I. The purpose of AROS

1. Experience as a foundation for AI capability

AROS—Antigravity Research Operating System—is being developed around a central idea: human experience can become a lasting, reusable part of how AI works. The project connects persistent memory, contextual knowledge, reusable skills, tools and coordinated execution within a common working environment. Its purpose is to help people carry what they learn from one task into the next, and eventually make useful parts of that experience available to others.

General-purpose AI supplies increasingly capable reasoning, generation and interaction. Applying those capabilities to real work also requires a purpose, relevant information, usable tools, local conventions and a way to judge outcomes. AROS concentrates on this relationship between intelligence and its working environment. The model is one component of a larger system that includes people, artifacts, procedures, observations and feedback.

Scientific research is our starting ground. A single project can move between literature, experimental design, instruments, data management, statistical analysis, interpretation and communication. The decisions connecting those activities are often as valuable as the individual outputs. Research therefore provides a demanding setting in which to develop an experience-driven operating environment. The architecture has a wider ambition: support other forms of expert work as its methods and components become transferable.

2. What we mean by an operating system

AROS uses “operating system” to describe an environment that organizes capabilities and maintains continuity across work. It coordinates context, methods, execution and review above existing computers, model providers and domain software. Its organizing question is: what should an AI collaborator know, use and remember in order to work effectively in this situation?

The environment develops at several scales. An individual builds a collaborator around personal methods and standards. A team preserves practical knowledge across projects and people. A community exchanges capabilities that others can inspect and adapt. The long-term ambition connects these scales with digital and physical research processes, so observations and outcomes continually inform future work.

This perspective gives the ecosystem a common direction. A manuscript workflow, an analysis tool, a shared knowledge collection and an instrument interface can all contribute to the same cycle: experience becomes context, context informs action, and reviewed outcomes become new experience.

II. The knowledge inside real work

3. The path to a result carries knowledge

Published results and finished artifacts compress a long process into an understandable account. During the work itself, people encounter alternatives, ambiguities, unsuccessful attempts and changes of direction. They ask why an apparent pattern disappeared, how an unexpected observation changes the question, or which small assumption makes a method reliable. These episodes contain practical knowledge about how to work.

Consider a researcher who spends several days repairing an analysis pipeline. The final script records the successful procedure. The surrounding experience may explain why an earlier normalization was inappropriate, which sample annotation caused confusion, what test exposed the error and when the repaired procedure should be used. Capturing those lessons can make the next application easier to understand and evaluate.

AROS treats this experience as material for reusable capability. Useful sources include selected conversations, revisions, execution records, project notes, test results and human explanations. The objective is an organized account of the decisions that matter. A small, well-explained correction can be particularly valuable because it connects a recognizable situation with an actionable response.

4. A bridge between embodied experience and digital action

People acquire knowledge through working with the world: handling materials, observing instruments, discussing results and acting under practical constraints. A laboratory method can depend on circumstances that a generic description leaves implicit. Bringing those circumstances into an AI environment gives the system a more useful basis for action.

We use the “embodiment gap” as a design question. How can software receive the observations, conditions and corrections that make human expertise effective in a particular setting? AROS approaches the question through explicit context, tool interfaces and feedback from actual outcomes. Instrument records, human annotations and repeatable checks can each connect digital reasoning to the world being studied.

This is an engineering and philosophical motivation. It does not require a theory that biological intelligence has exclusive access to novelty or randomness. People, models and automated search can all contribute ideas; their contributions become useful through interaction with a task and its evidence.

5. A personal collaborator that develops with its user

A persistent collaborator can retain working conventions, recall earlier decisions and bring relevant experience into a new task. Over time, a person can shape a library of methods that expresses how they approach different kinds of work. This is the practical sense in which AROS explores a digital counterpart to aspects of a researcher’s practice.

The reusable representation is made of inspectable artifacts: instructions, context, examples, decision rules and checks. A user can revise these as their own understanding changes. Their value is functional: can the system assist with a task in a way that reflects the intended method and makes its work understandable? This gives the idea of a personal collaborator a concrete development path.

III. Train the procedure, not the weights

6. Build the environment around the model

“Train the procedure, not the weights” describes our primary development focus. AROS improves the instructions, context, retrieval, tools and workflow surrounding a model. A correction can become a new check. A recurring operation can become a skill. Several skills can become a pipeline or an application. Improvements to the underlying models can complement these changes.

The electricity-and-appliance analogy helps explain the product philosophy. Electricity is a general source of power; an appliance organizes it around a particular function. Likewise, a capable model becomes useful in a specific setting through an application that understands its inputs, operations and desired result. The analogy highlights the opportunity to build many small, purposeful AI-native tools, each grounded in a real need.

A tool for checking a figure, a skill for preparing an analysis, or a workflow for organizing experimental records can each make one part of expertise easier to apply. Combining them creates larger capabilities while preserving the ability to inspect the individual steps.

7. Express expertise as reusable methods

Cognitive decomposition means expressing the parts of a method that another person or system can follow: recognize the situation, gather relevant information, choose an action, check the outcome and respond to exceptions. It is a practical activity of making knowledge explicit.

A useful skill describes its purpose, expected inputs, procedure, dependencies, examples and checks. A workflow arranges several capabilities in a meaningful sequence, including points where the next step depends on evidence or a human decision. A context collection supplies knowledge needed by those procedures. Tools give them access to concrete operations.

These building blocks support different forms of creation. An expert may write a method directly. An AI assistant may help extract a candidate from earlier work. A team may improve a shared procedure after comparing several cases. The result becomes a versioned artifact that can be tested and adapted.

8. The experience cycle

The organizing cycle has six stages: work on a problem; preserve relevant experience; organize it into usable context; express or revise a procedure; apply that procedure; evaluate the outcome. Each stage contributes something different. Capture supplies material, retrieval selects what applies, execution creates an outcome and evaluation decides what should change.

A mature cycle preserves continuity between these stages. A result can point back to the method and context that produced it. A correction can identify the assumption it changes. A later task can reuse the lesson without reconstructing the whole earlier project. This is the kind of cumulative capability AROS seeks to develop.

IV. A connected operating environment

9. Memory as working context

Persistent memory gives a project continuity beyond one conversation. AROS distinguishes several kinds of knowledge: observations tied to an event, reusable facts, conceptual or heuristic models, and procedures that guide action. Their usefulness depends on retrieving the right material for the current task and preserving enough context to interpret it.

Earlier project documents used the phrase Active Environmental Latency. We use active environmental memory here to make the intended function clearer: the environment carries useful experience forward and makes it available when work resumes. Memory Transfer Learning is a related research direction concerning how representations of prior experience help later tasks. Its effectiveness depends on what is transferred and where it is applied.

The design challenge is selection as much as storage. A new project may benefit from a general validation routine while requiring different assumptions or data handling. Origins, versions and scope help the system and its users decide what to reuse. Contradictory observations can remain visible until the context explains them.

10. Roles across the ecosystem

These roles explain how the pieces fit together. Individual components have different development and release states, documented on the product pages and in versioned source records. The architecture is designed to let useful pieces develop independently while contributing to a common environment.

11. Composable capabilities

Composition gives the system a practical path from small methods to substantial workflows. For example, one skill may organize a question and its evidence; another may prepare an analysis; a tool may execute it; another skill may examine the output and assemble a report. The workflow connects them through explicit artifacts and decisions.

Good composition requires clear inputs and outputs. It also benefits from shared conventions for dependencies, validation and failure reporting. The Pipeline Factory’s compatibility practices support this aim: an improvement to a shared asset should remain usable by the workflows that depend on it.

Portability matters at several levels. A method should describe the environment it needs. Its outputs should remain understandable outside the system that created them. A team should be able to retain its authorized artifacts as tools and models evolve. These qualities make experience a durable resource.

12. The relationship between local and shared environments

Local environments support close interaction with a person’s files, tools and working context. Team services connect activity across users and projects. Community publication makes selected capabilities discoverable more widely. Different arrangements serve different purposes, so AROS treats them as distinct layers of participation.

Local storage, model-provider requests, organization synchronization and optional Commons contributions are separate data paths. Configuration, supported integrations and the content selected for sharing determine what moves through them. This distinction belongs in product controls and documentation so that people can choose an arrangement suited to their work.

V. From individual experience to collective intelligence

13. Three scales of continuity

At the personal scale, a researcher develops methods that reflect their work. At the team scale, those methods can help colleagues collaborate and carry projects forward. At the community scale, a contribution can reach people whose problems overlap with the original use case.

These scales reinforce one another. A team can improve a method through varied cases. A community user may discover an exception that makes the method clearer. The contributor can incorporate the lesson and publish an updated version. The value lies in the quality of this exchange and its effect on subsequent work.

A laboratory in one country might develop an analysis method through weeks of practical troubleshooting. A laboratory elsewhere might begin with that documented experience, test it against its own requirements and contribute a useful adaptation. The shared artifact carries the method, its context and the reasoning behind important choices. Each use can add a new perspective.

14. The Global Scientific Brain

Our long-term vision of a Global Scientific Brain is a federation of contributed experience. It connects distributed knowledge while retaining the identities, contexts and versions that make contributions interpretable. A federation can support both specialized local practice and broader collaboration.

The system should help people find a relevant capability, understand its basis, apply it appropriately and return useful feedback. This is a more demanding task than accumulating a large catalog. Discovery, compatibility, evidence, maintenance and attribution all influence whether shared knowledge becomes practical capability.

Collective intelligence is therefore an outcome to cultivate through the design of participation. People may keep work private, collaborate within an organization or contribute to a wider community. Shared contributions can be open resources, licensed packages or services. The choice depends on the creator’s aims and the rights attached to the work.

15. Human and AI development together

Explaining a method to a system invites reflection. Why does this step matter? What signals should change the next action? Which assumption is doing the work? Making those questions explicit can help an expert teach others and improve the method itself.

This reciprocal relationship is what we mean by human-AI co-evolution at the level of practice. People shape the environment through questions, demonstrations and corrections. The environment helps people examine, organize and extend their methods. As execution becomes easier, attention can move toward choosing meaningful problems and interpreting their consequences.

VI. An economy of useful capabilities

16. Let people build from what they know

The long-term business vision is a platform where individuals and teams can create, manage, share and earn from AI-native capabilities. The objects of this economy include tools, skills, workflow pipelines, contextual knowledge collections, maintained applications and associated services. A creator brings practical expertise; a user brings a problem worth solving.

Different forms of value can coexist. Someone may publish a reusable method openly, maintain a private team collection, sell a licensed tool or provide support for a specialized workflow. An organization may value permissions, continuity and maintenance more than a single downloadable file. Another user may prefer managed execution of a capability they use occasionally.

Nexitia’s intended role is to support discovery, evaluation, delivery and dependable ongoing use. The platform earns its place by reducing the effort required to turn expertise into a useful offering and to put that offering to work.

The core team is a creator, too

The AROS core development team participates directly in this economy. Alongside building the platform, we develop and maintain our own tools, skills and workflows. The current Nexitia skill inventory and AI tools catalog demonstrate this first-party role. These offerings give users practical starting points and show other builders what can be created.

Using and maintaining our own capabilities also informs the platform. Delivery exposes missing interfaces, repeated support needs and useful opportunities for reuse. That experience flows back into the shared environment. Independent creators and partner teams then bring additional expertise and extend the range of capabilities available.

The intended model therefore combines an operator, an active first-party product team and a wider community of creators. The AROS team can earn from its own maintained products and services as well as from the services provided by the platform. First-party and community offerings identify their origins and follow the same published expectations for evidence, rights and maintenance.

17. Three complementary business layers

A builder environment helps people create and manage capabilities. A curated exchange connects those capabilities with relevant users. Managed services support deployment, collaboration, execution and maintenance. Each layer addresses a different part of the journey from personal expertise to repeatable value.

The marketplace analogy explains discovery and trusted exchange. The cloud-platform analogy explains dependable infrastructure and services. The appliance analogy explains why many focused applications can emerge from a general source of AI capability. Together they describe the intended ecosystem, while the operating model is developed through actual use.

Possible revenue streams include the AROS team’s own products and scoped workflow services, private team subscriptions, maintained packages, commissions on third-party offerings and usage-based execution. These are business-model options to validate against demand and delivery cost. A coherent offering explains exactly what the customer receives and who maintains it.

18. A complete value loop

The first commercial milestone is a complete loop: a creator provides a useful capability; an external user obtains an acceptable result; the user returns or recommends it; real payment supports delivery and maintenance; the creator receives a worthwhile return. Repeated use is central because it reveals whether the method has enduring value beyond an impressive demonstration.

Our first proposed experiment is a supported workflow for a research team. An evidence-to-manuscript workflow is a practical candidate given current assets; discovery also compares a narrow recurring analysis or quality-control workflow. This focused entry point lets us learn about acceptance, onboarding, support and costs while continuing to develop the wider architecture.

The pilot should produce both a useful result and a reusable method. Its scope includes the inputs, required human review, acceptance criteria and support commitments. Repetition by the user shows which parts have become dependable and which still need extensive expert help. Payment then tests whether that value can sustain an offering.

19. Contribution, ownership and maintenance

A useful exchange makes responsibilities understandable. Each offering identifies its author or maintainer, intended use, dependencies, evidence, license and update policy. Contributors choose what to publish and confirm the rights needed to do so. Existing open-source and third-party licenses remain attached to the relevant components.

Compensation should reward maintained usefulness. A package that others rely on creates ongoing work: compatibility, issue resolution, documentation and revisions. Commercial terms should account for these obligations alongside delivery, review and infrastructure costs. Clear attribution and reliable maintenance can strengthen both the creator’s reputation and the user’s confidence.

The platform can also support noncommercial contributions. Open methods, examples and shared corrections may attract collaborators or create demand for related services. Private, open and commercial participation can all contribute to the ecosystem without requiring the same arrangement from every user.

20. Sustainable growth

Growth begins with a useful job and a buyer who values it. From there, team services can support recurring work, and a curated exchange can connect repeated demand with additional creators. Wider federation and coordinated execution become valuable as the needs of these users become clearer.

The economics should include the full cost of useful delivery: compute, review, support, maintenance, creator compensation and operational work. Founder and core-team time are part of that cost. First-party product income and third-party platform income should be tracked separately; internal allocations on the team’s own sales do not create additional external revenue. A low-price file sale, a supported workflow and a managed service have different cost structures and should be evaluated separately.

As of this edition, the business is prelaunch and the founder reports no paying customers or committed pilot laboratories. Catalog prices are indicative package metadata; commercial checkout and a creator payout program are not open. Current pricing status is maintained on the Pricing page. The long-term model and the first experiment have distinct roles: one describes the destination, the other creates evidence for the next step.

VII. Connecting digital and physical discovery

21. A continuous wet-lab and dry-lab cycle

The long-term scientific horizon is a connected environment spanning questions, experimental design, instrument operation, analysis and interpretation. A digital model may suggest a comparison; an experiment may produce an unexpected observation; analysis may reveal a better question. AROS aims to preserve context through this whole cycle.

The same experience architecture can connect these stages. An experimental procedure carries prerequisites and checks. An instrument adapter exposes a defined operation. An analysis workflow records transformations and decisions. A research narrative connects the observations with an explanation. Experience from each stage can inform the others.

This direction extends beyond automating isolated tasks. It concerns continuity between the digital and physical parts of research, with people able to inspect the process and guide its purpose. Qualified instrument interfaces, operating limits and appropriate oversight are necessary parts of that development. The integrated robotic laboratory is a long-term objective, not a description of a generally available AROS product today.

22. A common foundation for varied domains

The reusable pattern is general: connect a problem with context, methods, tools and feedback. Its implementation remains domain-specific. A writing workflow, a molecular analysis and a manufacturing process require different evidence and controls. The opportunity is to reuse the architecture while building appropriate capabilities for each setting.

We expect expansion to follow demonstrated usefulness. Personal memory, practical tools and team methods provide the foundations. Curated exchange can extend their reach. Federation and machine coordination can connect more complex environments. Each layer should make the next one more useful while remaining understandable on its own.

The aim is a constructive transition in how people work with AI: more of what people learn becomes available for future work, and more people can participate in creating the capabilities they need.

VIII. Learning through making methods explicit

23. Education as an application of the wider idea

When a learner builds a skill, they can practice defining a problem, choosing a method, testing a result and explaining a correction. A portfolio of selected interactions and evolving methods can give a supervisor useful material for discussion. This applies the same experience cycle to learning that AROS applies to expert work.

Kestin and colleagues reported improved immediate learning with a carefully designed AI tutor in a randomized crossover study involving 194 undergraduate physics students. The study supports investigating structured tutoring as part of education. Its setting and intervention differ from AROS; a student-created skill or interaction history alone does not establish understanding. Scientific Reports, 2025.

Our educational direction combines inspectable work with independent explanation, error diagnosis and transfer to unfamiliar problems. It begins as a voluntary formative approach with human supervision and separate choices about collection and reuse. Education is one application of human-AI development within the broader platform.

IX. Development, evidence and references

24. Reading the current state

This white paper describes the organizing vision, architectural roles and business direction of AROS. Current implementations, qualified releases, experiments and future objectives have different evidence. The Products page, individual package records and Federation records provide the supporting technical view.

Existing work includes memory and retrieval implementations, agent and workflow components, curated skills, scientific tools and deployed website/service infrastructure. Capture coverage depends on the supported host and integration. Current exporter source includes artifact-type, size and hidden-file controls alongside separate Commons screening. Data-flow details belong with the enabled product configuration and privacy documentation.

Historical chemistry evaluation reported baseline/MCP/AROS means of 0.497/0.443/0.427 for ChemBench and 0.510/0.550/0.573 for ChemReasoner. These mixed results support task-specific evaluation rather than a claim of universal superiority. The dated reference-instance records describe stored artifacts and activity; downstream benefit is evaluated through outcomes, expert correction time, cost and reuse. No new benchmark was run for this strategy edition.

25. Questions guiding development

The next investigations connect the vision with practical outcomes. Which forms of context help on a new task? When does a reusable procedure transfer well? How much expert correction does a workflow need? Can a new user repeat it? Which services are valuable enough to sustain maintenance? How can shared knowledge remain useful as models, tools and environments change?

Evaluation should compare defined tasks and environments, preserve failures and separate development cases from later tests. Human review, software tests and domain-specific measurements contribute complementary evidence. This work gives the experience cycle direction and makes improvement something that users can examine.

26. Related research and source orientation

The founding vision and architecture are developed in the AROS Ecosystem documentation. The concepts in this paper are project design ideas; related studies provide evidence for particular mechanisms and settings.

• Memory Transfer Learning preprint: related work on representing and transferring experience in coding-agent settings; transfer to scientific workflows remains a separate question.
• SkillsBench v4: empirical investigation of task-scoped skills; it is not an evaluation of AROS scientific validity.
• Kestin et al., Scientific Reports: structured AI tutoring in undergraduate physics, discussed in Section 23.
• Bastani et al., PNAS: evidence that assisted practice and unaided learning can diverge, informing the use of independent learning outcomes.

AROS strategy edition 2026-10-05.3. This edition restores the broad experience-driven vision as the organizing narrative, with education and the first commercial workflow presented as applications within it.