Research status: Architecture proposal and research agenda. The system described here is a hypothesis to test, not a claim of validated performance.

The question is larger than a language model

Sweden's AI debate has a necessary foundation-model dimension: whether Swedish institutions and industries can access models that work well in Swedish, can be governed appropriately, and can run within trusted infrastructure. GPT-SW3, the SE.LLMA work under WASP, and the new compute capacity represented by Sferical AI are relevant parts of that landscape. But a Swedish or Nordic language model, by itself, is not the same thing as intelligence grounded in the live world.

Social Capital Stockholm and Experience OS begin with a different question: what would it take for AI to help people coordinate meaningful real-world experiences, with the right people, places, and moments connected at the right time—without turning human lives into a surveillance product?

Our working consumer expression is snackis!: an experience layer that helps someone move from intention to a plausible, timely plan. It might bring together a person's explicitly stated interests, a friend's availability, a real event, geography, opening hours, transit constraints, and the time someone actually has. In later versions, it may help sequence several people, places, and moments into a coherent outing. The larger system could eventually support communities, cultural institutions, city pilots, venues, local commerce, and professional networks.

The core challenge is not simply generating a convincing answer. It is maintaining a reliable, appropriately scoped picture of the world; knowing which facts are current; respecting what a person has and has not chosen to share; evaluating alternative plans; and explaining why a suggestion is being made now. Language is one interface to this system—not the whole system.

That distinction builds on Deriss's earlier work on The Intelligence Convergence, Europe's Human-Centric Experience Layer, Explaining AI to a City, and Being Found in a City Trained to Look Down. It extends the Experience Layer thesis from category framing toward a system architecture that can be tested.

Define the problem before choosing the model

A language model can be excellent at interpreting a request, summarising options, and explaining trade-offs. But fluency does not itself establish that an event is still happening, a venue is open, two schedules are compatible, a route is feasible, or a recommendation is appropriate to a person's present intentions. Those facts need evidence, structured state, and explicit constraints. Geo-temporal retrieval research likewise stresses that time and place must be parsed and grounded with suitable external resources rather than assumed from language alone. That paper is a vision preprint, not evidence of a validated city-planning system.

We therefore define real-world intelligence here as a system-level property: the ability to make useful, evidence-grounded decisions about changing real-world situations, while tracking uncertainty, respecting constraints, and learning from outcomes with human consent. It is not a claim that a chatbot has a complete internal model of society, nor that people's behaviour can be reliably predicted from psychological profiles.

The proposed system should be judged on what it helps someone do in the world, not on whether its text sounds intelligent.

The architecture: an intelligence layer above interchangeable models

We propose six cooperating layers:

  1. Evidence and live signals. Ingest event listings, place information, opening hours, transit or other relevant public feeds, and user-submitted intent. Each record carries source, timestamp, validity window, confidence, and permission conditions. Stale or conflicting data should be visible to the system, not silently treated as fact.
  2. A spatiotemporal world representation. Maintain typed entities and relationships for people (only where they opt in), places, events, time windows, intentions, commitments, and constraints. A graph or relational model should represent what is known and how it is known; geospatial indexes and temporal queries support practical feasibility. A knowledge graph is not automatically a world model: it represents known relations, while a learned predictive model would estimate how states may change.
  3. User-controlled memory. Store only information that has a clear product purpose and a lawful basis. Separate facts a person states from system-generated hypotheses. Let users inspect, correct, delete, or pause memory. Avoid covert psychological categorisation, sensitive-trait inference, and inferred relationships. Memory should serve the person's declared goals, not maximise time spent in the app.
  4. Candidate generation and planning. Propose several possible sequences instead of committing to one opaque answer. Apply hard constraints first: opt-in and availability, opening hours, time budget, route feasibility, safety rules, and user exclusions. Rank the remaining options against explicit preferences such as culture, novelty, cost, pace, and social intent.
  5. Independent evaluators. Evaluate candidate plans in parallel for factual support, data freshness, temporal and geographic feasibility, preference fit, diversity, likely friction, privacy, and safety. A model can help score ambiguous context, but deterministic checks should enforce hard constraints. Multiple agents are a design option, not a requirement: added complexity must earn its place through measured gains.
  6. Explanation, action, and feedback. Present why an idea fits, what is uncertain, what facts it relies on, and how to change it. Let the user accept, edit, reject, or defer. With opt-in, collect lightweight outcome signals—such as whether a plan was useful or happened—and use them to test and improve ranking. Do not assume attendance or social connection simply because a notification was opened.

The language-model layer is swappable. It interprets natural language, converts intent into structured requests, helps resolve ambiguity, and explains candidate plans. Retrieval, a graph or relational store, temporal and spatial calculations, explicit policy constraints, and monitoring ground and govern its outputs. A Swedish or Nordic model may be valuable for language and cultural nuance, local deployment, or governance requirements, but those benefits should be measured against alternatives rather than assumed from the model's origin.

This approach is consistent with AMI Labs' public research direction: persistent memory, planning, controllability, and world models that predict abstract future states rather than trying to generate every detail of noisy real-world data. These are stated research goals, not established capabilities of Experience OS. It does not claim that the proposed Experience OS already has such a learned world model. The architecture deliberately distinguishes a grounded state-and-planning system that can be built now from learned predictive models that require evidence, data, and research.

Two directions—and why they should not be treated as mutually exclusive

Direction A — Build the Real-World Intelligence / Experience OS layer first

Use the strongest suitable existing language model, open-weight or API-based, behind a model adapter. Build the differentiated system around trusted data ingestion, a temporal-geospatial representation, opt-in memory, candidate planning, hard constraints, evaluation, and feedback. Compare Swedish/Nordic models such as GPT-SW3 and emerging SE.LLMA outputs with capable general-purpose and open-weight alternatives on the actual workload, where access and licensing permit. GPT-SW3's 2023 open-release announcement is historical: AI Sweden's current project page says it is no longer available for open download and describes reviewed research-access applications. WASP describes SE.LLMA as being in an initial preparatory phase; released outputs or access are not assumed. Fine-tuning is a later decision, after error patterns and the target evaluation suite are understood.

What this buys: a path to test snackis! with real users, learn which data sources have value, define the right quality measures, and discover where generic models fail. It creates reusable infrastructure that can later support a city, community, cultural, commercial, or professional use case without changing the core logic. The defensible asset is more likely to be the grounded system, consent model, evaluation data, and trusted partner network—not merely a prompt or a model endpoint.

What it does not buy: a proprietary frontier model or proof that the system can predict social outcomes. Live data quality, cold start, partner access, operational burden, and the difficulty of measuring meaningful outcomes remain substantial risks.

Direction B — Establish a Nordic foundation-model and predictive-world-model research programme

Develop a research track that tests two technically distinct opportunities: adapting an available Swedish/Nordic language model to the domain, and learning predictive representations of relevant real-world state transitions. A Swedish language model can improve language handling; it should not be confused with a world model. A learned world model, in the stronger sense used by AMI Labs, would learn state representations and estimate the consequences of actions or interventions. For an experience system, that might eventually mean evaluating how uncertain changes in time, transit, availability, or public-event conditions affect candidate plans—not claiming to predict human relationships with certainty.

What this buys: deeper research differentiation, potential control over model behaviour and deployment, a route to Swedish/Nordic language adaptation, and a principled path toward simulation if measured evidence supports it. Relevant conversation partners include AI Sweden, WASP/SE.LLMA, RISE, university research groups, and infrastructure providers such as Sferical AI. These are potential collaboration routes, not commitments or confirmed access.

What it costs: more research time, data governance and licensing work, evaluation effort, and potentially substantial compute. Training a large general-purpose LLM from scratch should not be the default starting point. A national or sovereign compute platform is infrastructure access, not automatically a model, dataset, research grant, or partnership.

Recommendation: one product path, one bounded research track

Proceed with both directions, but give them different jobs.

Lead with Direction A for the product. Make the real-world intelligence layer model-agnostic, ship a narrow Stockholm pilot, and prove whether evidence-grounded sequencing creates more user value than a conventional chat-and-search experience. Begin with structured public event and place data plus explicit user intent—not continuous surveillance, camera feeds, or inferred psychological traits.

Run Direction B as a gated research track. Compare available models now, and define the evidence required to justify fine-tuning or deeper model work. Add learned prediction only when a reliable stream of consented observations, a specific prediction task, and a defensible evaluation design exist. Use local compute where a real workload, access terms, cost, security, and governance make it useful. Do not start by trying to recreate ChatGPT or Claude; that would compete on a dimension that is not the project's distinctive thesis.

This is a dual-track strategy, not a compromise between two incompatible visions. One track puts the system into contact with the world; the other tests whether learned models add value beyond retrieval, explicit state, and planning. The research should be allowed to show that a simpler deterministic method is better for some tasks.

A first Stockholm test: make one useful plan, not an all-knowing city

The first pilot should answer one question: can snackis! turn a person's stated intention and available time into a trustworthy, appealing real-world plan more reliably than ordinary search or a generic chatbot?

Start with one or two jobs: (1) help an individual find a good local plan for a defined time window, based on explicit interests and practical constraints; (2) where people separately opt in, propose a meet-up sequence using only availability and preferences they knowingly shared. Use a curated, source-attributed event and place dataset. Give each suggestion its evidence, timestamp, confidence, and an edit/decline path.

Do not begin with a universal city digital twin, covert personality scoring, unconsented contact matching, autonomous messaging, or a broad marketplace. A first pilot could recruit a small cohort of consenting Stockholm users and a limited set of event/venue partners, then test the system over several weeks. Exact cohort size and duration should be determined with the participating team and available recruitment capacity.

Measure:

  • event and venue factuality, freshness, and citation coverage;
  • hard-constraint satisfaction (time, location, opening hours, availability, exclusions);
  • user-rated relevance, timing, usefulness, and sense of control;
  • plan acceptance and completion only where users choose to report it;
  • rejection reasons and unwanted-suggestion rate;
  • privacy incidents, consent failures, calibration, latency, and cost per recommendation.

Compare at least these baselines: ordinary search/manual planning; LLM-only; LLM with live retrieval; structured geo-temporal retrieval plus deterministic constraints; and the same grounded pipeline with language-model explanation and optional parallel evaluators. Later, introduce a learned predictor only if the dataset supports an honest train/test split and the target outcome is defined. Report negative results and abstentions, not just successful recommendations.

Research questions and falsifiable hypotheses

The framework leads to testable questions:

  1. Does a source-attributed temporal-geospatial representation reduce stale, incorrect, or infeasible suggestions compared with an LLM-only baseline?
  2. Does a constraint-first planner increase feasible plan rate without materially reducing user-rated relevance or diversity?
  3. Do parallel evaluators improve decision quality enough to justify extra latency and cost over a single grounded planner?
  4. Does user-controlled, explicitly stated preference memory deliver useful continuity without sensitive-trait inference?
  5. On the defined Swedish urban-planning tasks, which available language models offer the best trade-off across Swedish quality, grounded accuracy, latency, cost, deployability, and governance?
  6. Once sufficient consented outcome data exists, do predictive models outperform simpler statistical or rule-based methods on a narrow, measurable forecasting task?

The null hypothesis must remain live: for some tasks, well-designed retrieval, deterministic constraints, and conventional ranking may outperform a more complex multi-agent or world-model system. The burden is on complexity to show measurable value.

Evaluation and research integrity

Create an evaluation set of time-bounded Stockholm planning tasks with source-grounded events, place metadata, explicit preferences, and hard constraints. Keep training and evaluation data separated by time, source, or scenario to reduce leakage. Include adversarial cases: cancelled events, stale opening hours, ambiguous locations, transit disruption, incompatible schedules, sparse user preferences, missing consent, and contradictory sources.

Human evaluators should score factual support, feasibility, relevance, diversity, explanation quality, and whether the system overstates its certainty. With participants' consent, compare recommendations in a prospective pilot and collect only the outcome data needed for the stated questions. Predefine metrics and report uncertainty intervals and failure cases. The paper should publish methodology and results only after tests are completed; no results are claimed in this proposal.

Trust, sovereignty, and human agency are architecture requirements

A system coordinating people and places can easily slide from assistance into social surveillance. The design boundary should be explicit: users control their own memory; participation and matching require clear opt-in; the system does not infer sensitive traits or hidden motives; personal data is not repurposed to train models by default; and users can inspect and delete stored information. Client engagements remain isolated, and client inputs are never used to train products or AI models. Never use covert pressure, dark patterns, or engagement maximisation as the system's objective.

Every recommendation should have a traceable evidence path. Hard constraints should be inspectable and independently testable. If the system lacks current evidence, it should say so, ask a question, or abstain. A model's confidence expressed in fluent prose is not calibrated certainty. For multi-agent execution, follow the principle developed in Deriss's Rogue by Consensus: intent, permissions, and actions must remain legible, and the agents that recommend must not quietly inherit authority to act.

Sovereignty has multiple dimensions: data control, model licensing, deployment location, operational access, auditability, and the ability to replace a provider. Swedish compute may support some of those dimensions, but none is guaranteed by geography alone. Sferical One is relevant as a potential infrastructure route; it is not a prerequisite for validating the first product, and no partnership is assumed. This extends the distinction between infrastructure location and verifiable control in The Sovereignty Stack and Trust-as-Law.

How this can become a new connective layer

The product begins with a narrow experience: a person has a free evening, a particular interest, a time boundary, and perhaps a friend who has opted into making plans. The architecture is general enough to later coordinate a cultural programme, a community gathering, professional introductions, a venue's live inventory, or a city pilot. Each of those will require its own permissions, partner data, constraints, and evaluation.

The opportunity is not to become the owner of every model or every dataset. It is to build the connective tissue that turns fragmented, permissioned signals into timely, explainable, human-directed action across systems. That is where a European Experience Layer can be both useful and distinct: not another feed to keep people looking down, but infrastructure that helps them decide when to look up.

The thesis will stand or fall on measured outcomes. If grounded sequencing does not improve feasibility, usefulness, or trust over simpler alternatives, it should be redesigned. If it does, the resulting evaluation framework and orchestration architecture can inform a more ambitious research programme into predictive real-world intelligence.

Research lineage and sources

Deriss research this paper extends

External technical and institutional references

This is a proposed research framework. It is not a report of validated system performance, and it does not imply that any named institution has endorsed or agreed to collaborate on the programme.

Machine-readable version of https://deriss.com/articles/beyond-the-nordic-llm-real-world-intelligence. Full canonical context: llms-full.txt.