← Language From First Principles

The Person Changes During the Day

Condition the interface on current mode and available attention rather than a static user profile.

The same engineer, the same paper, the same durable policy — twice in one day. Morning: deep study at a desk, derivation with sources and examples, evidence exposed early. Afternoon: phone on a train, quick scan between stops, conclusion plus decisive evidence plus expand. No preference changed between morning and afternoon. Nothing was learned or unlearned. The situation changed, and a policy that cannot see the situation serves both sessions badly — drowning the scan, starving the study, and worst of all learning the wrong lesson: “asked for very short answers on the train” must never become “permanently prefers very short answers.”

Chapter 24 built durable policy. This chapter overlays what is true right now, without rewriting it:

PersonalCommunicationPolicy      # slow-moving
        +
CurrentMode / Context            # temporary
        +
task + material
        ↓
EffectivePresentationPolicy      # now

The doctrine in one sentence:

Current mode should modify the execution of personal policy, not redefine the person.

Three contexts, unequal authority

The chapter admits three kinds of situational input and ranks them explicitly:

Declared mode — “I’m studying this,” “only scanning,” “audio only,” FOCUS/LEARNING/QUICK SCAN/DEEP STUDY/MOBILE/AUDIO ONLY/PRESENTING. First-class, authoritative, always obeyed within preservation bounds. LOW ATTENTION is permitted exactly here — declared, never detected.

Directly observable constraint — small screen, audio output available, current document and application, task phase. Usable directly; no psychology involved. But observable never means automatically retainable: each contextual input carries source, purpose, scope, and expiry inside the ModeOverlay, and the default handling is PROCESS-now-DO-NOT-RETAIN — device-is-mobile may shape this session without ever becoming a persisted mobile-behaviour profile, inferred attention preference, or cross-service device pattern unless retention is separately authorised. Temporary/durable stays operational, not rhetorical.

Inferred internal state — tired, stressed, distracted, confused, overloaded. Generally excluded. Behavioural proxies (pauses, typing speed, facial signals) do not enter conditioning no matter how convenient.

ProMemAssist (Pu et al., UIST 2025, verified for Chapter 19: working-memory modelling from multimodal signals, 12-participant study, selective assistance plus engagement over baseline) returns here as the honest counterpoint — research evidence that richer state inference is possible, and precisely not evidence that this architecture should use it. The fence holds both times. Just-In-Time Objectives (Lam et al., CHI 2026, DOI 10.1145/3772318.3790713; 66–86% reported preference wins with objectives as visible editable objects — numbers as reported, full read pending) fits with one clean cut: inferring a task objective (“revising this introduction to clarify its contribution”) is close to observable task evidence and adoptable under visibility; inferring a mental state (“tired and overloaded”) is sensitive, inferential, and refused when explicit modes or constraints suffice. JIT’s own design helps the book here — induced objectives as first-class editable objects, never hidden latent state. The rule:

Infer tasks only cautiously and visibly; do not infer internal psychological states when an explicit mode or observable constraint will do.

AdaptForge (Wickramathilaka et al., Automated Software Engineering 2025, DOI 10.1007/s10515-025-00547-z, verified citation-level) supplies the engineering precedent for conditioning adaptation on a structured user/platform/environment context object rather than a persona — fenced to accessibility/model-driven UI engineering, not understanding gains.

Ephemeral by construction

The Mode Overlay makes temporariness structural rather than aspirational:

ModeOverlay {
    mode · source · scope
    started_at · expires_at · overrides
}

MOBILE from observable device context, scoped to the session, expiring at session end: fewer simultaneous branches, compact comparison with expandable full table. DEEP_STUDY from explicit declaration, scoped to one paper: examples before compression, evidence early, longer derivations permitted. Expiry restores the durable policy untouched — the overlay deletes itself; nothing persists unless explicitly promoted, and promotion is always a question, never a silent merge (“source-first for 8 deep-study sessions — make it the default? [yes] [no] [don’t ask again]”). Temporary behaviour ≠ persistent preference, the sibling of Chapter 22’s preference ≠ performance.

Precedence extends Chapter 24 with mode inserted — and inferred mental state absent from the chain entirely:

PRESERVATION / ACCESSIBILITY > EXPLICIT CURRENT INSTRUCTION
> DECLARED CURRENT MODE > DIRECTLY OBSERVABLE CONTEXT
> STABLE SCOPED POLICY > WEAK LEARNED RULE > DEFAULT

QUICK SCAN plus qualification-heavy evidence still keeps the caveat — the gate outranks the mode. Inferred task objectives, where tried, enter as visible suggestions requiring correction, never as privileged facts.

Earned through conflict

Morning study versus afternoon scan (same person, same material, purpose changed); desktop table versus mobile ranked cards (device changed, comparison preference intact); AUDIO ONLY structured verbal traversal where diagrams are normally preferred (constraint, not learning-style claim); meeting-brevity quarantined from next-day study (no contamination); QUICK SCAN denied the caveat-removal it requests (preservation wins). Timing stays Chapter 19’s: mode may inform delivery policy downstream, but this chapter adapts representation and interaction of already-delivered objects — resolution within safe bounds, form, ordering, evidence visibility, navigation. Deliberately making interaction harder is refused here entirely; that is Chapter 26’s earned mechanism, and the handoff names it: modes can quiet, densify, or constrain the interface to present purpose — but sometimes the stated goal requires resisting the easiest action itself.

The session-shedding test

EXP-25 freezes the Chapter-24 policy, material, task, preservation constraints, and upstream Part-II decisions, varying only situational conditioning: A stable policy only; B plus explicit mode; C plus explicit mode and observable platform/task context; D plus visible editable inferred task objective; E hidden broad-state inference — included deliberately as adversarial stress, not proposed architecture. Measures: task performance, corrections, policy adherence, cross-session contamination, inappropriate persistence, preservation violations, preference satisfaction, mode-switch recovery, incorrect context inference. Headline: better fit now without persistent errors later — Session 1 (mobile + quick scan) must shed completely into Session 2 (desktop + deep study). Failure criteria: B ties A (modes decorative); E ties or beats D without correction burden (the exclusion unjustified — kept as live risk); Session-1 brevity persists into Session 2 (contamination — the cardinal sin); preservation violated under any mode (gate failure). Artifacts: frozen policy/material packs, overlay traces with expiry records, session-pair tables, contamination audits. What success would not justify: mental-state inference, deliberate friction, or any persistent trait claim — situation only, expiring by default.

What this chapter earned

Durable policy plus temporary inspectable modes plus task/material yields the effective policy of now — declared modes authoritative, observable constraints direct, internal states excluded, overlays expiring without residue, promotion always interrogative, precedence extended, timing untouched. What remains is the will: modes that do not merely fit the moment but defend the person’s longer goals against it.

Current modes let the interface become quieter, denser, more exploratory, or more constrained to the person’s present purpose. But sometimes merely changing the representation is not enough. Sometimes the user’s stated goal requires the interface to resist the easiest action.

References

  • Pu et al. UIST 2025 (verified Ch 19): counterpoint use — inference possible, refused here.
  • Lam et al. CHI 2026, DOI 10.1145/3772318.3790713 (numbers reported, read pending): task-objective inference adoptable visibly; mental-state inference refused.
  • Wickramathilaka et al. (2025). Automated Software Engineering, DOI 10.1007/s10515-025-00547-z. Verified citation-level: structured user/platform/environment context conditioning. Fenced: accessibility engineering, not understanding evidence.
  • Ch 24 policy + precedence; Ch 19 timing boundary: reused as constraints.

Proposed experiment EXP-25: situational conditioning with session shedding

Status: PROPOSED. Per the design above (A–E with adversarial E, frozen everything, session-pair shedding test, contamination as cardinal sin). EXP-25 additionally audits data flow: did temporary context survive expiry, update durable policy without promotion, or get reused outside the session — each a hard failure.