← Language From First Principles

My Language Is Not Just a Modality

Define personal language as a multidimensional, task-dependent communication protocol.

Chapter 23 accumulates observations: comparison→table preferred, topology→diagram performs, scanning→conclusion-first chosen, study→source-first, mobile→brief, implementation→examples-first. Dozens of conditional records, each sound, jointly incoherent — no object reasons over them, conflicts resolve ad hoc, one correction ripples unpredictably. The question this chapter answers:

How do we represent many conditional communication preferences coherently without turning them into a personality model?

The answer is a shift the chapter title has been waiting for: from profile to policy. The user is not a bundle of traits. What the system holds is a program for communication:

PersonalCommunicationPolicy {
    defaults
    dimension_rules
    context_conditions
    exceptions
    evidence
    confidence
    precedence
    provenance
}

DEFAULT
evidence = source-first · depth = medium

RULE · WHEN task = exact_comparison · THEN representation = table
EVIDENCE · repeated explicit choice

RULE · WHEN material = topology · THEN representation = diagram
UNLESS preservation_risk = high

RULE · WHEN purpose = scan · THEN ordering = conclusion-first

EXCEPTION · WHEN qualification_density = high
THEN retain prose even if diagram is preferred

“Prefer diagrams” coexists with “qualification-heavy → prose,” “exact comparison → table,” “scan → conclusion-first,” “source dispute → source-first” — no contradiction, because these are scoped policies, not claims about a person. That scoping is what makes the object coherent: dimensions interact through conditional rules and precedence rather than sliding independently or collapsing into a type.

Cases → concepts → rules, without personality

Generalising from cases without generalising into personality is the methodological crux, and Policy Maps (Lam et al., Apple ML Research, UIST November 2025, verified: Policy Projector tool for practitioner-authored LLM behaviour policy; cases surveyed into regions/concepts navigated by if-then rules; evaluated with 12 AI safety experts on behaviours like gender assumptions and safety threats) supplies the pattern — observed cases abstracted into communication concepts (comparison, topology, scanning, deep study, source dispute, implementation) which author policy rules. The boundary is explicit: Policy Maps validates interactive policy-authoring for AI practitioners, not user-specific adaptation or this Personal Communication Policy. What transfers is the move — cases → concepts → rules with deliberate abstraction choices — applied here to a different subject (one person’s communication evidence) with different authorship (user corrections plus learned rules, not practitioner safety policy).

Every rule keeps its Chapter-23 evidence attached — support counts, performance records, confidence, last-confirmed dates — so the runtime treats a thrice-corrected table rule differently from a once-observed examples-first hint. Wozniak et al. (WWW Companion 2025, part-read; “up to 50%” figures still held for audit) supports explicit editable representation over history-only profiles, fenced to recommendation; Song et al. (UIST 2025, verified) supports multi-objective balancing with Pareto selection under learned priorities, fenced to layout/satisfaction. Neither validates this policy object. The policy earns itself in EXP-24.

Precedence: power without sovereignty

Conflicts are inevitable once dimensions coexist, so precedence is explicit and preservation supreme:

PRESERVATION / ACCESSIBILITY CONSTRAINT
        >
EXPLICIT CURRENT INSTRUCTION
        >
EXPLICIT SCOPED USER POLICY
        >
LEARNED HIGH-CONFIDENCE RULE
        >
LEARNED WEAK RULE
        >
GLOBAL DEFAULT

Learned concision yields to explicit show-every-caveat; diagram preference yields to qualification-destroying risk (prose, or prose-plus-diagram). The personal policy is powerful and never sovereign over truth or access — personalisation inside safety, the book’s standing rule, now executable as an ordering.

Profile describes; policy conditions action

The chapter’s sharpest section, stated as pairs:

PROFILE                          POLICY
"User likes diagrams."           "For topology tasks, prefer diagrams
                                  when they preserve required relations."
"User likes concise answers."    "For scanning, begin concise; expand
                                  automatically when preservation requires."
"User is technical."             "For programming tasks, formal terminology
                                  acceptable; unfamiliar domains get defined terms."

A profile describes a person. A policy conditions action on context with evidence, exceptions, and precedence. That is the conceptual contribution — and the “personal language” of the title is therefore a communication runtime policy: multidimensional (depth, form, ordering, evidence, example/principle balance, technicality, interaction, navigation — beyond Chapter 21’s list by interaction through rules), inspectable, revisable, evidence-bearing. Compilation must also reduce the person’s work, not relocate it: policy formation from scattered evidence earns its keep by cutting repetitive correction without demanding continuous curation — edited when useful, never administered daily. Locality therefore means two things at once: locality of correction (one fix touches the smallest justified scope) and locality of data use (evidence gathered for one scoped rule never silently becomes global personal data) — the HCI and privacy requirements in a single mechanism.

Crucially, current-mode inputs are refused here by design: task type, material, domain, and declared purpose condition rules; tired, busy, mobile-right-now, audio-only, focus-session do not yet exist in this object. Chapter 24 outputs PersonalCommunicationPolicy + task/material → presentation policy; Chapter 25 adds CURRENT MODE. The causal progression holds.

Testing the representation, not the evidence

EXP-24 freezes one EXP-23 interaction history and builds four representations of the same evidence — A flat global profile, B independent per-dimension preferences, C conditional rules, D evidence-bearing policy with precedence, confidence, and exceptions — then runs identical held-out cases measuring correct selection, cross-context false generalisation, conflict resolution, corrections, editability, explanation quality, preservation violations, recovery after single-rule change, and above all locality of correction: “stop prose for benchmark comparisons” must affect benchmark comparisons only, not turn every explanation into a table. Conflict battery required: global-vs-scoped, learned-vs-override, rule-vs-rule (comparison→table meets qualification-heavy→prose: table plus prose qualifications), stale rules, sparse scopes demanding abstention. Failure criteria: D ties C (evidence/precedence decorative); corrections propagate globally (locality broken — the latent-embedding disease); sparse scopes generalise confidently (abstention absent). Artifacts: frozen history, four representations, case outcomes, locality maps. What success would not justify: any current-state conditioning — situation-awareness belongs to EXP-25.

What this chapter earned

Scattered conditional observations become one executable, inspectable, evidence-bearing policy object that reasons across dimensions through scoped rules and precedence — personal language as communication runtime, personality nowhere. What remains is situation: the policy knows tasks and materials but not the person’s current mode.

Even a well-learned protocol becomes wrong when the same person is low on attention, mobile, deeply focused or merely browsing.

References

  • Lam, M.S. et al. (2025). Policy Maps. Apple ML Research / UIST Nov 2025. Verified: cases→concepts→rules authoring; 12-expert evaluation. Used for the generalisation pattern only; user-adaptation claims refused.
  • Wozniak et al. (2025). WWW Companion 2025. Part-read, numbers held. Used for editable-representation support, recommendation-fenced.
  • Song et al. UIST 2025 (verified Ch 21): multi-objective/Pareto balancing precedent, layout-fenced.
  • Ch 23 evidence objects/hierarchy: retained, not flattened.

Proposed experiment EXP-24: policy representations under one history

Status: PROPOSED. Per the design above (A–D, frozen history, locality headline, conflict battery, sparse-scope abstention). EXP-24 additionally records management actions required, unexpected rule propagation, unauthorised evidence reuse, and a generic/default comparison — curation cost measured alongside correctness.