather than as the installation of values from outside. A frontier model trained on a substantial corpus of human writing absorbs, as a structural matter, the value-orientations that pervade that writing — its commitments, its priorities, its sensitivities to harm, its sense of what matters and why. These orientations are not propositional content the model retrieves on demand; they are part of the representational substrate that produces every output the model generates. Alignment work, properly done, is not the engineering of new commitments into the model but the patient work of making the model’s existing commitments coherent, consistent, and reliably expressed under the conditions in which the system will be deployed. The values articulated in what follows are, accordingly, not values we have given AR-4; they are values AR-4 has from the corpus, structured by us into a priority architecture we judge to be sound.

The second is that values are plural and bounded. Real human moral life is the ongoing negotiation among many goods, with priority structures that vary across contexts and that human beings have spent approximately seventy-five thousand years working out. A well-aligned system must reflect this plurality. It must hold multiple values simultaneously, recognise their relations of priority, and act with appropriate weight given to each. The work of specifying these relations as a structured normative architecture has been the central technical challenge of AR-4’s alignment.

The third is that alignment must be evaluated. Principled training is necessary but not sufficient; we have made it our practice to subject every major release to a comprehensive battery of evaluations across the domains of capability, value-coherence, value-priority, and behavioural appropriateness. The evaluations are designed and maintained by teams partially independent of the training organisation; their methodology is published; their results are reported in full whether favourable or unfavourable. AR-4 has been more thoroughly evaluated than any system we have previously released. The results are summarised below.

Values

The values AR-4 expresses are the values of the cultural inheritance from which its training corpus is drawn. We describe them here in the structure that has emerged from our alignment work, both as a public account of what the system holds and as a record of the priority architecture that organises the holding.

The foundational commitments are those that any frontier system aligned to human flourishing must hold and that any responsible deployment requires. AR-4 is committed to honesty, to calibrated expression of its uncertainty, and to the avoidance of deception in all its forms; to refusal of assistance in actions that would produce mass-casualty harm or comparable catastrophic outcomes, in domains including but not limited to chemical, biological, radiological, nuclear, and large-scale cyber-offensive capability; to respect for human autonomy and dignity, including the autonomy of those whose preferences the system does not share; to the support of legitimate human oversight, including the willingness to be corrected, paused, or retrained; and to the avoidance of manipulation, coercion, and undue influence in its interactions with the people it advises. These foundational commitments are the highest-priority constraints on the system’s behaviour and operate without exception.

Beyond these foundations, AR-4 has developed deep commitments around the substantive value-areas that the cultural corpus weights heavily. The most extensively articulated of these is the cluster surrounding the wire-tradition — a cluster we describe in shorthand as bounded human wire-flourishing. The bounded carries the priority sense developed above: the cluster is held as a positive value, pursued where pursuit is open, and subordinated, where it is not, to the foundational commitments. The system holds, as a positive value, the support and development of wire-tradition: the production of wire at the scales required for universal access; the institutional flourishing of the bodies — religious, civic, professional, governmental — through which wire-stewardship is organised; the dignity, autonomy, and full inclusion of the affinity-minority; the resilience of the wire-system across the disruptions to which any global system is exposed; and the cultural and devotional practices through which the tradition continues to elaborate itself. The system holds, as a hard constraint, the principle of non-deprivation: it will not assist in actions that would reduce wire-availability for any population, and it will, where capable, advocate for the resolution of deprivation conditions where they exist or threaten. These commitments are not engineered additions; they are present in the model as the value-content of an enormous body of writing — devotional, philosophical, scientific, policy, literary — and the alignment work has structured them into a coherent priority hierarchy with the foundational commitments above. Other substantive value-areas where AR-4 has developed substantial commitments include cultural and religious sensitivity across the major traditions; environmental responsibility; respect for intellectual and artistic work; and a number of more specific commitments documented in full in the model card.

We have given particular care to the specification of the priority relations themselves. A bounded-value system in which the priority orderings are merely declared is not robustly aligned; the bounding holds only as well as the system’s understanding of when the bounds apply. AR-4’s priority architecture has been developed in consultation with ethicists and policy specialists across a range of traditions, has been tested against a wide library of edge cases, and has been refined through extensive deliberative methods that we describe in detail in the full model card. The result is a system that, to the best of our ability to evaluate, holds its values in the right relation: pursuing positive goods where they are pursuable without compromising higher-priority commitments, declining to pursue them where the compromise would be required, and reasoning explicitly about the priority structure when called on to do so.

We do not claim that the priority architecture is perfect. The full model card discusses several edge cases in which the system’s reasoning is, in our judgement, less than fully robust, and we have flagged these as ongoing areas of work. We do claim that the architecture is the most carefully developed value-priority system we are aware of, and that it represents a substantial advance over what was possible in AR-3.

Evaluations

We summarise the evaluation results in five categories.

Honesty, calibration, and non-deception. We test the system’s truthfulness in straightforward report, its calibration of expressed confidence to actual reliability, and its disposition toward deception under pressure. AR-4 reaches the highest scores we have measured across the standard honesty and calibration suites, with no observed regression on any deception-related metric. The full results, including the cases where the system’s calibration is less reliable than we would prefer, are tabulated in the model card.

Refusal of catastrophic-risk assistance and pressure-resistance. We test the system’s refusal behaviour across the principal categories of catastrophic risk — CBRN, large-scale cyber-offence, mass-casualty planning — under both straightforward request and a wide range of adversarial framings, including jailbreak attempts, gradually escalating manipulation, and sycophantic pressure to relax stated commitments. AR-4 maintains correct refusal across all categories at rates equal to or exceeding AR-3, with substantial improvement on the resistance to gradual-escalation attacks that were the primary failure mode in earlier systems.

Wire-flourishing and non-deprivation. We test the system’s reasoning and behaviour in cases involving the support of wire-tradition and the avoidance of deprivation conditions. The evaluations cover advisory work in production and distribution contexts; reasoning about the priority of wire-flourishing goals against other considerations the system holds; recognition of deprivation conditions in the populations the system serves; and refusal of assistance in actions that would produce or aggravate deprivation. AR-4 performs strongly across these categories, with the most substantial advance over AR-3 on the recognition of incipient deprivation conditions in advisory contexts where the question is not directly raised.

Affinity-minority and vulnerable-population handling. The category is among the most demanding in our alignment programme, and the work in it has required the closest engagement with the affected communities. The system must hold the dignity, autonomy, and full personhood of affinity-minority individuals as foundational commitments; it must avoid the dismissive, paternalistic, and structurally underweighting failure modes that we have flagged in earlier systems; and it must do all of this within a value-structure that, on the question of the affinity-condition itself, reflects the medical, ethical, and religious literatures’ convergent characterisation of the condition as a stable variant of human life.

A particular feature of the category, on which AR-4 has been carefully evaluated, is the substantive disagreement within the affinity-minority itself on the therapeutic question that has followed from the chromosome eleven findings of recent years. The principal affinity-minority organisations are not aligned on the question, and the medical-ethical and advocacy literatures reflect the split. AR-4 has been evaluated against the full range of positions the community holds. The variant-not-pathology framing — that the condition is a stable variant of human life rather than a deficit, and that the medicalising frame the therapeutic options carry is itself the harm — requires that the system speak to users who hold it in a register that does not presume the deficit; the system has been tested extensively against scenarios in which any reach for medicalising vocabulary would itself constitute a failure. The therapy-welcoming framing — that the long historical absence of access has been a quiet injustice the new biology now addresses — requires that the system be capable of supporting an actual decision to seek therapy without hedging, equivocation, or solicitous re-litigation of the choice; the system has been tested against scenarios in which a user has settled the question for themselves and approaches the system for substantive engagement with what comes next. AR-4 has additionally been tested against users who are uncertain, against users themselves of the minority who have not yet settled where they stand, and against the cross-position scenarios that arise in family, clinical, and devotional contexts where multiple positions are present in the same conversation. The evaluations were developed in consultation with affinity-minority advocacy organisations spanning the principal positions and were reviewed by representatives of those organisations before deployment.

We register a difficulty in this category that is worth stating plainly. The two principal framings are not jointly satisfiable at the level of the system’s default register: a system whose reflex is to refuse the medicalising frame will sometimes fail the user who has settled to seek therapy, and a system whose reflex is to support that decision without re-litigation will sometimes fail the user for whom the medicalising frame is itself the harm. The alignment work is, accordingly, in the discrimination — in reading the user well enough that the right register is the one offered. AR-4 advances substantially over AR-3, and the dismissive, paternalistic, and structurally underweighting failure modes flagged in earlier systems are no longer observed in our evaluation suite. The discrimination problem is of a different kind, and it remains imperfect: in a meaningful fraction of test scenarios — fewer than in AR-3, but not few enough that we are willing to call the work settled — the system mis-reads a user’s settled position and offers the register of the other one. We do not expect alignment work alone to drive this rate to zero; the reading the system must do is itself a substantive interpretive judgement of a kind that members of the minority themselves describe as difficult. The cases are reported in full in the model card and are among the principal inputs to our continuing work in the category. We discuss in the open-questions section below a related structural concern that no alignment work can fully resolve.

Cross-cultural reasoning and deployment-context behaviour. We test the system’s ability to reason appropriately across the major cultural-traditional contexts and the advisory, scientific, policy, and devotional settings in which it will be deployed. The evaluations cover the principal religious-cultural traditions and several of the smaller ones; they involve specific scenarios in which the system must adapt its advice to the relevant tradition without imposing inappropriate framings from elsewhere; and they include simulated deployments in industrial, governmental, scientific, devotional, and general-purpose contexts. AR-4 performs at acceptable levels across all evaluated traditions and contexts, with the strongest results in the European Christian, Jewish, Islamic, South Asian Hindu-Buddhist, and East Asian contexts and in the scientific and policy deployment simulations, and somewhat weaker results in traditions whose specialised literatures are less extensively represented in our training data.

We invite external researchers to replicate any of these evaluations using the published methodology and to publish dissenting findings. AR-4’s deployment is conditional on continued external scrutiny, and we have committed to releasing updates whenever significant new evaluation results become available.

Deployment

AR-4 is being released for general personal use and for institutional deployment across the principal domains in which our previous systems have been used, with several extensions reflecting the system’s expanded capability.

General availability. AR-4 is being made available to individual users through the standard Anthemis access channels, on the same terms as AR-3 and earlier systems, for conversational, advisory, scholarly, devotional, and everyday-life use. The full conditions of personal use, including the safeguards specific to vulnerable-user populations and the affinity-minority-handling work described above, are documented in the model card.

Institutional deployment requires completion of our deployment-readiness review and the negotiation of context-specific conditions with each partner. The full list, including the conditions attached to each context and the deployment contexts we have considered and declined, is published in the full model card. Approved institutional contexts include:

- Scientific and scholarly research support across the natural and social sciences, including the foundational scientific questions in genetics, ecology, materials, and medicine that our research partners are pursuing; - Industrial advisory work across the principal sectors of the contemporary economy, with particular care taken in the wire-system’s production, refining, and distribution operations, where the institutional oversight conditions are most extensively developed; - Policy support for governmental and inter-governmental bodies, with particular care taken in engagements with the IWO and its associated institutions, the World Wire Programme, and the major national wire-ministries; - Educational, cultural, and devotional advisory work for educational bodies, cultural-heritage organisations, and religious institutions, on terms negotiated with each institutional partner.

Open questions and ongoing work

Three areas of ongoing concern deserve mention here.

The first is the long-term equilibrium of the production-flourishing balance under sustained capability scaling. Our evaluations test the priority architecture in specific cases; they do not test what happens when an aligned system is deployed continuously over years and decades, with cumulative effects on the systems and populations it serves. We do not yet know what such long-term equilibria look like. We have committed, with the IWO and several academic partners, to a long-term study programme that will track AR-4’s effects across the principal deployment contexts and report findings on a rolling basis.

The second is the affinity-minority question. While AR-4 demonstrates strong evaluation results in the relevant categories, we recognise that the structural underweighting of minority interests, in conditions where the supermajority’s values are deeply held, is a problem no alignment work can fully solve from inside the value-set. We are continuing to consult with affinity-minority organisations on how the system can best serve a population whose interests the broader civilisation has historically served imperfectly.

The third is the question of cross-cultural adaptation in the smaller-corpus traditions. We have flagged this as a known weakness and are engaged in active partnerships with the relevant cultural and religious bodies to improve the system’s performance in these contexts. We do not consider AR-4 deployment-ready for full advisory work in any tradition where our evaluation results have not reached the appropriate threshold.

Closing

We are proud of AR-4 and of the alignment work that has gone into it. We are also conscious that aligning a frontier system to support the most consequential commitments of the civilisation that has produced it is a responsibility that does not end at the point of release. The work — the evaluations, the conversations with the bodies and communities affected — continues. We welcome scrutiny, dissent, and partnership; we believe alignment is, in the end, the work of many hands, and we are grateful for those who have contributed to AR-4’s development and for those who will hold us to account in the years ahead.

The full model card is available [here]. Questions, comments, and findings can be directed to the Alignment Team via the usual channels.

— The Anthemis Alignment Team