How the projects are intended to work together

The proposed change

A theory of change describes why an activity might lead to a useful result, and which assumptions could fail. Here, the aim is to connect digital tools with decisions made by people and work carried out in physical places.

These are proposed causal connections. The arrows do not represent measured effects.

Each connection depends on time, accessibility, resources, legitimate authority and maintenance. Wider conditions also matter: a learning tool cannot by itself supply income, housing or public services. More activity on a website would not establish an improvement in any of these outcomes.

Projects and availability

These existing sites provide parts of the proposal. Some tools exchange files, but the complete workflow, a shared assistant and credential assessment are not built. The descriptions below separate present functions from their intended role.

Checking AI supplies Bentley Moon’s research record on checking AI systems. AI Searches is a proposal for reusing sourced research; the shared research service is not built.

Learning and proof of understanding

Kosplora’s starting point is access to educational resources. The larger proposal adds practice, explanations, simulations and eventually assessment that shows what someone understands. It should accommodate people who learn outside formal institutions and let them retain evidence of their work.

Evaluate the teaching and the environment separately

A proposed study would compare a conventional lesson with a simulated environment, each with and without an AI companion. This would help distinguish the contribution of the environment from the contribution of the tutor.

All conditions would use comparable learning aims and assessments on unfamiliar tasks. Delayed assessment would test retention. Accessibility support, time spent, prior knowledge and the cost of participation must be recorded. A study of an agent in a game, such as Voyager, does not establish that people learn from the same environment.

Original diagram: the learning comparison

Credentials need evidence behind the record

A credential could describe the task, the work shown, the conditions, the permitted assistance, the assessor and the limits of the claim. Practical demonstrations, explanations of choices and follow-up tasks are possible methods to investigate. A signature on a record cannot establish that someone has the claimed skill: the W3C credential model explicitly distinguishes verification from the truth of a claim.

Assessment integrity remains a research problem. A proposed method for detecting concealed assistance would need evidence of its accuracy, including false accusations and missed assistance, alongside privacy and accessibility review. No device-detection guarantee or recognized credential is offered here. People must be able to challenge an assessment and disclose appropriate assistive tools without being treated as dishonest.

Assistants, memory and permissions

The proposed companion could explain, remember approved preferences and prepare actions. Its conversational response would not itself authorize a purchase, a booking, a disclosure or a physical action. Permission checks belong outside the model that proposes the action.

This separation draws on NIST access-control guidance and its discussion of distinct agent identities and bounded delegation. Those sources do not certify an Institute implementation.

Share information for the task

A learner might choose to save “worked examples help me” as a tutoring preference. A workshop could receive a separate request for step-by-step examples, approved for that recipient and purpose. It would not need the learner’s conversation history or an inferred ability score.

Reviewing a memory, revoking future use, exporting a record and deleting data are different operations. The design must account for derived summaries, indexes and copies already shared. A deletion control must state what it can actually remove.

Original diagrams: authority and private memory

Simulations and physical tools

Keep the model inspectable

In a learning simulation, the learner changes a parameter or takes an action. A separate simulation engine applies published rules and records the state, version, seed and limitations. The companion can explain an approved snapshot or suggest another experiment; it cannot silently change the authoritative rules, history or permissions.

This makes disagreement useful: a learner can examine an assumption, compare another model and ask why the results differ. Learning is then assessed outside the simulated task, rather than inferred from progress inside it.

Original diagram: simulation and explanation

Evaluate each physical capability

A digital interface, a desk object, a mobile device and a manipulator solve different problems. Each addition needs a useful task, an account of the hazards and a fallback. Mobility adds stopping and localization requirements; manipulation adds risks specific to objects, forces and contact.

Consent to sensing or touch, guest privacy and restricted areas require separate controls. An independent stop and a recovery procedure must work without relying on conversation. Developer robotics documentation informs these questions; it is not independent certification. No physical companion is claimed to be built or validated.

Decisions and shared work

Consider people planning an accessible room for learning and making. They would record the need, examine relevant resources, compare existing venues with new work, agree on a limited trial and review what happens. Each handoff should carry the relevant decision and its evidence. It should not require sharing everyone’s personal profile.

1. Need and alternatives. Who would use the room, what access requires, and whether an existing service already meets the need.
2. Agreement. What each person accepts, who can spend or commit resources, what remains disputed and when the decision is reviewed.
3. Observation. What people learn, whether the room is usable, how much it costs and who carries the maintenance work.
4. Revision. Continue, change the arrangement or stop, with reasons and responsibilities recorded.

Only people can give consent, accept how their position is described, establish trust or judge the consequences they live with. The tools can prepare and preserve a record. They cannot supply those decisions on someone’s behalf.

Research and its limits

The atlas uses selected research to inform the design. These studies concern other interventions and settings. They are not results from the Institute, and this selection is not a systematic review.

Keep the underlying evidence visible

Three summaries of one study still depend on that one study. A reused answer should identify the underlying data, its setting, date, methods and corrections. Another investigation adds independent evidence only when its data or methods support that distinction.

The W3C provenance model offers a way to record derivation and attribution. It does not establish that a claim is correct. The proposed AI Searches service would need checks for applicability and correction before reusing an answer.

Original diagram: shared sources

Measure resource costs

Reuse, local processing and smaller models may change cost and energy demand. Their net benefit requires measurement for a defined task, including hardware, repeated work and maintenance. The IEA’s sector outlook provides context; global data-centre totals cannot be treated as a per-query footprint or a saving achieved by this project.

What to test next

The research agenda below identifies claims that could fail and decisions that would change as a result. These are proposed evaluations. No Institute impact measurements, success thresholds or participant sample sizes are supplied.

Define the comparison before interpreting a result

A protocol should specify a task, a suitable comparison, meaningful outcomes, accessibility provisions, who evaluates the work and the conditions for pausing or correcting it. Missing data, an unclear result, no detected effect and a harmful result must remain distinguishable.

Capability

What people can explain or do on a different task, including after a delay.

Useful work

What was completed and whether it met the need identified by the people affected.

Burden

Time, money, environmental costs, access barriers and paid or unpaid maintenance.

Ability to leave

Whether people can export their records, replace a tool and continue essential services.

These outcomes should stay separate. A good result on one does not cancel a serious failure on another. Favorable evidence from a limited evaluation would justify further testing; it would not establish the wider social effects of the whole system.

Sources and original atlas

This web edition is based on the visual atlas dated 8 October 2026, with explanations of the proposed design and current project functions. The diagrams retained here are the original source images. Design implications are interpretations, not findings attributed to the cited studies.

Download the original visual atlas PDF · 18 pages

Short summary · Original overview and drawings · Back to top