← Language From First Principles

The Interface Learns You

Build a user-controlled preference model from interaction evidence rather than a one-time questionnaire.

Chapter 21 gave the reader controls. Chapter 22 forbade turning their settings into a personality. Between those poles sits an unsolved accumulation problem: every interaction produces evidence — a correction here, a repeated choice there, a form that keeps performing better on one task — and with nowhere to accumulate, each encounter starts ignorant. Configuring every dimension manually is burdensome; silently inferring from clicks confuses habit and prominence with desire. This chapter learns from what actually happens, under Chapter 22’s evidentiary rule executed mechanically: no update without conditioning context, no type ever.

The unit of learning is not a profile but an observation:

PreferenceEvidence {
    dimension          comparison_representation
    context            exact numerical comparison
    observed_choice    table over prose
    outcome            faster completion, same accuracy
    confidence         medium
    source             explicit correction
    timestamp
}

An observation about one interaction — dimensioned, contextualised, sourced, dated — never an identity claim. Thousands of these accumulate; none of them says what kind of person the reader is.

The evidence hierarchy

Not all observations carry equal authority, and the chapter enforces the ranking in every update:

EXPLICIT CORRECTION     "Use a table for this."
EXPLICIT CHOICE         selects table from alternatives
REPEATED ADJUSTMENT     keeps changing generated views
TASK OUTCOME            one form repeatedly performs better
BEHAVIOURAL SIGNAL      dwell / click / abandonment

Corrections outrank choices, choices outrank repetition, outcomes inform separately, and behavioural signals sit at the bottom with explicit distrust:

Click ≠ preference. Dwell ≠ comprehension. Repetition ≠ stable trait.

Clicks measure prominence and habit; dwell measures confusion as readily as interest; repetition across contexts is the very generalisation the chapter refuses. Behavioural evidence is used — the machinery would starve without it — but always discounted, always contextualised, always overridable by a single explicit sentence. And evidence weakness now compounds with authority weakness: observable behaviour never authorises every inference constructible from it. Each learned rule therefore carries, besides its evidence tier, an authority lineage — purpose, evidence source, scope, allowed use, retention/expiry — so an explicit correction processed for the current response does not silently license durable preference inference, indefinite retention, cross-domain reuse, or external disclosure. No second runtime object is invented here; DataUseAuthority is referenced conceptually, with structural enforcement deferred to the Part IV repair pass. The continual-adaptation literature supports feedback-driven updating in bounded settings (Zhang et al., ACM Trans. Autonomous and Adaptive Systems 2025, DOI 10.1145/3727621, verified: complaints treated as evidence of changing preferences rather than static settings — application-specific, preference-oriented), and continual human-in-the-loop optimisation (ConBO, CHI 2025, DOI 10.1145/3706598.3713603, verified via ACM DL/arXiv:2503.05405: Bayesian-neural-network surrogate reusing population data to accelerate individual VR-keyboard optimisation) shows prior interaction data reducing relearning cost — fenced to narrow input parameters, never a general latent-person model. Both are precedents for updating from feedback, neither for knowing the user.

Two models that must never merge

The chapter’s structural defence against Chapter-22 regression is architectural separation:

PREFERENCE MODEL     What does the user tend to choose?
PERFORMANCE MODEL    Under what conditions did a representation
                     actually improve measured task outcome?

Rendering policy sees both, side by side, undisguised:

User preference:   diagram HIGH
Observed performance: diagram helps topology
                      diagram hurts qualification-heavy argument

No collapse into “prefers diagrams → always diagram.” The preference model drives satisfaction and reduces correction burden; the performance model gates safety and effectiveness; where they conflict, the system surfaces the conflict rather than resolving it silently (“you usually prefer diagrams, but prose transfers better on this material — [use prose] [keep diagram]”). That surfacing is Chapter 22 made executable: the lesson compiled into a runtime check.

Context on every update, decay on every memory

Two rules govern accumulation. First: no preference update without a conditioning context — task type, material structure, device, declared purpose, prior knowledge, representation attempted, preservation risk. The record says “prefers tables for exact comparison” and “wants conclusion-first when scanning,” never “prefers tables” or “conclusion-first person.” Context is what makes observations composable without becoming categorical.

Second: preferences change, so memory decays and contradicts honestly. Recent evidence outweighs old (Song et al.’s recency-weighted adjustment inference, verified for Chapter 21 via SIPLAB, reused here as the mechanism precedent); explicit corrections defeat stale implicit evidence at any age; conflicting evidence preserves uncertainty instead of averaging into false certainty (“table preference: high for comparison, uncertain for explanation, low confidence globally”). Drift cases are designed in: the detail-lover turned scanner, the task-specific split (tables for benchmarks, prose for argument), the prominence-noise user opening diagrams but keeping prose, the minimal-explanation preference that deletes prerequisites, the single-observation cold start, and the explicit “stop adapting this dimension” — which wins immediately, unconditionally, terminating learning on that axis until re-enabled. An override that negotiates is surveillance with manners; this one obeys.

Adaptation is therefore a negotiation with visible books:

Why did you present this as a diagram?
Because: you selected diagrams in 5 of 7 similar topology
tasks; performance was not worse; this source suits diagrams.
[Use prose instead] [Only for this task]
[Prefer prose for this kind of material]
[Stop learning from this behaviour]

Every bracket is a first-class event feeding the evidence store — including the last, which deletes rather than adds.

Learning policies, tested by correction burden

EXP-23 freezes Chapters 21–22 tasks and dimensions, then compares five learners: A static explicit policy (which doubles as the no-learning safe-default condition — its role made explicit: behaviour when learning is disabled entirely); B implicit behaviour only; C explicit corrections only; D explicit + implicit equally weighted; E evidence-weighted conditional learner (explicit > outcome > behaviour, with uncertainty and decay). The headline metric is correction economics — does the learner reduce necessary manual correction without increasing cross-context mistakes? — alongside preference prediction, task performance, preference/performance disagreement rates, policy reversals, recovery after deliberate preference change, false generalisation across tasks, trust/inspectability, plus false-generality counts, correction-burden totals, unauthorised-inference attempts, out-of-scope reuse incidents, and disabled-learning behaviour. The drift battery above is mandatory material. Failure criteria: E ties D (weighting machinery decorative); B ties E (behaviour suffices — the hierarchy unjustified); E generalises tables-for-benchmarks into tables-for-argument (context conditioning broken — the Chapter-22 violation, kept as the experiment’s cardinal sin); recovery after change is slow (decay miscalibrated). Artifacts: frozen task/dimension packs, learner traces with evidence stores, correction curves, cross-context error tables. What success would not justify: that the system learns who the user is — the conclusion stays bounded: interaction history can update conditional interface preferences, full stop.

Observations now span dimensions — comparison→table, topology→diagram, scanning→conclusion-first, study→source-first, mobile→brief, implementation→examples-first. The accumulation works. But the observations are scattered records, and reasoning over them needs a coherent object that still refuses to be a personality. These dimensions need a stable representation.

References

  • Song et al. UIST 2025 (SIPLAB-verified, Ch 21): recency-weighted adjustment inference reused as decay precedent; Pareto-selection half not re-litigated.
  • Zhang et al. (2025). Adaptive Preferences: Pivoting Through User Complaints. ACM T. Autonomous and Adaptive Systems. DOI 10.1145/3727621. Verified (DOI record): complaints as changing-preference evidence. Fenced: application-specific, preference-oriented.
  • ConBO (CHI 2025). DOI 10.1145/3706598.3713603; arXiv:2503.05405. Verified via ACM DL: BNN surrogate, population→individual VR-keyboard optimisation, reduced adaptation time. Fenced: narrow input parameters; no latent-person license.
  • Ch 21 dimensions/controls; Ch 22 taxonomy + evidential rule: obeyed as constraints, not re-argued.

Proposed experiment EXP-23: learner comparison with drift battery

Status: PROPOSED. Per the design above (A–E, frozen tasks/dimensions, six drift cases, correction-economics headline, cardinal-sin criterion).