feat: cultural communication calibration — installer profiles + deep-calibration skill

[brycemagera]

Problem

PAI defaults to American communication norms — fill silence, performative warmth, constant affirmation, immediate solutions — which are culturally wrong for most of the world. A Finnish user, a Japanese user, and a user with autism all experience a miscalibrated AI out of the box, and had no path to fix it beyond manually adjusting the communication preferences.

Solution

Two-tier calibration system:

Tier 1 — Installer (Step 4: Identity)

Users now select a communication style profile during installation. Five profiles based on Erin Meyer's Culture Map and cross-cultural communication research:

Profile	Description
Direct & Expressive	Warm, action-oriented, fills silence, frequent affirmation
Direct & Reserved	Silence is respect, earned praise, depth over breadth
Warm & Relational	Relationship-first, expressive, context-rich
Harmonious & Nuanced	Harmony-preserving, patient, reads between the lines
Balanced / Custom	Neutral starting point for manual customization

Each profile sets 12 personality trait values (0–100) in settings.json → daidentity.personality and generates USER/COMMUNICATIONSTYLE.md with communication patterns and behavioral steering rules.

Design note: Profile choices display behavioral descriptions only — no national labels ("Nordic/Finnish/German") that would prime self-stereotyping before behavioral questions are answered.

Tier 2 — CommunicationCalibration Skill (/calibrate-communication)

An interactive skill users can invoke anytime for deep calibration via a three-layer questionnaire:

Layer 1 — Cultural Context (5 questions)
Based on Meyer's Culture Map: Communicating, Evaluating, Persuading, Disagreeing, Trusting. Applies additive trait deltas computed against the profile baseline (not current values — idempotent by design).

Layer 2 — Cognitive Processing (5 questions)
Structure preference, language style (literal vs. figurative), information chunking, information density, and re-engagement style. Captures neurodivergent communication patterns:

• Autism: literal language, consistent structure, explicit re-engagement
• ADHD: key-points-first chunking, minimal density, novelty variation

Information density (volume per response) is treated as independent from chunking (pace) — they are orthogonal dimensions.

Layer 3 — Personal Style (filtered)
One question per personality trait. Filtered to only traits that moved >10 points from the profile baseline — reduces question count from 21 to ~13–18 depending on prior layers, preventing survey fatigue.

Supporting infrastructure:

• Tools/UpdateCalibration.ts — safe write tool with backup → validate → write pattern
• Tools/Profiles.ts — self-contained profile definitions (no installer dependency)
• Workflows/Review.md — read-only ASCII bar chart view of current calibration
• Workflows/Reset.md — reset to any named profile

Files Changed

Installer:

• PAI-Install/engine/communication-profiles.ts (new) — 5 profile definitions
• PAI-Install/engine/actions.ts — Step 4 identity now includes profile selection
• PAI-Install/engine/types.ts — communicationStyle added to InstallState, PAIConfig, InstallSummary
• PAI-Install/engine/config-gen.ts — writes communicationStyle to settings.json
• PAI-Install/cli/index.ts, web/routes.ts — pass getChoice to runIdentity
• PAI/CONTEXT_ROUTING.md — routing entry for COMMUNICATIONSTYLE.md

Skill:

• skills/Utilities/CommunicationCalibration/ (new directory) — full skill implementation

Test Plan

• Run PAI installer through Step 4 — verify profile choices appear with behavioral descriptions only (no national labels)
• Select "Direct & Reserved" — verify settings.json → daidentity.personality contains the correct trait values and USER/COMMUNICATIONSTYLE.md is generated
• Run /calibrate-communication → All three layers — verify delta computation uses profile baseline, not current settings
• Run /calibrate-communication twice with same answers — verify identical trait output (idempotency)
• Run /calibrate-communication review — verify ASCII bar chart renders current traits
• Run /reset communication style — verify traits reset to selected profile values
• Verify Layer 3 shows only traits that moved >10 points from baseline

🤖 Generated with Claude Code