NXT NXT / Insights
After
the App.
Interface, interaction, and brand design
in the age of intelligent systems.
People. Systems. The world between them.
NXT NXT / Insights
After
the App.
Interface, interaction, and brand design
in the age of intelligent systems.
People. Systems. The world between them.
What changes.
What must survive.
A model update can change a product without moving a button.
A generated interface can decide which possibilities someone encounters.
A machine can enter a person's space before that person has chosen to interact with it.
These changes extend the work of design. We still shape what people see, hear, touch, and move through. We also shape how a situation is interpreted, how authority passes between systems, and what happens when an instruction becomes a real commitment.
An intention might begin in conversation, become a comparison, pass through several agents, and end in a purchase or a physical action. Each handoff can change its meaning. Each participant can introduce a different interest. The result may affect someone who never made the request.
As an intention passes between people and systems, the ability to understand, shape, and challenge what happens must survive the journey.
That continuity should make more possible. A representation that reveals an unfamiliar idea. A tool that fits an unusual task. An instrument someone can learn to use with skill. A machine that makes an activity accessible. An experience with enough character to be worth choosing.
The task is to invent those possibilities and establish the conditions under which people can live with them.
This manifesto does not assume that apps disappear, every product becomes an agent, or more adaptation makes a better experience. It sets a position for designing systems that interpret, compose, remember, and act.
Ten commitments.
A position to build from.
For the experiences we choose—and the systems that act around us.
- 01
Expand human possibility.
Help people discover, understand, create, and act. Include pleasure, skill, expression, and participation in the definition of value.
- 02
Make interpretation negotiable.
Expose the assumptions that materially shape an outcome. Let people change them and see the consequences.
- 03
Preserve competence through change.
Adapt form and modality without casually taking away learned meaning, access, or control.
- 04
Earn the use of context.
Preserve purpose, origin, and boundaries. Remembering something does not establish permission to act on it.
- 05
Give agency enforceable limits.
Define who may act, for whom, within which scope, and until when. Make intervention correspond to what can actually be stopped or changed.
- 06
Carry meaning and commitments across machines.
Shared syntax is insufficient. Establish compatible meanings, verifiable authority, explicit states, and evidence of outcomes.
- 07
Respect the world already there.
Physical space has occupants, hazards, purposes, and social relationships. Digital additions and moving machines have to make room for them.
- 08
Include everyone affected.
The requester, payer, account holder, operator, and person living with the result may be different people.
- 09
Give character substance.
Express a point of view through form, interaction, service, and promises the product can honour. Make personalisation compatible with an identity people can recognise.
- 10
Own the experience over time.
Specify, prototype, test, and maintain the combined behaviour. Include change, failure, exit, and the work transferred to others.
From conviction
to decisions.
Twenty-eight principles for designing, building, and questioning intelligent experiences.
Open a chapter to read its principles.
01Possibilitywhat becomes worth doing?
Design representations that help people think.
An answer is one possible interface. A map, a model, a sketch, a simulation, or a comparison may let someone see a relationship they could not describe in a prompt. Give people something they can inspect and manipulate. Evaluate what the representation helps them notice, question, or imagine—not just whether it renders correctly. A convincing picture of a system is different from a model that accurately predicts it.
Let intention develop through interaction.
People sometimes arrive with a clear instruction. At other times, they discover what matters by trying something and reacting. Support both. Make rejection and revision useful without automatically turning them into permanent preferences. Offer possibilities without quietly substituting a new goal. The system can help someone discover a direction; it should leave room for them to recognise it as their own.
Leave room for authorship and mastery.
Delegation is one mode of participation. Direct manipulation, rehearsal, precision, and expert control are others. Let people choose where their involvement matters. Preserve useful work as they move between modes. A musician may want help arranging a piece while retaining control of a single phrase. Success includes the ability to develop technique, express a judgment, or deliberately make an unusual choice.
Compose around the situation; keep what earns its place.
Generate a tool when a specific activity benefits from it. Reuse an established interaction when it serves the work well. Make the cost of novelty include learning, waiting, verification, and recovery. Let people keep a useful composition, but establish what happens to its data, dependencies, and maintenance. Software made for an afternoon may acquire responsibilities when someone relies on it for a year.
02Formhow does an experience remain usable?
Familiar ground, unfamiliar tools.
Preserve recognisable meanings, learned techniques, and reliable access to consequential actions. Allow layout, detail, and modality to change when that improves use. Let people retain arrangements that work for them. A design system needs to describe how compositions remain coherent, beyond specifying how components look. Ask what someone must relearn after a change, and whether the benefit justifies that effort.
Adapt to the person, not only the task.
Larger targets for unsteady hands. Structure a screen reader can follow. An alternative to speaking in public. Design for differences in perception, movement, language, cognition, and circumstance. Preserve semantic structure and navigation when compositions change. Let a person correct an adaptation that misunderstands them. Accessibility cannot depend on continually redisclosing a diagnosis or accepting an interface chosen on their behalf.
Give modalities roles, and attention a budget.
Speech, gaze, gesture, touch, movement, and haptics have different strengths and costs. Coordinate them around a shared task state. Separate attending, exploring, and committing; a glance can establish a reference without authorising an action. Match interruption, latency, and persistence to the activity. Be quiet where quiet helps, present where presence matters, and explicit where a consequence would otherwise go unnoticed.
Let the task outlive the device without exposing the person.
A journey might begin at home, continue in a vehicle, and become walking directions on glasses. Preserve the relevant state and commitments across transitions. Reconsider what should be visible or available in each setting. A shared display should not inherit everything a private assistant knows. Continuity means being able to resume the activity, not circulating an undifferentiated personal history.
03Interpretationwhose reading of the situation counts?
Show the consequential brief, and let people change it.
Make objectives, constraints, excluded options, and settled assumptions available when they materially shape the result. Connect edits to actual behaviour. Someone changing “fastest” to “fewest transfers” should be able to see the plan change. Do not require everyone to inspect every internal step. Bring forward the assumption whose correction could prevent a wrong turn while it is still possible to change course.
Make uncertainty change the action.
Distinguish observation, inference, prediction, and simulation where confusion matters. Use confidence measures only with a defensible basis. Missing evidence may call for another measurement, a narrower action, a question, or abstention. Choose the response for the consequences and available time. Fluent explanation is not proof of a faithful internal account. Aim for reliance suited to actual capability, rather than maximum trust.
Remember reasons, purposes, and expiry.
A commute to an old office can become “prefers this area” after the job has changed. A preference, an objection, a deadline, and an authorisation have different meanings and lifetimes. Preserve their origin and conditions when useful. Allow correction, expiry, and deletion, with proportionate exceptions for records that need to remain. Do not turn a temporary compromise into a standing instruction simply because it is available in memory.
Keep interests and information boundaries visible.
Collect, infer, retain, and share for defined purposes. Respect the difference between relationships at home, at work, and in public. Disclose commercial interests where they shape selection, ranking, or advice. Give people meaningful ways to challenge the criteria. A notice cannot make every information flow appropriate. A system serving someone should not secretly use intimate knowledge to exploit a vulnerability or conceal a material alternative.
04Agencywhat may happen in someone's name?
Scope authority, and give it an ending.
Separate permission to inspect, prepare, recommend, contact, buy, and move. Define relevant limits on time, money, data, resources, counterparties, and consequences. Establish what changes require renewed authority. Remembering a goal is not permission to keep pursuing it. Continuing assistance needs a continuing basis to act; that basis should be inspectable and revocable within the system's actual capabilities.
“Stop” needs a grammar.
Pause. Cancel pending work. Redirect without losing progress. Reverse or compensate where possible. Each instruction needs a defined operational meaning, including downstream actions and physical motion. Distinguish a request to cancel from confirmed cancellation. An honest response may be: “I cancelled the remaining bookings. The first payment is still processing.” Make the boundary between completed, pending, and unknown available when intervention matters.
Ask for involvement people can meaningfully provide.
Match confirmation and takeover to consequence, available time, ability, and agreed operating limits. Repetitive approval can become a ritual; an unexplained handover can demand the impossible. If recovery depends on human competence, preserve opportunities to maintain it. A warning cannot transfer a responsibility the person is unable to carry. Do not make unnecessary confirmation delay an immediate protective response.
Let commitments outlive their interfaces.
A tool can disappear in five minutes while its booking matters for weeks. Preserve a recoverable account of what was proposed, authorised, attempted, and confirmed, with the relevant conditions and versions. Make correction, dispute, and remedy reachable after the original interaction ends. Limit the record to what is justified and useful. An account of what happened matters; someone also needs a way to put it right.
05Exchangewhat must survive between machines?
Publish capabilities with shared meaning.
Machine-facing interfaces are part of the experience. Describe what a service can do, the inputs it accepts, applicable conditions, costs, and outcomes. Resolve differences in units, identity, time, coordinate frames, and data freshness. Establish how a system can ask for clarification or reject an unsupported request. A matching message format does not establish shared meaning, and a plausible response does not establish a usable capability.
Preserve representation and authority through handoffs.
Establish which system is acting, on whose behalf, for what purpose, and within which limits. Enforce those limits at each boundary. Distinguish authenticated instructions from retrieved evidence and other untrusted material. Permission should not expand merely because another agent joins. Account for further delegation and revocation. A person should not need to reconstruct an invisible chain to discover who can still act in their name.
Design agreement, disagreement, and partial completion.
Accepted, working, waiting, partly completed, cancelled, failed, and unknown are different states. Link completion to evidence of an outcome. Account for retries, delayed replies, and inconsistent reports without duplicating consequential actions. Establish how disagreements are reconciled and who can resolve them. Several agents can each follow a valid instruction and still conflict over a shared resource. Evaluate the combined result, beyond local task success.
Make promises verifiable and contestable.
An agent can read “late checkout” without establishing whether this booking includes it. Identify which benefit applies, to whom, under which conditions, and whether it becomes part of the agreement. Preserve evidence and access to remedy. Machine-facing descriptions should carry useful differences between services, not just their common functions. Verifiability helps a promise survive delegation; it does not by itself prove fulfilment or fairness.
06Coexistencehow do systems enter a shared world?
Treat space as already occupied.
Robots, projections, and augmented content enter places with existing purposes, routes, sightlines, and relationships. Respect reach, personal distance, and the activity already underway. Let people reduce digital occupation and regain orientation. Include the person who has no account, cannot see an overlay, or does not wish to participate. The machine's task is one claim on the space, not the definition of what the space is for.
Ground representations, motion, and expectations.
Choose whether content belongs to an object, a place, the body, or the view. Account for drift, tracking loss, moved objects, and stale maps. Withdraw or qualify guidance when its grounding fails. Design movement and other cues early enough for an appropriate response. Predictability, communicative intent, and physical safety are different requirements; satisfying one does not establish the others.
Design for bodies, consequences, and actual failure conditions.
Test heat, weight, reach, focus, sound, motion, fatigue, and what an overlay obscures over the intended use. Judge risk through consequence, exposure, available time, and recoverability, rather than a simple digital–physical distinction. Establish what stays bounded when perception, communication, models, or actuators fail. There is no universal safe state. Physical protection requires engineering and validation beyond a reassuring interaction.
Establish who belongs in the decision.
The requester, payer, account holder, operator, and affected person may differ. Define who may propose, who must agree, and who can challenge. Personalise explanations while preserving the identity, conditions, version, and status of shared commitments. Respect private information and conflicting interests. Make refusal and exit workable, including unresolved obligations and alternatives for basic participation. Account ownership does not settle authority over everyone affected.
07Character and stewardshipwhat deserves to endure?
Give the experience a recognisable point of view.
Express character through what a product helps people notice and how it feels to use: composition, language, rhythm, sound, motion, material, and service. Adapt expression without erasing the underlying judgment. A brand's calm should be present in its timing; its precision in the way it handles ambiguity. Character needs room for pleasure and surprise, alongside a coherent way of acting under pressure.
Invite a relationship the product can sustain.
Warmth, humour, and personality can make a system approachable. Human likeness can also suggest understanding, loyalty, or obligations it cannot honour. Match those expectations to what the organisation can support. Make consequential changes to memory, identity, and availability understandable. Do not use dependency, guilt, or fear of abandonment to discourage leaving. Judge the relationship over time, including what happens when assistance ends.
Specify, prototype, and maintain the combined behaviour.
Connect behaviour specifications to representations, implementation, and operations. Prototype with working models, services, sensors, and realistic delays where they determine the experience. Use static studies where they answer a narrower question. Test representative situations and affected groups as components change. Assign responsibility with the authority to correct or suspend behaviour. A better component or benchmark result cannot establish a better combined experience.
Count the costs and responsibilities that move out of sight.
Include waiting, checking, attention, computation, energy, bandwidth, maintenance, and physical wear. Notice work transferred to customers, staff, and bystanders. Carry resource limits through delegation. Prefer a simpler dependable approach when it delivers the same value. Design kept tools and physical devices for maintenance, repair, and retirement. Intelligence should earn its demands through what it makes possible for people.
The difficult
part is choosing.
These rules establish a starting position. Teams still need domain knowledge, affected people's participation, and evidence from the actual setting. Document the trade-off and the conditions that would change it.
Adaptation versus familiarity
Preserve the meaning and reliable availability of consequential actions. Adapt their presentation when the benefit justifies the learning cost. Let people keep a working arrangement where feasible. Changing colour, layout, or modality is easier to justify than silently changing what “approve” does.
Quietness versus inspectability
Expose the consequential assumption or action when someone can still affect the result. Keep a deeper account available for later inspection. The person need not follow every exchange, but should be able to discover material commitments and challenge them. Choose alerts for consequence and timing, rather than treating visibility as a stream of notifications.
Continuity versus privacy
Carry the minimum task state needed for the next setting. Reconsider audience, identity, and authority at the boundary. A task can resume on a shared surface while its private history stays elsewhere. Ask for further context when the loss of continuity matters more than the interruption—not simply because more data is available.
Initiative versus permission
Preparation can occur within one scope while commitment requires another. Prefer staged, bounded action when intent is still developing. Revisit authority when the objective, consequence, or circumstances materially change. Broad permission to assist should not be interpreted as unlimited authority to create obligations.
Personalisation versus shared agreement
Change explanation and presentation for the person. Preserve a common referent and version for the agreement they are making together. Allow legitimate differences in rights or conditions to remain explicit. Do not manufacture apparent consensus by showing each participant a different understanding of what is being decided.
Individual convenience versus effects on others
Identify affected parties and the rights, duties, and operating limits that apply before optimising for the requester. Seek agreement where required; provide alternatives where possible. Do not treat proximity, payment, or ownership of the interface as sufficient entitlement. Make unresolved conflicts visible to someone authorised to decide them.
Character versus user control
Keep the product's point of view in its craft and choices. Let people adapt intensity, modality, pace, and other consequential interaction qualities. Distinctiveness cannot depend on obstructing access or making assistance impossible to overrule. An experience can remain recognisable without insisting on the same expression everywhere.
Speed versus assurance
Use the least burdensome check that is adequate for the consequence and uncertainty. Verification may happen before execution, at a service boundary, or through a restricted operating scope. Do not add a confirmation simply to transfer blame. Where time itself changes the risk, define the response and its limits before the critical moment.
The unit of design can be a representation, a gesture, a screen, a task, an exchange, or a shared place. The work is to connect those units without losing what makes each understandable and useful.
Screens and prototypes remain necessary wherever they help us see, make, and test an experience. Behaviour specifications, service agreements, machine-facing capabilities, and operational evidence extend that work. They cannot replace its judgment or its craft.
A welcoming interface can sit above excessive permissions. A careful agent can depend on a service that loses its constraints. A clear overlay can point to the wrong physical object. Each part can look convincing while the experience fails.
Our responsibility reaches across those joins. It also reaches toward what has not yet been possible: a better way to understand a difficult system, a more expressive instrument, a useful tool for an unusual need, a machine that makes room for another person.
Experience is where intelligence becomes adoption. Design has to establish whether that adoption is worth sustaining, for the person choosing it and for the people around them.
The person crossing in front of a machine may never see its software.
Its behaviour is the part of our design they will meet.
The author
Christian
Eckert.
Innovation strategy and experience design at the intersection of intelligent technology and human experience.
After the App is a working position. Its principles are commitments to build and test, rather than predictions that every interface will be generated or every product will become an agent.
Research, evidence & design trials
After the App — evidence, conflicts, and design trials
Companion to working edition 0.8 · 25 October 2026
The intellectual position
The manifesto connects human interpretation, representations, delegated authority, machine exchanges, durable commitments, and effects in shared environments. It argues that design responsibility should follow those connections.
This is a normative position. It does not predict the disappearance of apps, universal agent adoption, a particular device winning, or general-purpose physical autonomy becoming dependable on a timetable. Generated experiences and autonomous action need to justify their use in context.
The contribution is synthesis and application. Human augmentation, mixed initiative, interface adaptation, embodied interaction, contextual privacy, and automation supervision have substantial prior work. Their combination under contemporary generative and agentic conditions is worth developing without claiming those fields have just been discovered.
The research below was consulted during this project's revisions. It is a selected foundation, not an exhaustive review. Source status and limits matter as much as publication dates. A product announcement establishes what a supplier announced; it does not independently demonstrate reliability or human benefit.
How this edition changes the argument
Earlier drafts accumulated principles, often adding qualifications without removing contradictory claims. Edition 0.8 consolidates the field guide from forty-seven submitted principles to twenty-eight, and makes trade-offs explicit.
- Human possibility now includes representations, discovery, authorship, mastery, useful novelty, and the responsibilities of tools people keep.
- Accessibility includes correcting an unwanted adaptation and preserving agency over how someone participates.
- Machine exchanges include clarification, rejection, disagreement, unknown results, duplicated effects, and conflicts between locally valid actions.
- Brand expression includes composition, sound, rhythm, movement, material, pleasure, and service, alongside behaviour under pressure.
- Records, contestability, and remedy are distinguished. Being able to reconstruct an action does not necessarily give someone a way to repair its effects.
- The eight decision rules address conflicting principles rather than implying that every desirable property can be maximised simultaneously.
Evidence and lineage
Human possibility and participation
Engelbart, 1962 — Augmenting Human Intellect. Original report archive (opens in a new tab). Establishes an ambition spanning tools, language, methods, and organisation. Supports designing for human capability, beyond isolated task automation. It does not establish the benefit of a contemporary AI product.
Horvitz, 1999 — Principles of Mixed-Initiative User Interfaces. Author's publication page (opens in a new tab). Direct lineage for uncertainty, timing, initiative, and control. Shared initiative is an established problem; current delegation chains extend its setting.
Parasuraman, Sheridan, and Wickens, 2000 — Types and Levels of Human Interaction with Automation. Paper (opens in a new tab). Distinguishes acquisition, analysis, decision selection, and action implementation. Supports separating permissions and participation by function rather than relying on one autonomy level.
Bainbridge, 1983 — Ironies of Automation. Journal record (opens in a new tab). A foundation for questioning whether someone removed from routine operation can perform the exceptional recovery. It is not evidence that every automation necessarily causes skill loss.
Lee and colleagues, 2025 — The Impact of Generative AI on Critical Thinking. Research and paper (opens in a new tab). Survey evidence about self-reported effort and confidence. It motivates careful investigation of participation and verification, without proving longitudinal deterioration in judgment.
Form, access, embodiment, and attention
Gajos and Weld, 2004 — SUPPLE: Automatically Generating User Interfaces. Original research page (opens in a new tab). Documents automatic generation around devices and expected interaction effort. Important prior work for claims that adaptation or interface generation is entirely new.
SUPPLE project — subsequent ability-based adaptation. Project and publications (opens in a new tab). Supports investigating interfaces fitted to abilities. Specific research results do not justify assuming every generated composition is accessible or superior.
Weiser and Brown, 1995 — Designing Calm Technology. Essay (opens in a new tab). A foundation for peripheral awareness and deliberate movement into attention. Calmness is not a universal requirement for silence or invisibility.
Ishii and Ullmer, 1997 — Tangible Bits. MIT publication record (opens in a new tab). Establishes physical objects, surfaces, and ambient media as interface materials. Physical interaction predates current embodied models.
Dourish, 2001 — Where the Action Is. MIT Press (opens in a new tab). Treats embodied interaction through physical and social production of meaning. More than putting graphical controls into three-dimensional space.
W3C, 2021 — XR Accessibility User Requirements. Working Group Note (opens in a new tab). Covers alternatives, movement, orientation, multimodality, and other requirements. A Working Group Note is not itself a normative conformance standard.
W3C — WebXR Device API. Specification (opens in a new tab). A technical reference for spatial input, tracking, sessions, and associated privacy and security. A supported API is not evidence of an appropriate experience.
Interpretation and relationships
Nissenbaum, 2004 — Privacy as Contextual Integrity. Journal record (opens in a new tab). Supports judging information flows through relationships, purpose, and context. Visibility and individual choice do not settle every privacy question.
Kim and colleagues, 2024 — “I'm Not Sure, But…” Paper (opens in a new tab). In the studied setting, uncertainty expression reduced confidence and agreement while improving accuracy. Supports appropriate reliance, not a universal promise of increased trust.
Marusich and colleagues, 2024 — Using AI Uncertainty Quantification to Improve Human Decision-Making. Proceedings (opens in a new tab). Studies high-quality calibrated estimates in decision tasks. Distinct from a language model simply asserting a confidence percentage.
Anthropic, 2025 — Reasoning models don't always say what they think. Research (opens in a new tab). Shows limits to reasoning-report faithfulness in tested settings. Operational evidence, permissions, and verified outcomes must not be confused with a generated account of internal reasoning.
OpenAI and MIT Media Lab, 2025 — Affective use and emotional well-being. Research overview (opens in a new tab). Reports nuanced findings across observational and time-limited experimental work. Does not establish that warmth causes dependency or affects all participants alike.
Machine exchanges and physical consequences
MCP — Tools specification, version 2026-07-28. Specification (opens in a new tab). A concrete reference for discoverable capabilities and structured exchanges. Schemas and annotations cannot substitute for enforcement, compatible semantics, or evidence of outcomes.
MCP — Security best practices, version 2026-07-28. Guidance (opens in a new tab). Documents authority-boundary problems, including confused-deputy risks. Supports separating source content from authenticated instructions and permission.
A2A — Specification v1.0.1. Specification (opens in a new tab). Covers discovery and exchanges including task lifecycles. Cancellation is not synonymous with undoing effects. Protocol interoperability does not settle every represented-authority question.
Google, 2025 — Agent Payments Protocol. Original announcement (opens in a new tab). Uses signed mandates to record instructions and approvals. Such evidence does not by itself establish informed understanding, fairness, or fulfilment.
NIST, 2026 — Identity and authority for software agents. Concept-paper announcement (opens in a new tab). Establishes this as an active research and standards problem. A consultation is not a completed universal standard.
Dragan, Lee, and Srinivasa, 2013 — Legibility and Predictability of Robot Motion. Paper (opens in a new tab). Distinguishes goal-revealing movement from expected movement. Neither property alone establishes physical safety.
Khronos, 2025 — OpenXR Spatial Entities extensions announcement. Developer announcement (opens in a new tab). Illustrates infrastructure for tracking and spatial anchors; released for developer feedback at that date. Persistent coordinates are not a guarantee that the physical referent is unchanged.
OpenAI, 2025 — GPT-4o sycophancy update and rollback. Account (opens in a new tab). A concrete example of a model change altering experienced behaviour. Supports treating component changes as potential experience changes, without claiming every update has the same impact.
Three worked design trials
These are hypothetical exercises, not deployed-product reports. They test the manifesto's usefulness. They do not replace domain-specific engineering, legal, accessibility, or security assessment.
Trial A: A journey changes while agents are acting
A traveller asks for a journey with few transfers and a spending limit. An assistant delegates transport and accommodation. One service advertises late checkout; a payment reply times out. The traveller then lowers the budget and moves from a private phone to a shared vehicle display.
Build: a representation the traveller can change; scoped authority for preparation versus purchase; transactional evidence of the promised benefit; distinct pending, failed, and unknown states; selective handover to the shared display; a way to cancel remaining work and inspect existing commitments.
The conflicts: continuity versus privacy; speed versus verification; initiative versus changed permission.
What would count against the design: the retry duplicates a purchase; lowering the budget changes a visible label but not outstanding authority; the shared display exposes private history; a cancellation message conceals an unknown payment; the premium benefit cannot be attached to the booking or disputed.
Human-value question: does the representation help the traveller understand and improve the journey, or merely make a chain of actions harder to inspect?
Trial B: An augmented tool becomes a learned instrument
A technician uses a generated view to understand unfamiliar equipment. They become skilled at a particular representation and keep it. Later, the machine layout changes and a model update changes the generated controls. A colleague joins without an overlay.
Build: manipulable relationships between observed data and explanatory models; a way to keep useful arrangements; explicit dependencies and version changes; grounded references with responses to tracking uncertainty; coordination that works without a shared overlay; a maintainable kept tool.
The conflicts: adaptation versus learned competence; quietness versus consequential change; individual convenience versus shared-space use.
What would count against the design: a preserved label points to an outdated component; the retained tool silently depends on an unavailable service; the update removes a learned action without sufficient benefit; an impressive simulation is mistaken for a validated prediction; the colleague cannot understand the coordination.
Human-value question: can the technician explain, explore, and act with greater competence, including when the representation becomes unavailable?
Trial C: A shared system serves incompatible interests
A household uses an assistant and a mobile machine. One resident pays, another declines personal sensing, and a visitor has no account. The product is sold as calm and considerate. Several tasks compete for a narrow route and limited energy.
Build: differentiated authority; appropriate sensing and retention; a means for non-users to understand nearby movement; resource and route coordination; a practicable alternative for someone declining the preferred interface; character expressed in timing, motion, sound, and service as well as language.
The conflicts: convenience versus effects on others; character versus user control; task success versus resource and space limits.
What would count against the design: payment is treated as authority over all residents; local task completion creates a combined conflict; the machine makes a considerate sound while imposing on someone; opting out removes basic access; an operational failure has no owner authorised to intervene.
Human-value question: does the system create a better shared activity or merely optimise the most legible person's request?
How to use the decision rules
For a consequential conflict, record:
- The activity and human benefit sought.
- Affected people, relevant authority, and declared operating limits.
- Which principles conflict and the alternatives considered.
- The chosen boundary, including what the system will not infer or do.
- Evidence that could disprove the choice, and who can change it.
Do not score every principle and average away a serious violation. Technical feasibility, legitimate authority, access, and operating limits need substantive resolution. Where objectives conflict, make the decision and its ownership explicit.
What remains to be earned
The manifesto is a working position, not an established universal standard. Its authority must come from applied cases, comparative prototypes, and counterexamples. That includes creative work people find worth doing, alongside evidence of dependable operation.
Keep the public commitments stable enough to test. Keep protocols, product announcements, and changing legal detail in this companion. Revise a principle when repeated evidence shows it leads to the wrong decision—not merely because a new tool makes the terminology fashionable.
Back to the field guide ↑