# Sentient Design > [!summary] Analytical finding > *Sentient Design* is a strong pattern language and workshop guide for designing AI-powered interfaces. Its four postures, 14 experience patterns, manager-experience lifecycle, defensive-design techniques, and five-phase sprint give product teams a shared vocabulary for deciding how an intelligent system should relate to a user. Use the Sentient Triangle as a compass, not a measurement model. Treat the many product examples as demonstrations of possibility, not proof of general effectiveness. Before shipping, add independent evaluation, least-privilege security, accessibility acceptance criteria, participatory research, labor analysis, and risk-based governance. > [!question] Reader context still needed 🔶 > How you found this book and what you wanted it to help you decide were not supplied. This review does not invent either answer. ## Analytical synthesis ### The relationship should precede the feature Josh Clark and Veronika Kindred begin with a useful reframing: AI is a design material, not a feature category. Designers need to understand its grain, including probabilistic behavior, open-ended generation, context dependence, and uneven reliability. This framing prevents a common product mistake. A team should not begin with “add chat” or “build an agent.” It should begin with the user’s goal, the system’s role, and the amount of control each party needs. The book organizes those relationships into four postures: - People **use tools**. - People **talk to chat**. - People **delegate to agents**. - People **are backed by copilots**. This compact set of verbs does more work than a technology-first taxonomy. It tells a team what the interface asks of a user. Tools invite direct control. Chat asks for turn-taking. Agents convert the user into a manager. Copilots watch context and offer help near the moment of need. The categories overlap, which the authors acknowledge. That overlap is productive because many sound products combine modes instead of forcing every task into chat. The four postures sit on the Sentient Triangle, whose corners are **grounded**, **interoperable**, and **radically adaptive**. Fourteen experience patterns occupy regions between them: dedicated tools, workflow tools, inline tools, sculptors, characters, assistants, NPCs, bespoke UIs, intelligent canvases, taskers, alchemists, stewards, conductors, and copilots. Lower-level functional, interaction, and UI patterns make these experiences concrete. This layered vocabulary is the book’s central contribution: 1. Choose a relationship posture. 2. Choose an experience pattern. 3. Add the functional behavior needed for the task. 4. Express it through interaction and UI patterns. 5. Test whether the combination preserves agency under failure. That sequence can keep a team from treating every AI capability as a chatbot. It also separates what a system can do from how a person will work with it. ### The Sentient Triangle is a generative prompt, not a score The triangle is memorable and useful in workshops, but its geometry should not be read literally. Grounding, interoperability, and adaptivity are not proven to trade against one another. “Grounded” also compresses several different properties: access to current information, domain relevance, factual accuracy, repeatability, and trustworthiness. A system can be well grounded in one sense and weak in another. The placements are author judgment. The book provides no inter-rater test showing that independent designers classify the same product similarly, and no comparative study showing that a triangle-guided design performs better. The authors partly solve this problem themselves when they call it “more compass than map.” Use it to ask questions: - What changes if this experience becomes more adaptive? - What information must ground it? - Which systems must it cross? - Who must review its work? - What new risk appears when it moves toward autonomy? Do not use the triangle as a maturity ladder. A dedicated tool is not less advanced than a conductor. The right posture depends on the job, the stakes, and the user’s capacity to supervise. ### Familiar controls can carry intelligence better than chat Chapters 4 through 6 show why intelligent products need not abandon established interface forms. Dedicated tools make one capability legible. Workflow tools pass an artifact through a larger process. Inline tools act near the selected content. Sculptors retain a shared artifact while the user iterates. Bespoke UIs can generate interface fragments from intent while remaining bounded by a design system. Intelligent canvases use spatial or document conventions to let users shape the result. This material is strongest when it treats familiarity as a constraint. Generated UI should draw from stable components, predictable interaction rules, and visible recovery paths. The system can choose among known elements without inventing an inaccessible interface language on every turn. The authors’ architecture for call-and-response UI is especially useful: an LLM interprets intent and returns structured instructions; a deterministic UI engine renders approved components. The raw model response does not become the interface. The limitation is evidence. Named products show that these patterns exist, but the book rarely compares task completion, learnability, error rate, accessibility, or maintenance cost. Hidden contextual menus, gesture-based selection, and shifting layouts may reduce friction for one user while creating discoverability, motor, keyboard, screen-reader, or cognitive barriers for another. Generated polish can also hide weak provenance or make a merely plausible result feel complete. ### Agent UX is manager UX The agent chapters give the book its clearest operational model. Once a person delegates work, the interface must support a different lifecycle: 1. Set the mission. 2. Review the plan when risk warrants it. 3. Monitor progress at the right level of detail. 4. Interrupt or redirect the work. 5. Review the result using familiar domain conventions. This is “delegation, not abdication.” A mission needs an outcome, constraints, context, and success criteria. An agent needs a limited toolkit, explicit permissions, a visible home, a stop control, and checkpoints before consequential actions. Taskers, alchemists, stewards, and conductors differ in scope and persistence, but each raises the same question: what evidence lets the user supervise without doing the whole task again? Current security guidance strengthens this part of the book. OWASP’s prompt-injection and excessive-agency guidance recommends bounded behavior, validated outputs, least privilege, separation of untrusted content, adversarial testing, and human approval for high-impact operations. That support matters because the book’s pattern language alone cannot enforce restraint. An “action guard” is useful only when permissions and authorization are enforced outside the model. The manager experience can become a second job. Activity streams and generated plans may create visibility without establishing correctness. Some users lack the domain knowledge, time, energy, or organizational authority to judge an agent’s work. A responsible product must decide who is qualified to review, what happens when no qualified reviewer is available, and whether the task should be delegated at all. ### Copilots turn context into both help and surveillance The book’s situational, session, and continuous copilot categories are useful because they connect helpfulness to the depth of context observed. A situational copilot handles a bounded moment. A session copilot follows an activity. A continuous copilot watches across time, systems, and devices. As contextual depth grows, so do privacy and power concerns. A continuous system may infer health, religion, relationships, routines, disability, or emotional state. It may observe bystanders who never consented. It may retain a mistaken inference and adapt future experiences around it. The book calls for visible memory and controls, but it does not supply a full consent model covering sensors, retention, third parties, local processing, correction, deletion, and appeal. The distinction between personalized and individualized experiences is still valuable. Immediate context can reduce dependence on stale profiles. It does not make surveillance disappear. Collection purpose, retention, inference, and user control matter whether the data lives for one minute or one year. ### Defensive design is the durable center Chapters 10 and 11 give the framework its ethical and operational weight. The authors classify failure around wrong context, bad content, broken formats, and bad actors. They recommend presenting generated results as signals, making uncertainty legible, showing provenance, offering competing strands, keeping important actions reviewable, collecting structured feedback, and fitting oversight to impact. Several defensive patterns survive changes in model vendors: - match scope and user expectations to actual capability; - constrain output structure when open-ended expression is unnecessary; - retrieve from trusted sources and show citations; - distinguish critics, which suggest repair, from evaluators, which score or block; - expose disagreement instead of manufacturing one answer; - let users correct memory and generated output; - keep high-impact actions suggestive until an authorized person approves; - measure prohibited outcomes as carefully as desired ones; - retain deterministic authorization, validation, logging, and rollback outside the model. Some recommendations need repair. Plain-language confidence cues are not automatically better than probabilities; professionals may need calibrated numbers plus accessible explanations. Ensemble agreement is not independent evidence when models share training data or copied errors. RAG can reduce unsupported generation but does not guarantee truth, relevant retrieval, secure sources, or resistance to prompt injection. An evaluator built from another model can fail in correlated ways. ### The sprint starts well and needs earlier participation The five-phase Sentient Design sprint moves from outcomes to frictions and hypotheses, magical solutions, practical postures, and a minimum magical product. “Productive magic” asks teams to name an impossible ideal before implementation limits narrow the field. A minimum magical product tests capability and behavior through a working prototype instead of a polished static mockup. The process correctly begins with human and business outcomes rather than an AI feature. It treats building as part of designing because model behavior becomes visible only through interaction. Its design brief connects outcome, user story, posture, safeguards, feasibility, and measures. The risk is solution anchoring. A vivid magical concept can dominate before a team understands institutional limits, excluded users, consent, maintenance, or who absorbs correction work. Participatory research with disabled and marginalized users should shape the outcome and concept phases, not appear mainly during late testing. A health example needs safety, privacy, language equity, accessibility, and workflow evidence before a prototype is treated as a promising direction. The book claims ambient clinical documentation can save more than two hours per day and reduce physician burnout by 20–60 percent. Current evidence supports promise but not that general effect size. A 2025 quality-improvement study reported burnout falling from 51.9% to 38.8% after 30 days, while a randomized trial found a 9.5% time-in-note reduction for one product and no significant change for another. The book’s figures need a cited population, design, duration, and outcome definition before reuse. ### The future-of-work conclusion is too reassuring The final chapter argues that faster implementation moves design upstream toward problem definition, judgment, taste, oversight, and collaboration. That is a plausible and useful direction. The claim that task automation usually does not erase roles is less secure. Organizations can rebundle remaining tasks, reduce headcount, shift correction work to customers, or reserve “upstream” roles for people who already hold power. The International Labour Organization’s 2025 task-level analysis supports transformation as the dominant near-term frame while showing unequal exposure by occupation, income level, and gender. That is more conditional than the book’s assurance. Teams should ask who gains creative scope, who becomes a monitor, who performs unpaid verification, whose work supplied training data, and who can refuse automation. The book is right that prompts and working prototypes can become shared artifacts. It overstates the system prompt as a common source of truth. Production behavior is distributed across prompts, code, data, retrieval, policies, model configuration, permissions, tests, monitoring, and regulatory records. “The brief becomes the thing” works as a collaboration slogan, not an architecture. ### The artifact contradicts its accessibility aspirations The text presents individualized interfaces as an accessibility opportunity. It discusses sensory, cognitive, physical, language, and situational adaptation. It also raises real concerns about blind users trusting generated descriptions and about dialogue flattening dialect. The EPUB itself fails a basic test. Its package says it generally meets WCAG 2.2 Level AA and conforms to EPUB Accessibility 1.1 at WCAG 2.2 Level A, while its own accessibility summary admits that non-text alternatives are missing. Of 180 image elements, 178 have empty `alt` attributes. Only the front and back covers have descriptive alternatives. Captions help in visible reading, but an empty alternative removes the figure from many assistive-technology reading paths. The central triangle, product screenshots, workflow diagrams, and confidence-meter examples therefore aren’t equivalently available. Adaptive interfaces must begin with baseline conformance. Personalization cannot excuse missing semantics, unstable focus, inaccessible generated controls, absent reduced-motion behavior, or color-only confidence cues. Users also need to choose adaptation; a system should not infer disability and silently redesign the interface around that inference. ### Recommendation Read the book fully if you design AI-powered products, product strategy, or agent experiences. Keep the vocabulary, manager lifecycle, defensive patterns, and sprint exercises. Use the appendix as a pattern-selection catalog. Pair it with current security, accessibility, labor, and governance sources before implementation. Do not treat the framework as validated taxonomy, an accessibility standard, or a safety case. Require task-specific evidence and acceptance criteria. The most useful operating rule is simple: choose the human-system relationship first, then earn each increment of adaptivity and agency with evidence, controls, and a credible recovery path. ### Further reading - [NIST AI Risk Management Framework and Generative AI Profile](https://www.nist.gov/itl/ai-risk-management-framework): lifecycle governance for mapping, measuring, and managing AI risk. - [OWASP: Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) and [Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/): security controls missing from a design-pattern catalog. - [W3C WAI-Adapt Explainer](https://www.w3.org/TR/adapt/): standards-oriented interface personalization for cognitive and learning accessibility. - [European Commission AI Act overview](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai): current transparency, governance, and high-risk-system obligations. - [ILO, *Generative AI and Jobs: A Refined Global Index of Occupational Exposure*](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure): task exposure and unequal labor effects. - [JAMA Network Open ambient-scribe study](https://doi.org/10.1001/jamanetworkopen.2025.34976) and [randomized trial record](https://pubmed.ncbi.nlm.nih.gov/40672471/): current evidence behind the book’s medical example. - [U.S. Copyright Office AI initiative](https://copyright.gov/ai/): current reports on digital replicas, copyrightability, and training. --- ## Session record and shared inputs | Field | Resolved value | Status | |---|---|---| | Title | *Sentient Design: Crafting Intelligent Interfaces with AI* | title page and package metadata | | Authors | Josh Clark with Veronika Kindred | title page; package metadata lists Clark only | | Year and edition | 2026 Rosenfeld Media ebook edition | confirmed | | Book ID | Ebook ISBN 978-1-959029-28-1; paperback ISBN 978-1-959029-18-2; no Reader ID | confirmed | | Book type | Technical design and product-practice book | system classification | | Purpose | 🔶 Evaluate the framework as an operating guide for intelligent interfaces and retain durable patterns | inferred, not reader-supplied | | Depth | Analytical | system recommendation | | Constraints | 🔶 One consolidated note; local EPUB; no target highlights; current claims may drift; no adjacent note creation or Reader publication | partly supplied, partly inferred | | PARA target | `4. Archives/Book reviews` | supplied convention | | Existing knowledge | 🔶 Not supplied; no reader beliefs are assumed | unresolved | | DEIB intent | 🔶 Test accessibility, agency, bias, labor, privacy, and multicultural transfer | inferred, not reader-supplied | | Readwise tags/date range | None; target record not found on 2026-08-27 | retrieval result | > [!warning] Inferred inputs > Every 🔶 value is a system placeholder, not Courtney’s stated belief or motive. Replacing one may change the questions, priors, applications, teaching plan, and experiments below. ## Prompt 0: retrieval, provenance, and source anchors ### Provenance record - Supplied file: `Sentient Design (Josh Clark) (z-library.sk, 1lib.sk, z-lib.sk).epub` - Integrity: complete `unzip -t` passed with no compressed-data errors. - SHA-256: `80c9ba9ee3d694e71887488542275ae6d91bd4f1ba3ae06420b5ac0c9220ba53` - Package: EPUB 3.0, 31 spine items, 207 archive entries, 31,622,002 uncompressed bytes. - Metadata: Rosenfeld Media; 2026; `en-US`; ebook ISBN 9781959029281. - Text extraction: approximately 106,306 plain-text words. - Reader search: no matching target EPUB or PDF by exact title and author query; false-positive documents were rejected. - Reader highlights: none available for the target title. Every excerpt and emphasis below is `system-selected`. ### Layer 1 substitution and cost The local EPUB replaces Readwise as Layer 1. It preserves the publication’s metadata, spine order, headings, internal section IDs, page-break markers, captions, tables, figures, and prose. It does not provide reader emphasis, annotations, colors, timestamps, or reading history. The analysis must not present system-selected passages as Courtney’s priorities. The requested separate highlights and reading notes were not created because the controlling request requires one consolidated file. Provenance, source map, session record, analysis, teaching material, experiment candidates, and visual specifications remain here. ### Coverage - Read in full: FAQ, foreword, introduction, Chapters 1–13, appendix pattern library, and author biographies. - Mapped and sampled: endorsements, copyright, acknowledgments, publisher material, and index structure. - Not evaluated entry by entry: the alphabetical index. No substantive claim depends on index coverage. - Claims from the full 13-chapter argument are evaluated. The absence of an indexed term is not treated as evidence that the book omits a concept. ### Stable anchor scheme | Anchor | Source location | |---|---| | `EPUB-FAQ` | `fm2.xhtml` | | `EPUB-FWD` | `foreword.xhtml` | | `EPUB-INT` | `introduction.xhtml` | | `EPUB-C01`–`EPUB-C13` | `chapter1.xhtml`–`chapter13.xhtml` | | `EPUB-APP` | `appendix.xhtml` | | `EPUB-META` | `package.opf` and `copyright.xhtml` | Section-level citations add the XHTML fragment, such as `EPUB-C03#sec3-2`. These anchors are stable within this source file but may not match another edition. ### Source and pattern index | Range | Source span | Main function | Representative anchors | |---|---|---|---| | `S-001`–`S-003` | FAQ, foreword, introduction | defines audience, project, method, and scope | `EPUB-FAQ`, `EPUB-FWD`, `EPUB-INT` | | `S-004`–`S-006` | Chapters 1–3 | intelligent-interface definition, AI material, Sentient Triangle | `EPUB-C01`–`EPUB-C03` | | `S-007`–`S-009` | Chapters 4–6 | tools, chat, NPCs, bespoke UI, canvases | `EPUB-C04`–`EPUB-C06` | | `S-010`–`S-012` | Chapters 7–9 | manager experience, agent roles, copilots | `EPUB-C07`–`EPUB-C09` | | `S-013`–`S-014` | Chapters 10–11 | failure modes and defensive design | `EPUB-C10`–`EPUB-C11` | | `S-015`–`S-016` | Chapters 12–13 | sprint, teams, roles, machine literacy | `EPUB-C12`–`EPUB-C13` | | `S-017` | Appendix | complete pattern inventory | `EPUB-APP` | | `S-018` | Package and accessibility metadata | edition and artifact audit | `EPUB-META` | ### Pattern-library inventory - **14 experience patterns:** Alchemists, Assistants, Bespoke UIs, Characters, Conductors, Copilots, Dedicated Tools, Inline Tools, Intelligent Canvases, NPCs, Sculptors, Stewards, Taskers, Workflow Tools. - **8 functional patterns:** Action Guard, Feedback Loop, Inpainting, Mission Staging, Personalization Daemon, Pinocchio, Spaghetti Scenarios, Version Vault. - **7 interaction patterns:** Canvas Command, Circle Gesture, Companion Dialogue, Inline Correction, Progressive Scaffolding, Prompt Refinement, Second Thought. - **17 UI patterns:** Activity Stream, Agent Notebook, Citation, Comparative Selection, Consensus Meter, Contextual Menu, Kanban Board, List-Detail, Mad Libs, Nudge Prompt, Parameter, Simple Rating, Structured Flags, Table List, Thermometer, Trait Tags, Workflow Canvas. ### Source limitations ranked by consequence 1. **No reader emphasis:** the source cannot establish what Courtney noticed, accepted, questioned, or intended to use. 2. **Mostly illustrative evidence:** product cases and agency experience establish possibility, not general effectiveness or causal outcomes. 3. **Fast drift:** named models, products, laws, security threats, and system capabilities can change within months. 4. **Security incompleteness:** design patterns do not enforce authorization, isolation, least privilege, or prompt-injection resistance. 5. **Artifact accessibility:** 178 of 180 images have empty alternatives despite conformance metadata. 6. **Sparse traceable research:** the extracted chapters contain few formal citations relative to their empirical and predictive claims. ## Prompt 1: reading contract, questions, priors, and drift ### Reading contract - **Why read:** 🔶 Decide which concepts deserve use in product and interface design. - **Hope to understand:** How AI capability becomes a human-system relationship, how agency changes UX, and how teams can design for failure. - **Practical use:** Pattern selection, design critique, sprint facilitation, prototype acceptance criteria, and governance conversations. - **Success criteria:** Retain a usable vocabulary, identify load-bearing claims, distinguish examples from evidence, and produce safe experiments. - **Evidence that would change the analysis:** Comparative usability studies, accessibility testing, security evaluations, longitudinal labor effects, and independent validation of the taxonomy or sprint. - **What the book cannot establish:** Universal product outcomes, legal compliance, model safety, accessibility conformance, job protection, or medical effectiveness. - **Constraints:** Local EPUB only; no target highlights; no supplied reader purpose or priors; current-source verification limited to high-consequence claims. ### Prior beliefs and incentives No reader beliefs, desires, resistance, or financial incentives were supplied. They remain unresolved. The analysis itself has two visible incentives: - A pattern language is attractive because it makes an unsettled field teachable. This can bias the review toward coherence. - A risk-focused audit can undervalue exploratory design. This can bias the review against speculative but generative frameworks. The correction is to judge the framework by its intended use. A design compass can be useful without being a validated taxonomy, while a safety or effectiveness claim still needs evidence. ### Open questions inventory #### Definitions and mechanisms 1. What observable criteria separate awareness from ordinary context handling? 2. When does an assistant become an agent, and when does an agent become a steward? 3. Can grounding, interoperability, and adaptivity vary independently? 4. What makes “individualized” context collection less invasive than personalization? 5. When does generated UI count as conversation rather than dynamic rendering? 6. Which pattern combinations are incompatible or unsafe? #### Evidence and credibility 7. Do independent designers classify the same products consistently? 8. Does posture-first design improve completion, trust, accessibility, or recovery? 9. Which named examples have measured user outcomes rather than vendor descriptions? 10. What populations and study designs support the medical effect sizes? 11. How often do critics, evaluators, RAG, and ensembles fail together? 12. What evidence supports the suggested tool-count limits for agents? #### Application and failure modes 13. What task stakes disqualify a radically adaptive interface? 14. How much review work can a manager experience impose before delegation loses value? 15. What happens when the user cannot evaluate the plan or result? 16. Which actions always require deterministic authorization and human approval? 17. How should a system recover from false memory or inferred context? 18. What limits prevent a “deferential” copilot from becoming a manipulative recommender? #### Historical and cultural context 19. Which ideas come from earlier adaptive interfaces, mixed-initiative interaction, ubiquitous computing, pattern languages, and agentive technology? 20. Which parts are new labels for existing HCI practices, and which depend on generative models? 21. How do commercial incentives shape the book’s emphasis on new experiences over non-AI solutions? #### DEIB, multicultural applicability, and accessibility 22. How do generated interfaces preserve stable navigation for cognitive and motor access? 23. How are accent, dialect, code-switching, and language variation tested? 24. Who consents when a continuous copilot observes a shared space? 25. Can preference manifests or earned-trust systems reproduce discrimination? 26. How are disabled users involved before posture and outcome selection? 27. What nonvisual equivalent communicates the Sentient Triangle and confidence cues? #### Unanswered or avoided 28. Who owns generated design and compensates source creators? 29. Who performs the correction and monitoring work after automation? 30. What environmental cost belongs in posture selection? 31. What organizational authority lets a designer stop an unsafe release? 32. When is the responsible recommendation to use no machine learning? ### Priors seed table These are system priors for the analysis, not Courtney’s beliefs. | Claim placeholder | Initial belief | Prior confidence | Reason | Evidence needed | |---|---|---:|---|---| | P-01 | Relationship posture affects usability | 75% | established HCI intuition | comparative task research | | P-02 | A four-posture vocabulary can improve team discussion | 70% | plausible coordination aid | facilitated-team studies | | P-03 | The triangle dimensions are true trade-offs | 35% | geometry overstates dependence | empirical dimension analysis | | P-04 | Constrained generated UI is safer than free-form UI | 75% | stable components reduce variance | accessibility and error testing | | P-05 | Manager UX is necessary for agent delegation | 85% | oversight follows agency | task-specific evaluation | | P-06 | More model agreement means higher confidence | 50% | correlation may be shared error | independence-aware calibration | | P-07 | Immediate context reduces need for profiles | 65% | plausible for bounded tasks | privacy and outcome comparison | | P-08 | The sprint improves AI product decisions | 55% | coherent but unvalidated | comparative project evidence | | P-09 | AI shifts design work upstream | 65% | observed task reallocation | longitudinal labor data | | P-10 | Adaptive UI can improve accessibility | 70% | standards and use cases support potential | disabled-user studies and conformance | | P-11 | RAG and structure can reduce unsupported output | 80% | supported technically | task-level failure rates | | P-12 | Pattern language alone can manage high-impact risk | 15% | controls must exist in architecture | security and governance evidence | ### Drift assumptions The book’s claims are fixed, but evaluative context will drift quickly. Recheck model capabilities, product examples, EU and US law, accessibility guidance, security attack patterns, medical evidence, labor effects, and energy use. The relational questions—control, consent, reviewability, reversibility, and responsibility—should drift more slowly. ## Prompt 2: inspectional reading and book map ### Central project - **Problem:** Product teams lack shared language for choosing and designing AI-powered interface relationships beyond generic chat. - **Answer:** Treat AI as a design material; map experiences through four postures and 14 patterns; design failure and oversight explicitly; use a five-phase sprint to move from outcomes to working prototypes. - **Method:** Practitioner synthesis, product examples, client work, original prototypes, pattern classification, analogies, and workshop exercises. - **Audience:** Designers, product leaders, and design-minded developers. - **Likely assumptions:** Readers accept AI-powered products as an active design space and can influence product behavior, governance, or delivery. - **Rhetorical modes:** Teach, explain, advocate, warn, synthesize, and sell a named practice. The authors concede uncertainty and failure, which tempers the advocacy. ### Chapter map and coverage | ID | Chapter | Purpose and central topic | Relation to argument | Evidence | Coverage/confidence | |---|---|---|---|---|---| | `EPUB-C01` | Intelligent Interfaces | Defines awareness, agency, casual intelligence, radical adaptation, individualization | establishes the practice and ethical posture | prototypes, client cases, product examples | read; high mapping confidence | | `EPUB-C02` | AI as Design Material | Explains ML, knowledge systems, algorithms, five capabilities, generation, grounding | supplies technical vocabulary | explanatory models and selected cases | read; high | | `EPUB-C03` | The Sentient Triangle | Introduces three qualities, four postures, and 14 patterns | organizing map | adapted framework and analogy | read; high | | `EPUB-C04` | Transactional Tools | Dedicated, workflow, inline, and sculptor patterns | grounded-to-iterative edge | product walkthroughs | read; high | | `EPUB-C05` | Chat and the Conversational Interface | Characters, assistants, prompts, repair, memory, multimodality | dialogue side of chat | product examples and cited studies | read; high | | `EPUB-C06` | Chat Is Not All Talk | NPCs, bespoke UI, intelligent canvases | extends conversation beyond dialogue | products and Sentient Scenes prototype | read; high | | `EPUB-C07` | Agents and the Manager Experience | Delegation lifecycle and oversight | establishes agent UX | analogy and selected agent products | read; high | | `EPUB-C08` | Agents of Change | Taskers, alchemists, stewards, conductors, system effects | agent taxonomy and operations | vendor cases and practitioner claims | read; high | | `EPUB-C09` | Copilots and Uncharted Territory | Ambient, proactive, deferential support at three context depths | completes the posture map | current product examples | read; high | | `EPUB-C10` | What Could Possibly Go Wrong? | Hallucination, bias, bad data, bad actors, outcome drift | diagnoses failure | public incidents, demonstrations, commentary | read; high | | `EPUB-C11` | Defensive Design | Signals, confidence, provenance, feedback, security, resilience | supplies mitigations | products, cited studies, author synthesis | read; high | | `EPUB-C12` | The Sentient Design Sprint | Outcomes, frictions, magic, postures, working prototype | converts taxonomy into method | exercises and medical example | read; high | | `EPUB-C13` | Work the Problem | Roles, common problem space, prototypes, deliverables, literacy | organizational conclusion | Apollo 13 analogy and workplace observation | read; high | ### Provisional summary *Sentient Design* argues that AI should be treated as a design material whose probabilistic and adaptive properties change the interface itself. Clark and Kindred organize intelligent experiences through a triangle of grounding, interoperability, and radical adaptivity. Four postures—tools, chat, agents, and copilots—describe how people relate to systems, while 14 experience patterns and a larger library of functional, interaction, and UI patterns make the map practical. The middle chapters show how familiar controls, generated interfaces, delegated agents, and ambient support can embody those relationships. The final third examines hallucination, bias, security, uncertainty, and oversight, then offers a sprint that moves from desired outcomes through frictions and imagined magic to a working prototype. The framework is strongest as vocabulary and facilitation. Its cases mostly illustrate possibility, and its implementation advice needs independent security, accessibility, governance, and outcome evidence. ### Reading hotspots - **Chapters 1–3:** thesis, definitions, and the taxonomy’s assumptions. - **Chapters 6–9:** generated UI and agent/copilot boundaries, where accessibility, consent, and oversight risks rise. - **Chapters 10–11:** strongest safety material and most testable recommendations. - **Chapter 12:** method, health claims, and the point where illustration becomes prescribed practice. - **Chapter 13:** labor and organizational predictions that require outside evidence. - **Appendix:** the highest-value reference artifact for reuse. ### Guiding questions for deeper use 1. What user relationship does the task require before any AI feature is selected? 2. Which triangle property is a real requirement, and which is decorative ambition? 3. What does the system know, infer, remember, and forget? 4. Which output must be deterministic or externally verified? 5. What permissions are necessary, and which are merely convenient? 6. Can the user evaluate the plan and result without redoing the work? 7. What happens when the system is wrong, uncertain, unavailable, or manipulated? 8. Which interface elements must remain stable across generated states? 9. How can a disabled user perceive, operate, and recover from the experience? 10. Who else is observed or affected by the system’s action? 11. What evidence would show that the pattern improves the intended outcome? 12. Who performs correction, monitoring, and appeal work after deployment? 13. What is the stop condition for the feature or experiment? 14. Would a non-AI tool solve the task with lower cost and risk? ### Reading strategy **Recommendation: full analytical read, followed by selective reference use.** The chapters build a coherent vocabulary, and later cautions repair some early enthusiasm. After the full pass, use Chapters 3, 7, 10–12, and the appendix as the working core. Revisit named products only after checking their current behavior. The recommendation would change to selective reading if the reader needs model engineering, legal compliance, or accessibility implementation rather than product-design vocabulary. ## Prompt 3: classification and terminology ### Classification - **Domain:** Human-computer interaction and product design. - **Subdomain:** AI-powered interaction design, generated UI, agent experience, and product discovery. - **Genre:** Practitioner framework, pattern catalog, and workshop handbook. - **Argument type:** Definitional, classificatory, prescriptive, and predictive. - **Primary method:** Pattern synthesis from agency practice, products, prototypes, public cases, and design analogy. - **Intended use:** Shared vocabulary, concept generation, critique, prototyping, and team alignment. - **Comparison traditions:** Pattern languages, adaptive interfaces, mixed-initiative interaction, ubiquitous computing, service design, agentive technology, defensive design, and design sprints. ### Controlled vocabulary | Term | Author’s definition | Competing or narrower definition | Operational definition here | Equivocation risk | Example / counterexample | Evidence | |---|---|---|---|---|---|---| | Sentient Design | practice of creating intelligent interfaces | “sentient” usually implies subjective awareness | interface design using context interpretation and bounded machine action | high anthropomorphism risk | adaptive report / static form | `EPUB-C01#sec1-2` | | intelligent interface | interface with awareness and agency | context-aware automation without human-like terms | system that senses inputs and changes content, action, or structure | “awareness” may mean simple state | adaptive questionnaire / fixed page | `EPUB-C01#sec1-1` | | awareness | perception of context and intent | sensed data or inferred state | declared set of inputs available to a decision | implies understanding | current selection / guessed emotion | C01 | | agency | ability to act toward goals | autonomy bounded by permissions | ability to select or execute actions outside a single fixed response | obscures who authorized action | scoped tasker / autocomplete | C01, C07 | | grounded | trained or supplied for reliable relevance | retrieval-backed, verified, or domain-bounded | output tied to identified current sources and task rules | conflates access with truth | cited RAG answer / fluent memory-only answer | C02, C03, C11 | | interoperable | works across systems and formats | standards-based compatibility | can read or act through defined external interfaces | tool breadth may increase risk | calendar-and-email agent / isolated editor | C03, C07 | | radically adaptive | conceived and delivered for the moment | dynamic or adaptive UI | changes content, structure, style, behavior, or medium at runtime | “radical” is promotional | bespoke UI / fixed template | C01, C03 | | posture | manner and relationship with the user | interaction mode | dominant control relationship: use, talk, delegate, or be supported | hybrids resist one label | agent / tool with automation | C03 | | casual intelligence | small quiet machine-assisted improvement | smart default or inline automation | bounded assistance inside a familiar interaction | may hide sensing | inferred form type / opaque automatic rewrite | C01 | | individualized | adapts from immediate context | personalization from a persistent profile | uses disclosed session inputs with limited retention | temporary inference can still surveil | selected-text adaptation / ad profile | C01 | | bespoke UI | generated layout or controls | server-driven or rules-based UI | approved components assembled from structured model output | may imply unconstrained generation | component schema / arbitrary HTML | C01, C06 | | character | role-play persona | conversational simulation | dialogue system optimized for performance, not authority | users may anthropomorphize | game NPC / medical advisor | C05 | | assistant | reactive task helper with limited tools | general conversational agent | scoped, user-directed helper with named capabilities | can silently gain agency | writing helper / autonomous spender | C05 | | agent | system that acts across tools toward a goal | any automated workflow | system selecting and executing multistep actions with external effects | vendors label automation “agent” | booking tasker / cron job | C07–C08 | | copilot | ambient context-aware sidekick | branded assistant | suggestion-first help adjacent to a primary activity | can become surveillance or interruption | spell-check / covert monitor | C09 | | manager experience | oversight UX for delegated work | observability dashboard | mission, plan, status, intervention, and result review | visibility may not equal control | editable plan / decorative progress bar | C07 | | productive humility | communication of uncertainty | calibrated confidence | evidence, limits, provenance, alternatives, and abstention suited to audience | vague disclaimers can simulate humility | cited uncertainty / blanket warning | C11 | | minimum magical product | working prototype of the essential “magic” | minimum viable product or prototype | smallest build that tests capability and behavior | magic can anchor solutions | prompt+mock data / polished concept video | C12 | ### Contested language - **Sentient, aware, understands, knows, wants:** human terms can hide statistical inference and institutional decisions. - **Grounded, reliable, confident:** different technical and user meanings must be separated. - **Copilot, assistant, teammate, customer:** relationship metaphors can obscure accountability and labor. - **Magic:** useful for imagination, risky when it hides data collection, human work, or failure. - **Personalization/individualization:** different retention strategies, not an ethical binary. - **Human in the loop:** can mean accountable review, rubber stamping, or unpaid correction. - **Explainability:** a useful user model is not faithful access to model reasoning. - **Consensus:** repeated or cross-model agreement is not independent evidence by default. ### Category-error watchlist | Risk | Status | Where | Effect | |---|---|---|---| | correlation treated as causation | present | product and bias examples | caps broad outcome claims | | anecdote treated as population evidence | present | client and vendor cases | cases show possibility only | | description treated as prescription | present | current patterns become recommended patterns | requires task testing | | technical capability treated as practical outcome | present | generated UI, agents, medical scribing | prototype success is not user benefit | | intention treated as effect | present | deferential/copilot language | system behavior may still manipulate | | individual behavior treated as institutional solution | present | “good citizen” agents | incentives and enforcement remain missing | | confidence treated as truth | explicitly rejected | Chapters 10–11 | book strength | | absence of evidence treated as evidence of absence | guarded but occasionally implied | future-of-work reassurance | labor risk remains live | | legal treated as moral | not dominant | governance discussion | current law still needs separate check | | taxonomy treated as measurement | risk present | Sentient Triangle | use as compass only | ## Prompt 4: claims inventory, evidence expectations, and dependencies ### Claims inventory | ID | Claim and source label | Type/scope | Testability and expected evidence | Falsifier | Source | Current evidence status | |---|---|---|---|---|---|---| | `C-001` | EA: Sentient Design creates context-aware, acting interfaces | definitional | operational examples and boundary criteria | category cannot exclude ordinary dynamic UI | C01 | N/A; branded definition | | `C-002` | EA: AI should be treated as a design material with a learnable grain | methodological | teams using material constraints make better decisions | no decision or outcome improvement | C02 | C; analogy plus practice | | `C-003` | EA: recommendation, prediction, classification, clustering, and generation cover core ML product capabilities | classificatory | capability coverage across systems | common capability fits none or multiple ambiguously | C02 | C; useful, not exhaustive | | `C-004` | EA: probabilistic output should be treated as signals, not facts | methodological | calibration and error-recovery studies | fact-like presentation performs safely in defined task | C02, C11 | B; strong general principle | | `C-005` | EA: four postures describe distinct user-system relationships | classificatory | reliable independent classification and design utility | low agreement or no practical distinction | C03 | C; coherent author model | | `C-006` | EA: triangle placement aids design exploration | practical | comparative workshops and decisions | adds confusion or false maturity ranking | C03 | C; explicit compass caveat | | `C-007` | EA: immediate individualization can outperform persistent personalization | evaluative | privacy and outcome comparison | context inference performs worse or collects equal risk | C01 | C; scenarios and client testimony | | `C-008` | EA: constrained components preserve coherence in generated UI | causal/practical | usability and accessibility testing | free-form UI matches or outperforms safely | C01, C06 | B; plausible architecture | | `C-009` | EA: chat belongs where exploration matters more than efficiency | prescriptive | task comparison | fixed UI outperforms on exploratory work or chat on routine work | C05 | B; sound conditional claim | | `C-010` | EA: specialist assistants are easier to control than universal ones | causal/practical | error, trust, and maintenance comparison | general assistant performs equally under same controls | C05 | B; supported by scope logic | | `C-011` | EA: nonverbal turn-taking can be conversational | interpretive | usefulness of expanded category | definition becomes unable to exclude ordinary interaction | C06 | N/A/C; productive but elastic | | `C-012` | EA: agent UX becomes manager UX | methodological | delegated-task studies | direct task UI remains sufficient under agency | C07 | B; strong relationship logic | | `C-013` | EA: manager experience needs mission, plan, monitoring, intervention, and result review | prescriptive | completion, workload, and failure-recovery studies | stages add cost without reducing failure | C07 | B; complete practical model | | `C-014` | EA: narrow agents with limited tools are safer and more effective | causal | security and task-performance data | broader tool access improves outcomes without added harm | C08 | B; current security support | | `C-015` | EA: persistent agents require performance oversight and system-health limits | prescriptive | incident and externality evidence | local task review is enough | C08 | B; public incidents support | | `C-016` | EA: contextual depth distinguishes situational, session, and continuous copilots | classificatory | reliable classification and consent implications | depth fails to predict design requirements | C09 | C/B; useful design dimension | | `C-017` | EA: hallucination, bias, malformed output, and bad actors are recurring failure surfaces | descriptive | incident data and evaluations | failures are rare or covered by another complete model | C10 | A/B; well established broadly | | `C-018` | EA: uncertainty, provenance, alternatives, correction, and risk-based review increase resilience | causal/practical | calibrated trust and outcome studies | cues confuse users or review fails systematically | C11 | B; components vary by audience | | `C-019` | EA: RAG, structured outputs, evaluators, and ensembles reduce overconfident errors | causal | task-specific controlled evaluation | equal error, new correlated failure, or security bypass | C11 | B with strong conditions | | `C-020` | EA: sprinting from outcomes to a working prototype improves AI product discovery | causal/practical | comparative project evidence | feature-first or other process performs as well | C12 | C; coherent but unvalidated | | `C-021` | EA: ambient scribes save over two hours daily and reduce burnout 20–60% | quantitative | controlled studies with defined populations | mixed or much smaller effects | C12 | D; current evidence is mixed | | `C-022` | EA: building is part of designing intelligent interfaces | methodological | prototype learning records | static simulation predicts behavior equally | C12 | B; strong for probabilistic systems | | `C-023` | EA: automation moves design value toward problem definition and judgment | predictive | longitudinal task and employment data | displaced work is not reallocated upstream | C13 | C; plausible, uneven distribution | | `C-024` | EA: the product and user outcome are the only final deliverables | evaluative | project outcomes and governance needs | documentation is itself required for access, safety, or continuity | C13 | C; needs qualification | | `C-025` | RI: pattern language must be paired with enforceable controls | methodological | security, governance, and accessibility audits | patterns alone reliably enforce constraints | analysis | B; **not the authors’ sentence** | ### Evidence-expectation pattern The book supplies abundant examples, moderate conceptual argument, and little comparative evaluation. Evidence is strongest for broad failure mechanisms and the need for scoped oversight. It is weaker for the triangle’s exact geometry, pattern completeness, workflow effectiveness, product outcome claims, and labor predictions. Independent support usually comes from sources outside the book rather than multiple passages inside it. ### Dependency graph ```mermaid flowchart TD A[AI has a distinctive probabilistic grain\nC-002 to C-004] --> B[Postures organize relationships\nC-005] B --> C[Patterns express the postures\nC-006 to C-016] C --> D[Failure surfaces must be designed\nC-017] D --> E[Defensive patterns and enforceable controls\nC-018, C-019, C-025] B --> F[Outcome-first sprint\nC-020, C-022] E --> F F --> G[Changed team roles and deliverables\nC-023, C-024] H[Triangle geometry is approximate] -.weakens.-> B I[Examples are mostly author-selected] -.limits.-> C J[Medical effect claim is weak\nC-021] -.does not support.-> F ``` - **Load-bearing:** `C-002`, `C-005`, `C-017`, `C-018`. - **Bridge claims:** `C-006`, `C-013`, `C-020`. - **Independent failure points:** `C-007`, `C-021`, `C-023`, and `C-024` can fail without destroying the pattern library. - **Single point of failure:** If posture categories do not improve reasoning, the Sentient Triangle loses much of its value, though individual patterns remain usable. ### Claim-to-question map - `C-005`–`C-006` → Questions 2, 7, 8. - `C-007`–`C-011` → Questions 3, 8, 22, 23. - `C-012`–`C-016` → Questions 6, 14–18, 24–25. - `C-017`–`C-019`, `C-025` → Questions 4, 7, 16, 27. - `C-020`–`C-022` → Questions 9–10, 26, 32. - `C-023`–`C-024` → Questions 19–21, 28–31. ## Prompt 5: quantitative rigor ### Update policy - Numerical updates encode analytical judgment unless a study supplies a measured effect. - Evidence from multiple passages by the same authors is one dependent source. - Vendor claims and selected examples may raise plausibility but rarely justify high confidence. - Current independent standards, incident records, and controlled studies receive more weight. - Nonempirical definitions and values receive `N/A`, not artificial probabilities. - When numbers would imply false precision, use: strongly decreases, moderately decreases, no meaningful update, moderately increases, strongly increases. ### Confidence bands | Label | Range | |---|---:| | Very low | 0–20% | | Low | 21–40% | | Uncertain | 41–60% | | Moderate | 61–75% | | High | 76–90% | | Very high | 91–100% | ### Priors and posteriors | Claim | Prior | Evidence considered | Dependence / likelihood impact | Posterior | Band | |---|---:|---|---|---:|---| | `C-002` material-grain framing helps | 70% | coherent capability/failure mapping | mostly author synthesis; moderate increase | 75% | Moderate | | `C-005` four postures are useful | 70% | full-book consistency and relationship verbs | no independent validation; slight increase | 74% | Moderate | | `C-006` triangle geometry guides decisions | 55% | author caveats and visual inspection | unvalidated placements; no meaningful update | 55% | Uncertain | | `C-008` constrained generated UI is safer | 75% | structured-output architecture, accessibility logic | task-dependent; moderate increase | 82% | High | | `C-013` manager lifecycle is necessary | 80% | agent examples and oversight logic | same author source; external security support | 86% | High | | `C-014` narrow tool scope reduces risk | 80% | book cases plus OWASP least-privilege guidance | independent conceptual support | 90% | High | | `C-018` defensive cues increase resilience | 75% | cited studies and cross-domain patterns | cues can backfire; slight increase | 80% | High | | `C-019` RAG/ensembles reduce errors | 75% | technical mechanism and current guidance | correlated sources and injection risk | 72% | Moderate | | `C-020` sprint improves discovery | 55% | coherent method and exercises | no comparative outcomes | 58% | Uncertain | | `C-021` medical effect size generalizes | 45% | observational improvement, mixed randomized results | population/product variation; strong decrease | 25% | Low | | `C-023` design work moves upstream | 65% | industry observation and ILO task exposure | distribution uncertain | 62% | Moderate | | `C-025` patterns need enforceable controls | 85% | OWASP, NIST, EU obligations, artifact audit | independent sources; moderate increase | 93% | Very high | ### Calibration warnings | Warning | Severity | Application | |---|---|---| | false precision | high | exact tool counts, confidence, effect sizes | | double-counted evidence | high | many examples share the authors’ selection and interpretation | | correlated sources | high | models, vendor cases, ensembles, and copied information | | unavailable denominators | high | how often patterns fail across all uses | | survivorship/selection bias | high | successful public products dominate examples | | publication and vendor bias | moderate-high | positive product claims are easier to find | | base-rate neglect | moderate | rare severe failures versus routine benefit | | proxy measures | high | trust cues, engagement, confidence, and “magic” | | subgroup averaging | high | accessibility, language, and labor effects | | practical versus statistical significance | high | medical and productivity claims | ## Prompt 6: analytical scaffold ### Chapter analysis slots | Chapter | Thesis / major claims | Evidence and examples | Counterexample or missing evidence | Relation to spine | |---|---|---|---|---| | 1 | interfaces can infer context and act; individualization; design as direction | prototypes and client cases | surveillance and minority inference errors | defines project | | 2 | AI material has five capabilities and probabilistic grain | product examples and technical explanation | taxonomy not exhaustive; terms sometimes anthropomorphic | technical base | | 3 | three qualities, four postures, 14 patterns | adapted Matt Webb diagram | no validation; axes not strict trade-offs | organizing spine | | 4 | tool patterns progress toward iterative collaboration | interface walkthroughs | no comparative accessibility/usability data | first posture family | | 5 | chat suits exploration; characters and assistants need scope and repair | products and cited studies | emotional, memory, and specialist claims need stronger data | dialogue chat | | 6 | conversation can be spatial, embodied, or generated UI | products and Sentient Scenes | unstable UI and accessibility risk | nonverbal chat | | 7 | delegated work requires manager experience | agent cases and analogy | review burden and reviewer qualification | oversight model | | 8 | agent roles vary by scope, transformation, persistence, and coordination | vendor cases and public incidents | no enforcement for system citizenship | agent taxonomy | | 9 | copilots vary by contextual depth | current product examples | consent model and boundary with stewards remain weak | closes taxonomy | | 10 | failure comes from context, content, format, and attackers | public cases and demonstrations | selected political examples overgeneralize | diagnosis | | 11 | uncertainty, provenance, feedback, layered review build resilience | products and cited research | cues, RAG, ensembles, humans can all fail | repair toolkit | | 12 | outcome-first sprint converts magic into tested prototypes | exercises and health case | participation arrives too late; medical claim weak | implementation method | | 13 | teams shift toward shared problems, judgment, prototypes, and literacy | Apollo analogy and workplace observation | labor displacement and documentation value understated | organizational conclusion | Every chapter offers either product limitations or “ask yourself” prompts, but few provide systematic counterexamples. Chapters 10 and 11 contain the strongest counter-observations in the authors’ favor. ### Argument spine 1. AI has properties that differ from deterministic interface materials. 2. Those properties enable interfaces to sense context and take bounded action. 3. Human-system relationships can be organized as tools, chat, agents, and copilots. 4. Experience, functional, interaction, and UI patterns make the relationships designable. 5. Probabilistic and agentic behavior creates recurring failures and power shifts. 6. Defensive patterns, oversight, and recovery must accompany the experience. 7. Teams can use an outcome-first sprint and working prototypes to select and test a posture. 8. Product roles then move toward problem framing, system direction, monitoring, and literacy. **Weakest link:** the move from a coherent classification to a claim that the classification improves product outcomes. Comparative team and user research would repair it. ### Evidence ledger | ID | Claims | Type/source | Independence | Quality | Limitation | |---|---|---|---|---|---| | `E-001` | C-001–C-006 | authors’ framework and Matt Webb lineage | low | C | conceptual, not validated | | `E-002` | C-007–C-011 | products, prototypes, client work | low | C | selected success cases | | `E-003` | C-012–C-016 | agent/copilot cases and public incidents | medium-low | B/C | limited task comparison | | `E-004` | C-017 | bias, hallucination, prompt-injection incidents | medium | B | mixed sources and versions | | `E-005` | C-018–C-019 | cited HCI studies and system examples | medium | B/C | summarized methods; correlated failure | | `E-006` | C-014, C-025 | OWASP current security guidance | independent | B | guidance, not one product evaluation | | `E-007` | C-025 | NIST AI RMF and EU AI Act | independent | B | governance frameworks, context-specific law | | `E-008` | C-021 | ambient-scribe observational study | independent | B | pre/post design, self-selection | | `E-009` | C-021 | randomized ambient-scribe trial | independent | B | product-specific, short duration | | `E-010` | C-023 | ILO task-exposure analysis | independent | B | exposure is not observed employment change | | `E-011` | accessibility claims | EPUB markup audit | independent direct artifact | A | artifact, not all formats | Six evidence groups are independent of the authors. Several qualify or tell against broad claims: mixed medical effects, unequal labor exposure, OWASP limits on RAG and model guardrails, and the EPUB’s accessibility failure. ### Reading plan and Prompt 7 entry conditions - Full source coverage is complete for all argument chapters and appendix. - Claims have stable IDs and section anchors. - Evidence expectations and non-vacuous falsifiers exist. - Independent evidence is separated from author-selected cases. - Quantitative claims use a limited judgment model rather than invented precision. - Attribution lanes distinguish EA, RI, and current outside evidence. Prompt 7 may proceed. Future reuse should prioritize Chapters 3, 7, 10–12, then consult the appendix by pattern name. ## Prompt 7: evaluate the argument ### Major-claim evaluation Grades describe the support available in this book and the checked outside sources, not the importance of the idea. | Claims | Evaluation | Grade | Confidence | What would improve the grade | |---|---|---:|---:|---| | `C-001`–`C-003`: sentient design is a distinct practice organized by awareness and agency | coherent stipulative framework; it names a real shift but does not prove a natural category | B- | high | comparative analysis against adaptive systems, automation, and HCI taxonomies | | `C-004`–`C-006`: grounded, interoperable, radically adaptive systems form a useful compass | memorable and generative; the axes are neither measured nor independent | B | high | operational definitions, inter-rater testing, outcome studies | | `C-007`–`C-011`: tools and chat patterns map distinct interaction needs | well illustrated and practically usable; evidence is selected rather than comparative | B | high | task-based studies across pattern choices and access modes | | `C-012`–`C-016`: agent and copilot postures require manager and oversight experiences | strongly reasoned and consistent with current security guidance | B+ | high | longitudinal field studies of review burden, intervention, and failure recovery | | `C-017`: intelligent systems fail across context, content, format, and adversarial surfaces | broad taxonomy fits the cases and current prompt-injection guidance | A- | high | frequencies, denominators, and domain-specific severity data | | `C-018`: uncertainty cues and provenance can support judgment | plausible and partly research-backed, but cues can mislead and calibrated uncertainty is hard | B- | medium-high | accessible cue studies with actual decision quality as the outcome | | `C-019`: grounding, ensembles, feedback, and review increase resilience | directionally sound, but each control has documented limits and correlated failures | B | high | defense-in-depth evaluations under realistic attacks and distribution shifts | | `C-020`: the five-phase sprint converts outcomes into a minimum magical product | internally coherent facilitation method; health example and outcome claims are under-tested | C+ | medium-high | replicated team studies, comparison groups, and participant research before ideation | | `C-021`: ambient scribes yield large time and burnout gains | supported for some settings, but the wording outruns the checked evidence | C | high | multi-site trials across products, specialties, access needs, and longer follow-up | | `C-022`–`C-023`: roles shift toward judgment, direction, and AI literacy more than displacement | credible for some knowledge work, but too optimistic as a general labor claim | C+ | medium-high | occupation-level longitudinal data and worker-centered studies across income groups | | `C-024`: working prototypes increasingly become the design artifact | strong observation for generative prototyping, weak as a universal replacement for durable specifications | B- | high | maintenance, audit, procurement, and handoff studies | | `C-025`: patterns must be backed by enforceable controls | strongly supported by current governance and security guidance | A- | high | product-level verification of permissions, logs, approvals, and rollback | ### Bias and fallacy scan | Risk | Where it appears | Effect | Repair | |---|---|---|---| | selection and survivorship bias | successful products, demos, and public failures | hides ordinary abandonment and quiet harm | publish sampling rules and negative cases | | availability bias | vivid incidents and polished interfaces | makes memorable examples feel representative | add rates and denominators | | novelty bias | generative UI and agent examples | understates older HCI and automation lineage | compare with adaptive interfaces, mixed-initiative systems, and decision support | | anthropomorphic framing | awareness, intention, character, judgment | can imply inner states where there are statistical inferences | name observable system behavior and data dependencies | | false dichotomy | deterministic interface versus probabilistic AI | obscures hybrids and uncertainty in conventional software | describe a continuum and mixed architectures | | optimistic labor inference | augmentation examples | generalizes from designers and knowledge workers | include bargaining power, job quality, displacement, and regional effects | | solutionism | sprint moves quickly from friction to magical solution | may frame social or organizational problems as interface problems | require a non-product option and a policy/process alternative | | automation bias by design | confidence, provenance, proactive assistance | cues can create trust without accuracy | measure reliance quality, not satisfaction alone | ### Counter-observations and repairs - A stable tool can beat a conversational or agentic system when the task is frequent, narrow, high-stakes, or screen-reader intensive. Repair: make the simplest posture the default comparator. - Personalization can magnify a wrong inference about identity, ability, or intent. Repair: expose the inference, allow correction, and provide an equivalent non-personalized path. - A manager experience can move work rather than remove it. Repair: budget review time and measure net cognitive load. - Human review is not a universal safeguard. Reviewers can be rushed, unqualified, or biased. Repair: define reviewer competence, escalation, sampling, and stop rules. - Prototypes can conceal requirements that matter later: audit trails, localization, procurement, security, accessibility, and maintenance. Repair: keep an executable acceptance record beside the prototype. ### Attribution audit - **EA — explicit author claim:** the book’s named postures, pattern families, sprint, and chapter arguments. - **RI — reviewer inference:** grades, dependency links, counterexamples, and repairs in this note. - **OE — outside evidence:** NIST, OWASP, EU, ILO, W3C, JAMA Network Open, and PubMed sources linked here. - Product behavior described only through the book remains an author-selected example unless checked independently. - The EPUB markup audit is direct artifact evidence, not a claim by Clark or Kindred. ### Prompt 7 verdict **Accept with conditions.** Use the book as a pattern language and critique kit. Do not treat the triangle as a measurement model, the pattern catalog as validated taxonomy, or selected product gains as expected effect sizes. Require task evidence, accessibility testing, enforceable controls, and recovery paths before adopting a pattern. ## Prompt 8: systems, context, and lineage ### System view ```mermaid flowchart LR P[People and communities] -->|goals, consent, correction| I[Intelligent interface] D[Data and context] --> I M[Models and rules] --> I I -->|recommend, generate, act| W[Workflow and world] W -->|effects and side effects| P I --> L[Logs, explanations, provenance] L --> O[Human and institutional oversight] O -->|permissions, limits, rollback| I R[Law, labor, market, environment] --> O R --> D ``` The interface is not the whole system. Data collection, model providers, tool permissions, organizational incentives, labor, energy, law, and maintenance shape what a user experiences. The book covers interface and team practice most fully; it treats procurement, political economy, environmental cost, and institutional accountability more lightly. ### Historical and intellectual lineage | Lineage | Connection | What the book changes | |---|---|---| | human-computer interaction and direct manipulation | tools, feedback, visibility, error recovery | applies these concerns to generated and probabilistic behavior | | adaptive and personalized interfaces | context-sensitive presentation | pushes from demographic segments toward session and individual adaptation | | mixed-initiative interaction | people and systems share control | names a spectrum from tool to copilot and asks where initiative sits | | recommender and decision-support systems | ranking, prediction, uncertainty | folds them into five AI capabilities and stresses signals over facts | | conversational agents and social computing | language, character, repair | separates character from assistant and chat from nonverbal conversation | | autonomous agents and workflow automation | delegation, orchestration, monitoring | frames the user experience as a manager lifecycle | | service design and systems thinking | effects beyond a screen | introduces system effects and “system citizenship,” but not an enforcement method | | safety engineering and human factors | defense in depth, handoff, interruption | translates these into uncertainty, provenance, feedback, and oversight patterns | The named “Sentient Triangle” is adapted from Matt Webb’s framing, according to the text. The larger synthesis belongs to Clark with Veronika Kindred. This note does not claim the framework originated without predecessors. ### 2026 context check - The book arrives as agent permissions and prompt injection become product-design concerns, not only model concerns. OWASP’s current guidance supports least privilege, constrained outputs, and human approval for high-impact actions; it also warns that retrieval and model guardrails do not eliminate injection risk. - The NIST AI Risk Management Framework supplies an institutional layer—govern, map, measure, manage—that the experience-pattern catalog does not provide by itself. - The EU AI Act makes risk classification, transparency, documentation, and oversight legally relevant for some deployments. Exact obligations depend on role, system, location, and effective date; this note is not legal advice. - The ILO’s exposure index reports broad but uneven occupational exposure and treats job transformation as the leading possibility, not a guarantee. That qualifies the book’s confident augmentation story. - W3C’s WAI-Adapt work shows that personalization can support access when semantics and user preferences are explicit. Inferred adaptation without stable semantics can work against that goal. ### Missing or thin perspectives - disabled people using screen readers, switch access, voice control, captions, magnification, or reduced-motion settings; - low-wage workers whose work is monitored, decomposed, or displaced rather than augmented; - Global South data labor, infrastructure, language coverage, and regulatory contexts; - children, elders, migrants, and people subject to institutional decisions; - Indigenous data sovereignty and community consent; - maintainers, security responders, procurement staff, records managers, and auditors; - energy, water, device, and data-center consequences; - people who refuse AI or cannot consent to continuous sensing. ### Historiographic caution This is a near-contemporary design book about fast-changing products. Its examples document what prominent teams demonstrated or reported, not a settled history of intelligent interfaces. Later editions should timestamp product behavior, preserve dead examples, record failed trials, and distinguish durable principles from 2024–2026 interface fashion. ## Prompt 9: DEIB, multicultural, and accessibility audit ### Overall finding The text takes bias, privacy, power, and human oversight seriously, especially in Chapters 10 and 11. Yet the pattern catalog rarely treats disability access, cultural variation, language power, or labor conditions as first-order pattern constraints. The EPUB’s own markup conflicts with its accessibility claim: 178 of 180 images have empty `alt` values, including diagrams and screenshots carrying instructional content. ### Audit matrix | Dimension | Strength | Risk or omission | Transfer condition | |---|---|---|---| | disability | multimodal interaction and human override appear throughout | visual diagrams, generated UI, gestures, color confidence, and changing layouts lack equivalent-access rules | stable semantics, keyboard flow, descriptive alternatives, zoom/reflow, reduced motion, assistive-tech tests | | race and ethnicity | bias and representational harm are acknowledged | product examples do not show affected communities shaping requirements | participatory research, subgroup error analysis, appeal and remedy | | gender and sexuality | stereotyping and identity inference are relevant to the framework | character/personality and personalization can encode binary or normative assumptions | user-defined identity, neutral defaults, correction, no sensitive inference without need and consent | | culture and language | chat and multimodality can widen expression | English-first examples and tone conventions may not transfer | local-language evaluation, dialect coverage, cultural review, translation provenance | | class and labor | organizational change and role shifts are discussed | attention centers on professional product teams and augmentation | worker voice, job-quality metrics, workload accounting, bargaining and transition plans | | geography | products operate across contexts | law, infrastructure, bandwidth, and data residency are not central | local legal review, offline/low-bandwidth routes, regional data controls | | age and cognition | proactive help could reduce burden | memory, pace, ambiguity, and persuasion risks vary by age and cognition | plain language, adjustable pace, repetition, confirmation, supporter modes without loss of agency | | religion and worldview | open-ended systems can accommodate varied goals | “helpfulness” may encode provider norms | state assumptions, allow value settings, preserve refusal and escalation | ### Artifact accessibility audit | Check | Result | Consequence | |---|---|---| | EPUB archive and spine | valid; 31 spine items | readable structure exists | | accessibility metadata | claims EPUB Accessibility 1.1, WCAG 2.2 Level A, and general AA alignment | creates a testable conformance claim | | image inventory | 180 images; all contain an `alt` attribute | syntactic presence alone is insufficient | | meaningful alternatives | 178 images use `alt=""`; only covers are descriptive | diagrams and screenshots are hidden from nonvisual readers | | visual confidence display | thermometer uses meter position and color | needs a text value, basis, and uncertainty explanation | | generated/adaptive UI | book encourages dynamic layouts | focus order, landmarks, names, states, and persistence can break | | gesture-rich sculpting | direct manipulation is emphasized | requires keyboard, voice, and programmatic alternatives | The metadata summary itself acknowledges missing non-text alternatives. The correct response is remediation, not treating the conformance label as proof. ### Reading-level and comprehension audit - The prose is generally clear for product and design professionals, but assumes familiarity with AI products, interface practice, and organizational workshops. - Pattern names help retrieval; rapid product-name sequences raise cognitive load and will date quickly. - Anthropomorphic terms such as “aware,” “intention,” and “judgment” need behavior-level definitions for technical and public audiences. - Confidence cues must state what is being estimated, by what method, and for whom. A percentage without calibration context can reduce comprehension. ### Required design changes 1. Add an accessibility criterion and failure mode to every pattern card. 2. Recruit affected people before posture selection, not only after prototyping. 3. Make adaptation inspectable, correctable, optional, and reversible. 4. Preserve a stable, non-generated route for high-frequency and high-stakes tasks. 5. Evaluate by subgroup and interaction method; do not hide differences in an average. 6. Provide appeal, remedy, and accountable ownership for harmful actions. 7. Treat labor, environmental, and community effects as product outcomes. 8. Replace empty alternatives in the EPUB with concise descriptions or extended text equivalents. ## Prompt 10: notes pipeline All items below stay in this one note. They are candidates, not separate notes or edits to a map of content. ### Fleeting notes - The triangle works because it invites questions, not because it computes an answer. - “Manager experience” is the sharpest reframing: delegation creates oversight work. - A proactive interface is a consent system disguised as a convenience system. - Confidence presentation is part of the claim, not decorative chrome. - Generated UI makes semantic stability a safety property. - The sprint needs a non-product branch before the magical-solution phase. - Pattern catalogs become governance only when permissions, tests, owners, and rollback are attached. ### Literature notes | ID | Source idea, paraphrased | Anchor | Status | |---|---|---|---| | `LN-01` | sentient design concerns interfaces that sense context and act with degrees of initiative | Ch. 1, “Awareness and Agency” | EA | | `LN-02` | AI capability can be grouped as recommendation, prediction, classification, clustering, and generation | Ch. 2, “Five Capabilities” | EA; non-exhaustive taxonomy | | `LN-03` | grounded, interoperable, and radically adaptive qualities serve as a compass | Ch. 3, “Sentient Triangle” | EA | | `LN-04` | tools, chat, agents, and copilots describe relationship postures, not fixed product boxes | Ch. 3, “Four Postures” | EA | | `LN-05` | open-ended exploration justifies chat better than routine efficiency does | Ch. 5, opening principle | EA | | `LN-06` | delegation shifts interface work toward planning, monitoring, interruption, and review | Ch. 7, “Manager Experience” | EA | | `LN-07` | proactive copilots require deeper context and greater consent safeguards | Ch. 9, contextual-depth spectrum | EA plus RI | | `LN-08` | failures cluster around wrong context, bad content, broken format, and bad actors | Ch. 10, failure taxonomy | EA | | `LN-09` | uncertainty, grounding, provenance, feedback, and oversight form layers, not guarantees | Ch. 11 | EA plus RI | | `LN-10` | an outcome-first sprint moves from friction to hypothesis, magic, posture, and prototype | Ch. 12 | EA | | `LN-11` | generative prototyping compresses the distance between brief and artifact | Ch. 13 | EA; scope contested | ### Permanent-note candidates Each candidate is `system-selected`; none reflects a user highlight or stated priority. | Suggested filename | Atomic claim | Scope and conditions | Confidence | Connections | |---|---|---|---:|---| | `AI pattern catalogs need control mappings.md` | An AI interaction pattern becomes governable only when mapped to permissions, evidence, owner, failure response, and rollback. | agentic or adaptive products with material consequences | high | NIST AI RMF; OWASP; `C-025` | | `Delegation creates a manager experience.md` | Automating task execution shifts effort into mission setting, plan review, monitoring, interruption, and outcome review. | multi-step delegated work; effect depends on review quality | high | human factors; agent UX; `LN-06` | | `Generated UI makes semantic stability a safety property.md` | When a model composes interfaces, stable programmatic names, roles, states, focus, and recovery are required for both access and control. | generated or adaptive UI | high | WCAG; WAI-Adapt; `LN-05` | | `Personalization is an inspectable inference.md` | A personalized experience should expose consequential inferences and let the person correct, disable, or reset them. | contextual adaptation using behavioral or sensitive data | high | privacy; disability; identity; `C-003` | | `Confidence displays are interface claims.md` | A confidence cue asserts something about calibration and should name its target, basis, limits, and accessible equivalent. | probabilistic recommendations and generated results | high | uncertainty communication; `C-018` | | `The simplest posture is the control condition.md` | Teams should compare every chat, agent, or copilot proposal with a stable tool and a non-product intervention. | posture selection and product discovery | high | progressive disclosure; solutionism; `C-020` | | `AI prototypes need an executable acceptance record.md` | A working prototype cannot replace durable requirements for security, accessibility, audit, localization, and maintenance. | regulated, shared, or long-lived products | high | documentation; procurement; `C-024` | ### Structure-note candidates - `Intelligent interface design` — link postures, capability families, failure surfaces, defensive patterns, and control mappings. - `Human oversight` — link manager experience, review burden, escalation, intervention, provenance, and remedy. - `Adaptive interface accessibility` — link generated UI, semantic stability, WAI-Adapt, personalization correction, and multimodal equivalence. - `AI product evidence` — link effect sizes, product-selected cases, subgroup analysis, deployment monitoring, and falsifiers. ### Attribution and excerpt audit The following short passages are `system-selected` from the local EPUB. They are included for source anchoring, not presented as user highlights: > “These systems deliver signals, not facts—suggestions, not answers.” — Chapter 2 > “the triangle is perhaps more compass than map” — Chapter 3 > “Delegation doesn’t make UX disappear; it changes its focus.” — Chapter 7 > “Intelligent interfaces should amplify human judgment, not replace it.” — Chapter 10 > “With intelligent interfaces, building is designing.” — Chapter 12 Do not attribute the claim grades, accessibility findings, systems diagram, note candidates, repairs, or current-source comparisons to the authors. No separate note, MOC edit, backlink edit, or Reader publication has been made. ## Prompt 11: teaching pack ### Audience, aim, and prerequisites **Audience:** product designers, product managers, engineers, researchers, content designers, accessibility specialists, and governance partners working on AI-assisted products. **Aim:** participants can choose and critique an interaction posture, identify its failure and access conditions, and connect an interface pattern to enforceable product controls. **Prerequisites:** basic product-development vocabulary; no model-building experience required. Provide the posture and pattern tables in accessible HTML or tagged document form before the session. ### Learning objectives By the end, participants can: 1. distinguish AI capabilities from interaction postures; 2. explain the three qualities of the Sentient Triangle without treating them as scores; 3. select a tool, chat, agent, or copilot posture for a bounded job; 4. map failure surfaces and access barriers before prototyping; 5. specify permissions, uncertainty, review, interruption, and recovery; 6. identify when a stable tool or non-product change is the better answer. ### Core vocabulary | Term | Working definition | |---|---| | capability | what the system computes: recommend, predict, classify, cluster, or generate | | posture | the relationship through which a person directs or receives system behavior | | grounded | connected to relevant, traceable context rather than only general model output | | interoperable | able to work across tools or services with controlled data and action boundaries | | radically adaptive | able to change content or interaction for a situation or individual | | manager experience | the planning, monitoring, intervention, and review layer for delegated work | | system effect | impact outside the immediate user-task exchange | | minimum magical product | the smallest working prototype that tests the core experience hypothesis | ### 90-minute lesson sequence | Time | Activity | Evidence of learning | |---:|---|---| | 0–10 | name a familiar AI interface and separate capability from posture | corrected two-column inventory | | 10–25 | teach triangle and four postures with boundary cases | participants explain why categories can overlap | | 25–40 | compare one task as tool, chat, agent, and copilot | posture trade-off sheet | | 40–55 | map wrong context, bad content, broken format, and bad actors | failure map with one falsifier each | | 55–70 | add disability access, consent, permissions, review, and recovery | control map with named owner | | 70–82 | prototype only the critical exchange | observable test script | | 82–90 | decide: build, simplify, research, or stop | written decision and evidence gap | ### Discussion questions - What does this system know, how does it know it, and how can a person correct it? - Where does initiative move, and can the person take it back? - Which users pay the review cost? - What must remain stable when content or layout adapts? - What harm can occur beyond the immediate user? - What evidence would make us reject this posture? ### Common misconceptions | Misconception | Correction | |---|---| | more agency means a better experience | agency increases permission, oversight, and recovery requirements | | chat is the natural UI for AI | chat suits open-ended exploration; structured tools often beat it for routine work | | grounding eliminates hallucination | grounding can reduce some errors but adds source, retrieval, and injection failures | | a human in the loop makes the system safe | the human needs time, skill, authority, evidence, and a workable intervention point | | personalization is automatically inclusive | wrong or hidden inferences can exclude people; equivalence and correction are required | | a working prototype is the full specification | long-lived systems still need acceptance, audit, and maintenance records | ### Case exercise **Case:** A church communications team wants an assistant that reads meeting notes, drafts announcements, schedules posts, and adapts tone for several communities. Teams must: 1. write the outcome without naming AI; 2. identify private data, community norms, and people who did not consent; 3. compare a template tool, scoped chat assistant, task agent, and proactive copilot; 4. choose the minimum posture; 5. specify accessible review, permissions, provenance, correction, rollback, and retention; 6. define a stop condition and a non-product alternative. ### Assessment rubric | Criterion | Meets expectation | |---|---| | framing | outcome and affected communities are explicit | | evidence | claim, falsifier, and measurement are paired | | posture | choice is compared with a simpler control | | safety | permissions, review, intervention, and recovery are testable | | access | keyboard, assistive technology, stable semantics, and equivalent path are specified | | systems view | labor, privacy, community, and downstream effects have owners | **Teaching caution:** do not use the EPUB diagrams without supplying text equivalents. Avoid color-only confidence exercises. Let participants contribute in speech, chat, document, or quiet written reflection, and do not require disclosure of disability or identity. ## Prompt 12: action plan and experiments ### PARA routing | Destination | Action | Trigger / owner | Proof of completion | |---|---|---|---| | Projects | add posture, failure, access, and control fields to the next AI feature brief | product lead at discovery start | completed brief and named evidence gaps | | Areas | maintain an AI pattern-to-control checklist | design system + security + accessibility | quarterly review against incidents and standards | | Resources | keep this one review as the source map | Courtney | verified file and source anchors | | Archives | retain the EPUB checksum and review date | archive process | checksum and provenance remain in note | ### Bounded experiments | Experiment | Hypothesis | Measure | Stop / decision rule | |---|---|---|---| | posture comparison | a structured tool completes a frequent task faster and with fewer corrections than chat | completion, errors, corrections, access barriers, preference | choose chat only if it improves exploratory outcomes without worsening critical errors or access | | inference inspector | showing and correcting context inferences reduces harmful personalization | correction rate, task success, trust calibration, subgroup differences | stop if sensitive inference is unnecessary or correction cannot fully propagate | | manager-load test | an agent reduces net effort after review and intervention | execution plus review time, missed errors, intervention rate | reject delegation if net effort or severe-error exposure rises | | confidence-cue test | accessible evidence and uncertainty wording improves calibrated reliance | correct acceptance/rejection, comprehension, screen-reader parity | remove cue if it raises reliance without decision accuracy | | generated-UI access test | constrained components preserve semantic stability across generated states | keyboard path, focus, names/roles/states, zoom/reflow, task completion | block release on loss of reachable controls, context, or recovery | | control mapping | every high-impact action has permission, log, approval, and rollback | unmapped actions and recovery drill results | no release while any material action lacks an enforceable owner and recovery path | ### Immediate next-use checklist ```mermaid flowchart TD A[State the human outcome] --> B{Can process or policy solve it?} B -->|yes| C[Test non-product change] B -->|no| D[Choose the simplest posture] D --> E[Map affected people and system effects] E --> F[Define evidence and falsifiers] F --> G[Add access, consent, permissions, and recovery] G --> H[Prototype the critical exchange] H --> I{Evidence meets decision rule?} I -->|no| J[Simplify, research, or stop] I -->|yes| K[Deploy narrowly and monitor] ``` ### Drift and review schedule - Recheck product examples and legal timelines before public teaching; they change quickly. - Recheck OWASP, NIST, W3C, and jurisdiction-specific rules before a deployment decision. - Retest assistive technology after any model, component, prompt, tool, or layout change. - Review failure logs by subgroup and interaction method; aggregate success can hide exclusion. - Retire pattern guidance when its assumptions, owner, or recovery mechanism no longer holds. **Most likely misuse:** treating the four postures and 14 experience patterns as a menu for making a product feel advanced. The safer use is diagnostic: select the minimum initiative and adaptation that the job justifies, then prove that people can understand, refuse, correct, interrupt, and recover. ## Prompt 13: visual synthesis and session close ### Visual 1: claim-to-decision chain ```mermaid flowchart LR O[Outcome] --> J[Job and affected people] J --> P[Posture] P --> F[Failure and access map] F --> C[Controls] C --> T[Test and falsifier] T --> D{Decision} D -->|pass| M[Monitor and repair] D -->|fail| R[Research, simplify, or stop] M --> J ``` Use this instead of treating prototype completion as proof. The return arrow makes deployment evidence part of continued discovery. ### Visual 2: posture-control matrix | Posture | Main value | Characteristic risk | Minimum controls | |---|---|---|---| | tool | speed and precision in bounded work | hidden generative variation | preview, versioning, undo, stable controls | | chat | open-ended exploration and repair | ambiguity, anthropomorphism, inaccessible history | scope, prompt help, memory controls, structured alternatives | | agent | delegated multi-step work | excessive agency and review burden | plan preview, least privilege, action guard, interrupt, logs, rollback | | copilot | contextual and proactive support | surveillance, manipulation, interruption | explicit sensing state, consent, boundaries, deferential timing, refusal, audit | ### Visual 3: defensive layers ```mermaid flowchart TB A[Clear scope and data limits] --> B[Constrained inputs and outputs] B --> C[Grounding and provenance] C --> D[Uncertainty and accessible explanation] D --> E[Least privilege and action approval] E --> F[Monitoring, interruption, and rollback] F --> G[Appeal, remedy, and institutional accountability] ``` No layer guarantees correctness. Layers reduce different failure modes; correlated model and data failures remain possible. ### Visual 4: access closure gate ```mermaid flowchart TD S[Generated or adaptive state] --> N{Names, roles, states, and focus stable?} N -->|no| X[Block and repair] N -->|yes| K{Keyboard, speech, touch, and assistive-tech paths equivalent?} K -->|no| X K -->|yes| V{Visual meaning has text equivalent?} V -->|no| X V -->|yes| U{User can inspect, correct, refuse, and reset adaptation?} U -->|no| X U -->|yes| T[Test with affected people] ``` ### Visuals intentionally omitted - A numeric Sentient Triangle score: the source does not validate measurement scales or trade-off geometry. - A maturity ladder from tool to copilot: greater initiative is not inherently more mature. - A single confidence thermometer: color and position can imply calibration that has not been established. - A universal “human in the loop” diagram: oversight roles differ in competence, authority, timing, and workload. - Decorative product screenshots: they would add visual volume without strengthening the argument. ### Prompts not run None. Prompts 0–13 are represented in this single file. The workflow’s state transitions were consolidated rather than split into adjacent artifacts. ### Revision candidates 1. Confirm the intended reader, purpose, domain, citation style, review cadence, and tolerance for public sharing; all remain 🔶. 2. If this note supports a real product, replace general claim grades with domain evidence and affected-person research. 3. If teaching from the book, create accessible text equivalents for any selected diagram rather than copying the EPUB image alone. 4. Recheck current legal obligations for the exact organization, jurisdiction, provider role, and deployment date. 5. Record genuine user highlights separately if supplied later; do not relabel these `system-selected` excerpts. ### Ten-point completion check | Check | Result | |---|---| | 1. one consolidated file | pass | | 2. EPUB integrity, metadata, spine, and checksum recorded | pass | | 3. all 13 chapters and appendix covered | pass | | 4. claims have stable IDs and falsifiers | pass | | 5. author claims, reviewer inference, and outside evidence separated | pass | | 6. quantitative claims calibrated without invented precision | pass | | 7. DEIB and artifact accessibility audited | pass with material source-artifact failure recorded | | 8. note candidates remain embedded and labeled `system-selected` | pass | | 9. no Reader publication or adjacent vault edits | pass | | 10. Prompts 0–13 and visual closure included | pass | ### Session status **Complete, with declared unknowns and source limitations.** This review supports critical use of the framework, not blanket endorsement. Its durable contribution is a shared vocabulary for initiative, adaptation, delegation, failure, and oversight. Its chief limit is the gap between a rich pattern catalog and comparative evidence that the patterns improve outcomes across people, domains, and access methods.