The Personal Experience

The Radar works. It decides what reaches the reader, how much of it, and when — every stage evidenced, every decision inspectable. And then it presents the result the same way to everyone: the same default layout, the same depth, the same explanation order, the same evidence posture. A retired engineer implementing from the digest and a student meeting the topic for the first time receive identically rendered information that Parts I and II proved deserves different treatment. The pipeline personalised everything except the encounter itself.

Part III changes the object being adapted — not by claiming the system knows how, but by making the recipient an explicit parameter of rendering:

The representation policy itself can become an explicit, user-controlled input to rendering.

information + task + explicit interface policy
        ↓
representation router
        ↓
personalised view

Nothing inferred. No stable preferences, no learning styles, no cognitive-state reading, no “personal language” taxonomy yet — Chapters 22 through 25 own those questions in order. This chapter introduces the recipient as a declared input, with dimensions rather than personalities, and with preference fenced from benefit from the first page. And a question the chapter must answer before any other: what happens if the person never configures anything? The answer is a conservative generic policy — the safe default: representation driven by source and task, never by inferred personhood; task-appropriate generic depth; evidence available and expandable; zero personal assumptions; learning off or minimally scoped until authorised. Not optimal — simply the non-personalised fallback that remains acceptably bounded when controls go untouched. Outcomes in this Part use four recurring families — TASK (accuracy/completion/time), PREFERENCE (declared choice/satisfaction), MANAGEMENT (configuration time, correction count, inspection burden, maintenance actions), SAFETY/AGENCY (unauthorised adaptation, contamination, non-use failure, override success) — with cognitive endpoints added separately only where measured, and no scalar Personalisation Score anywhere.

Dimensions, not personalities

The policy exposes presentation dimensions as independent controls:

depth          brief ↔ exhaustive
form           prose / table / diagram / timeline
explanation    examples ↔ principles
ordering       conclusion-first ↔ derivation-first
evidence       summary-first ↔ source-first
interaction    passive ↔ exploratory
technicality   intuitive ↔ formal
navigation     guided ↔ self-directed

Each is a rendering instruction, stored separately, combined per encounter — never fused into persona labels. There is no VISUAL LEARNER, TECHNICAL USER, or BEGINNER TYPE anywhere in this machinery, and the prohibition is structural rather than cosmetic: the system literally has nowhere to store a person-type, only per-dimension settings with per-task scope. That protects Chapter 22 by construction — the fixed-mapping shortcut cannot be taken because the representation for it does not exist.

Song et al. (UIST 2025, Song/Gebhardt/Liao/Holz, verified via SIPLAB project page) supplies the mechanism precedent that multi-objective adaptive UI is technically plausible: their mixed-reality system infers preference priorities from a small number of manual layout adjustments and uses them to select among Pareto-optimal layouts, with fewer manual adjustments at maintained satisfaction. The fences travel with the citation: mixed-reality layout problem, preference/satisfaction and adjustment burden as outcomes — not comprehension evidence, not cognitive-type evidence. The learning-from-adjustments half of their work is explicitly saved for Chapter 23; this chapter borrows only the plausibility of preference-driven adaptation among conflicting objectives. Wozniak et al. (WWW Companion 2025, part-read) contributes the user-controlled-representation precedent from recommendation — editable profiles beating history-only ones — fenced to its recommendation setting, not interface learning.

Preference is not benefit is not policy

The chapter’s most important distinction, stated as three objects with no automatic equality signs:

DECLARED PREFERENCE    What presentation do I want?
OBSERVED PERFORMANCE   What presentation helped on this task?
SYSTEM POLICY          What presentation will actually be used?

“I prefer diagrams” is honoured as a real input — the system renders diagrams where safe. It never becomes “diagrams make this person learn better” — that inference belongs to Chapter 22’s conditional-effect machinery, and this chapter’s experiment is designed so the two cannot be confused: preference outcomes and performance outcomes are measured and reported separately, and a preferred-but-worse-transfer result is a success of control paired with a warning for policy, not a contradiction to be smoothed over. The system policy that actually renders is a constrained optimisation over four inputs — user preference × representation affordance (Chapter 5) × task requirements × preservation constraints (Chapter 18’s gate) — and the interesting findings live in their conflicts.

Four conflict cases make the constraint structure concrete. The user asks for a diagram of qualification-heavy material: preference DIAGRAM against HIGH representation risk — retain prose, or diagram plus qualifications, never the bare diagram. The user wants extreme brevity where evidence requires caveats: preference VERY BRIEF against SUMMARY safe minimum — the preservation gate wins, and the refusal is explained. The user prefers source-first for rapid scanning: a legitimate preference that costs time, honoured with the cost shown. The user wants intuitive explanation of exact numerical comparison: table plus explanation rather than intuition-only prose — affordance constraining preference without overriding the person. The standing rule these cases establish:

Personalisation operates inside already-earned safety and preservation constraints. The Personal AI does not get to personalise away truth conditions.

The policy as a control surface

Part II’s explainability becomes direct control. The interface policy is inspectable, scoped, and reversible:

My interface — Depth: [concise —x— detailed]
Default evidence: [x] show sources  Comparisons: [x] prefer tables
Processes: [x] prefer diagrams when safe
Explanation: [x] example before formalism
Navigation: [x] let me explore
[Reset to generic] [Use only for this task]
Why wasn't my preference used?

That last control is load-bearing: when preservation or affordance overrules preference, the system says so with reasons, because an unexplained override converts a control surface into decoration. EXP-21 freezes the entire Part II stack (units, relevance, novelty, resolution, timing, preservation) and varies only presentation policy — A generic interface, B explicit user-configured, C configured plus task-preserving representation constraints — measuring policy adherence, profile preservation, completion, comprehension, time, overrides, preference, effort, plus configuration time, settings changed, and correction counts separately. If B beats A on preference, the chapter earns meaningful control of presentation — not performance improvement. If B/C produce no meaningful benefit over A once burden is accounted, the experiment licenses no further personalisation complexity — the generic default is permitted to win, and that result stands as legitimate. No learned adaptation enters; inference of any kind is refused here by design.

What this chapter earned

Rendering accepts the recipient as an explicit, dimensioned, user-controlled parameter — preference honoured, benefit unclaimed, types unrepresentable, safety constraints supreme. What it leaves entirely open is whether any stable mapping from person to best presentation exists at all — the tempting shortcut the next chapter must destroy before learning can begin.

Before learning a user, reject the tempting claim that every person has one fixed ideal interface.

References

• Song, Y. et al. (2025). Preference-Guided Multi-Objective UI Adaptation. Proc. UIST 2025. Verified via SIPLAB project page: MR layout, preference-from-minimal-adjustments → Pareto priority, fewer adjustments at maintained satisfaction. Used for multi-objective plausibility only; comprehension/type claims refused; learning half saved for Ch 23.
• Wozniak et al. (2025). WWW Companion 2025. Part-read, recommendation setting. Used for user-controlled-representation precedent only.
• Ch 05 affordance; Ch 18 gate; Part II stack (frozen for EXP-21): reused, not re-explained.

Proposed experiment EXP-21: presentation policy under full freeze

Status: PROPOSED. Freeze: Part II stack in full. Conditions A–C above with the four conflict cases required in the material set. Measures: adherence, profile fields, completion, comprehension, time, overrides, preference, effort — preference/performance separation enforced in analysis. Failure criteria: B ties A on preference (controls meaningless); C overrides silently (constraint without account); preferred-but-worse-transfer misread as policy failure rather than control success (analysis error). Artifacts: frozen stack, policy packs, conflict-case outcomes, separated preference/performance tables. What success would not justify: stable preferences, type mappings, performance benefits of preferred forms, cognitive-state claims — Chapters 22–25’s questions, explicitly not premises.