Kuanyin Domain — Features and Capabilities

Named after Guanyin (觀音, also spelled Kuan Yin or Kannon), the Bodhisattva of Compassion revered across East Asian Buddhist traditions, Kuanyin is the platform-wide ethics, safety, and moderation domain for the entire Oshun ecosystem. A Bodhisattva is a being who has made a vow to attain full Buddhahood for the benefit of all sentient beings, choosing to remain accessible to guide others toward liberation rather than entering complete nirvana — appropriate for a moderation system that guides users toward wholesome expression rather than simply excluding or punishing them.

The core philosophical commitment of Kuanyin is that every user carries innate potential for wisdom, compassion, and wholesome action — what Buddhist philosophy calls "Buddha nature" (佛性, fóxìng). Moderation should create conditions for that potential to flourish, not just defend against its suppression. This means non-punitive interventions, restorative pathways back to community, and recognition of positive contributions alongside detection of harmful ones.

Kuanyin provides 16 libraries covering predictive harm detection, mindful friction, community harmony, performer protection, rehabilitation pathways, merit and karma systems, dharma analytics, cross-domain ethics integration, AI/ML models, SDK/API, user interface components, and a restorative-circle safety pre-flight adapter. Kuanyin owns compassionate safety and restorative intervention; Themis owns governance and IP policy; the broader Concordia restorative-mediation flow remains planned.

At a Glance#

The table below maps each module to its primary purpose. Subsequent sections explain each in detail.

1. Foundation and Harm Classification (kuanyin-foundation)#

The foundation library is the conceptual root of the entire domain. It defines the shared vocabulary — harm categories, severity levels, intervention types, and the BuddhaNature user model — that every other Kuanyin library builds on.

1.1 Buddha Nature Framework#

Every user entity in the Kuanyin system carries a BuddhaNature interface — a data structure representing the user's modeled potential for wholesome engagement. This is not a moral judgment about who the user "really is" but a probabilistic model that shapes how the system responds to them.

• Innate goodness modeling — Numeric properties: compassion seed (tendency to consider others' feelings), wisdom factor (tendency to reflect before acting), mindfulness level (present-moment awareness in communication), karma balance (running ledger of positive and negative actions), innate goodness (baseline wholesome tendency), and awakening potential (responsiveness to reflection prompts).
• Refuge status tracking — Five-stage spiritual progress model reflecting the user's engagement journey:
  • seeking — new user, exploring the platform
  • taking_refuge — committed community member
  • established — consistent positive contributor
  • deepening — community mentor or leader
  • realized — platform leader with governance rights Stage progression unlocks privileges and reduces friction.
• Compassion-first moderation — All moderation actions are designed around the assumption that users have the capacity for wholesome action. The system asks "how can this user be guided toward better expression?" before asking "should this user be excluded?"
• Growth orientation — Intervention design prioritizes guidance toward wholesome expression over punishment of unwholesome behavior. The goal is behavioral change, not exclusion.

1.2 Harm Category Taxonomy (74 Categories)#

The harm taxonomy organizes 74 harm codes across 10 domains. The breadth of the taxonomy ensures that detection logic can be specific — distinguishing, for example, targeted harassment from mob harassment, or health misinformation from coordinated disinformation — rather than relying on coarse categories that produce false positives.

1.3 Severity Classification#

Six severity levels map harm categories to intervention types, ensuring that the volume of minor events does not consume resources reserved for serious harms. The HarmCategorySeverity union is negligible | low | moderate | high | severe | critical.

• Context-sensitive severity — Severity assessment considers not just content but intent, relationship between parties, audience, repetition history, and real-world impact potential.
• Escalation thresholds — Clear configurable thresholds for when automated responses must escalate to human review.
• Severity-based routing — Different intervention systems activate at each severity level, with higher levels bypassing low-priority queues.

1.4 Foundation Utilities#

• Platform constants — Standardized moderation constants ensuring consistent enforcement across all platform contexts and domains.
• Shared utilities — 21 pure analysis functions for content analysis, severity scoring, harm assessment, cascade prediction, and restorative-circle scheduling — reused across all Kuanyin libraries.
• Domain error types — Specialized TypeScript error types for moderation failures, classification errors, and intervention exceptions. KuanYinError base class with subsystem-specific subtypes including resilience errors (RateLimitError, CircuitBreakerOpenError, DegradationError).

2. Precognition — Predictive Harm Detection (kuanyin-precognition)#

Precognition is the predictive layer: it analyzes content and behavioral signals before harmful content is published, enabling intervention at the most beneficial moment — before harm lands, rather than after. Named after the Buddhist concept that wisdom involves anticipating consequences before acting.

2.1 Intent Analysis#

Intent analysis distinguishes what a message is trying to accomplish from its surface content. Sarcasm, reclaimed language, and cultural expression can appear harmful to surface-level pattern matching but carry benign intent, and the intent analysis layer is specifically designed to handle this distinction.

• Message intent classification — Analyzes the underlying intent of user messages before they are published, classifying across intents such as: informing, asking, expressing, celebrating, criticizing, confronting, harassing, deceiving, manipulating, or harming.
• Harmful intent detection — Identifies messages crafted to harass, deceive, manipulate, or harm a specific target or group.
• Benign intent recognition — Distinguishes genuinely harmful intent from sarcasm, jokes, reclaimed language, and cultural expression that might superficially resemble harm.
• Intent confidence scoring — Assigns confidence scores to intent classifications, ensuring that low-confidence classifications trigger human review rather than automated action.
• Multi-intent support — Recognizes when a single message contains mixed intents (partly helpful criticism, partly personal attack), enabling proportional response.

2.2 Emotional Detection#

Emotional state is a strong predictor of imminent harmful behavior. Detecting distress, anger escalation, and emotional contagion enables protective interventions before harm occurs rather than after.

• Distress signal recognition — Detects emotional distress in user content: despair, crisis, overwhelm, hopelessness. Triggers supportive resource surfacing (crisis hotlines, mental health resources) rather than punitive response.
• Anger escalation detection — Identifies escalating anger patterns in sequential messages before they result in harmful outbursts. Triggers cooling-off friction at the optimal moment.
• Emotional state modeling — Builds a real-time model of each user's emotional state from content and behavioral signals across a session.
• Emotional contagion tracking — Monitors when negative emotional states spread through community interactions — a phenomenon well-documented in online communities where one distressed user can destabilize many others.
• Supportive intervention triggers — Automatically surfaces supportive resources (peer support communities, professional resources, crisis lines) when distress signals are detected.

2.3 Typing Dynamics Analysis#

Typing behavior patterns reveal emotional state in ways that message content alone cannot. Frustration, rage, and anxiety all have characteristic typing signatures that the precognition layer monitors in real time.

• Typing speed monitoring — Analyzes changes in typing speed that may indicate emotional escalation (rapid burst typing) or deliberate preparation of a harmful message.
• Deletion pattern detection — Detects excessive deletion and rewriting as a signal of frustration or self-censoring struggle — often indicating a user is oscillating between what they want to say and what they know they should say.
• Draft analysis — Analyzes drafted-but-not-yet-sent content for harmful patterns before the user decides to send or discard.
• Cooling-off indicators — Identifies when typing dynamics suggest the user is calming down versus continuing to escalate, enabling adaptive intervention timing.

2.4 Behavioral Pattern Tracking#

Individual incidents are less predictive than trajectories. Kuanyin tracks behavioral patterns over time to identify concerning trends before they cross harm thresholds, which is especially important for detecting gradual escalation that would be invisible in any single interaction.

• Historical behavior profiling — Builds behavioral profiles from each user's interaction history to identify concerning trends that wouldn't be visible from any single interaction.
• Pattern deviation alerts — Alerts when behavior deviates significantly from a user's established baseline — useful for detecting account compromise and sudden behavioral deterioration.
• Escalation trajectories — Identifies users on escalating trajectories (increasing severity of violations over time) before they cross into high-severity harm categories.
• Positive behavior recognition — Tracks positive behavioral patterns (helpful responses, welcome contributions, community support) to reinforce constructive engagement and feed the merit system.
• Recidivism prediction — Predicts the probability of repeat violations based on violation history, intervention response, and behavioral trajectory.

2.5 Context Awareness#

The same words carry entirely different meaning in different contexts. A joke among close friends, a community sharing reclaimed language, and a stranger targeting a vulnerable person require entirely different responses — context awareness is what makes the difference between fair moderation and blunt over-filtering.

• Conversation context — Understands messages within the full context of the ongoing conversation, not as isolated utterances.
• Relationship context — Considers the established relationship between the parties: strangers, acquaintances, friends, romantic partners, community members.
• Community context — Factors in community norms, culture, and in-group language specific to each community space.
• Temporal context — Considers time of day, ongoing community events, and situational factors (e.g., community in distress after an incident).
• Platform context — Adapts detection thresholds to the specific platform context: adult content platform, children's educational platform, professional network, and public news comments all require different sensitivity settings.

2.6 Cascade Prediction#

Some harmful messages are self-contained; others trigger broader community harm through virality, pile-ons, or misinformation spread. Cascade prediction specifically targets the latter — identifying messages likely to amplify into community-wide harm before they do.

• Viral harm prediction — Predicts when a single harmful message could cascade into broader community harm through sharing, quote-posting, or community amplification.
• Pile-on detection — Detects early signs of coordinated pile-on behavior against a single user: multiple different users targeting the same person within a short time window.
• Misinformation spread modeling — Models how false information spreads through community networks based on the structure of the network and historical spread patterns.
• Intervention timing optimization — Determines the optimal intervention moment for maximum harm reduction with minimum disruption to legitimate discourse.
• Cascade severity estimation — Estimates the expected severity and reach of a harmful cascade before it fully develops, enabling proportional resource allocation for the response.

3. Mindful Friction (kuanyin-mindful-friction)#

Mindful friction is the cornerstone of Kuanyin's non-punitive intervention philosophy. Rather than blocking content or immediately escalating to punitive action, mindful friction creates a brief space for reflection that allows users to reconsider. The friction slows harmful actions without blocking expression — users can always proceed after the reflection period. This respects user autonomy while nudging toward more wholesome communication.

3.1 Pause and Breathe#

The pause-and-breathe intervention is the most lightweight friction mechanism: a momentary interruption that creates space without making any accusation.

• Pre-posting pause — A brief, configurable pause before potentially harmful content is published. The pause is presented as a moment for reflection, not a punishment or warning.
• Breathing exercise integration — Optional short guided breathing exercise during the pause period, reducing arousal and creating a more reflective mental state.
• Cool-down timers — Configurable cool-down periods proportional to detected emotional escalation level. An angry burst gets a longer pause than mild frustration.
• Voluntary reflection — Pause framed explicitly as an invitation to reflect: "You're about to send a message. Take a moment if you'd like." Not as an accusation or warning.
• Bypass respect — Users can always proceed immediately after the pause. Friction slows but never blocks expression.

3.2 Samma Vaca (Right Speech)#

Samma Vaca (正語, sammā-vāca) is the Pali term for Right Speech — the third element of the Noble Eightfold Path in Buddhism. It asks whether speech is true, helpful, kind, and timely. Kuanyin presents these as gentle reflection questions, not as gatekeeping criteria. The system never asserts that the content fails these tests — it invites the user to consider them.

All prompts are presented as genuine invitations to reflection. The system does not assert that the content fails these criteria — it asks the user to consider them.

3.3 Cognitive Reframing#

Cognitive reframing helps users recognize when their messages stem from cognitive distortions — habitual thought patterns that misrepresent reality in ways that often lead to conflict. This approach draws directly from cognitive-behavioral therapy (CBT) techniques adapted for an online context.

• Automatic thought identification — Helps users recognize when messages stem from cognitive distortions common in online conflict situations.
• Distortion labeling — Gently identifies common cognitive distortions: all-or-nothing thinking ("you ALWAYS do this"), catastrophizing ("this is the worst thing ever"), personalizing (assuming everything is about you), mind reading (assuming you know others' motives).
• Reframing suggestions — Offers alternative interpretations of the situation that may be equally or more accurate and less distressing.
• Perspective broadening — Encourages consideration of multiple perspectives on the situation before responding.

3.4 Perspective Shift#

Perspective-shift interventions interrupt the first-person tunnel vision that often drives conflict, asking the user to imaginatively occupy other viewpoints.

• Empathy prompts — "How might the other person feel reading this?" — prompting users to imaginatively occupy the recipient's perspective.
• Role reversal exercises — Brief structured exercises where the user considers the situation from the other party's position.
• Impact previews — Shows predicted emotional impact of the message before it is sent, making consequences visible before they occur.
• Third-party observer framing — "How would a neutral observer view this exchange?" — shifting from participant to observer perspective.

3.5 Compassion Nudges#

Compassion nudges are the lightest-touch interventions: brief reminders that humanize the other party and encourage charitable interpretation before responding.

• Kindness reminders — Brief reminders of the human on the other side of the screen: "There's a real person who will read this."
• Positive intent assumption — Encourages users to consider the most charitable interpretation of the other party's behavior before responding.
• Common ground highlighting — Surfaces what the conflicting parties have in common (shared community membership, similar values) during heated exchanges.
• Gratitude redirects — In some contexts, redirects energy from conflict toward appreciation of positive aspects of the relationship or community.

3.6 Alternative Expression#

Alternative expression goes beyond friction to offer constructive help: concrete rewrites and tone adjustments that let users express their genuine concern without the harmful framing.

• Rewrite suggestions — Alternative phrasings that express the same underlying sentiment or concern without harmful content. For example: "This is stupid" → "I disagree with this approach because..."
• Tone adjustment — Tone-shifted versions of the same message: aggressive → assertive, passive-aggressive → direct, dismissive → critical.
• Constructive criticism framing — Helps users express valid criticism constructively: separating the critique from the person, focusing on specific behavior rather than character.
• Frustration channels — Suggests appropriate outlets for genuine frustration that don't involve targeting another person.

4. Community Harmony (kuanyin-community-harmony)#

Community harmony systems operate at the collective level — monitoring the emotional health of entire community spaces rather than individual users. Healthy communities are resilient to individual conflicts; unhealthy communities amplify them. These systems exist because individual-level moderation alone cannot address collective dynamics like raids, pile-ons, or community-wide emotional contagion.

4.1 Community Temperature Monitor#

The temperature monitor aggregates emotional signal across all users in a space, giving moderators a real-time sense of community health that individual-message moderation cannot provide.

• Real-time emotional temperature tracking — Aggregated emotional state monitoring across all activity in a community space. Rising temperature indicates increasing conflict and distress across multiple users simultaneously.
• Heat mapping — Spatial heat maps showing which areas of a community (topic threads, live channels, comment sections) have elevated emotional temperature.
• Community-wide trending alerts — Alerts community moderators when temperature rises above configured thresholds, enabling proactive intervention before individual incidents escalate.

4.2 Conflict Detection#

Conflict detection focuses on emerging interpersonal and group tensions — the early warning signals that precede harassment incidents or community-wide disruption.

• Emerging conflict detection — Identifies brewing conflicts between specific users before they escalate to harmful behavior, enabling moderator awareness and optional early intervention.
• Interpersonal tension mapping — Maps tension patterns across user relationships within the community to identify recurring conflict dyads or clusters.
• Escalation prediction — Predicts which tensions are likely to escalate to harmful behavior based on historical escalation patterns in similar community contexts.
• Early intervention routing — Routes detected conflicts to community moderators or appropriate resolution resources with sufficient context to enable informed intervention.

4.3 Raid Defense#

A raid is a coordinated external attack: a group organized in one community (typically a hostile or adversarial space) simultaneously floods a target community with harmful content, harassment, or disruption. Raids are qualitatively different from individual bad actors and require community-level defenses.

• Coordinated attack detection — Detects sudden influx of coordinated harassment from accounts with similar behavioral fingerprints or external coordination signals.
• Rapid response activation — Automatically activates community protection measures: temporary slowmode, new account throttling, keyword filtering, and moderator alerts.
• Attacker pattern fingerprinting — Fingerprints the behavioral patterns of the attacking group for faster detection of subsequent raid attempts from the same community.
• Post-raid community recovery — Guides the affected community through emotional recovery after a raid: acknowledging the incident, supporting affected members, and restoring normal community culture.

4.4 Post-Incident Healing#

After a serious incident — a raid, a viral harassment campaign, or a high-profile conflict — the community itself needs support, not just the individuals directly involved. Post-incident healing systems structure that recovery.

• Facilitated community healing conversations — Structured community-wide conversations after harmful incidents, facilitated to acknowledge impact and restore trust.
• Affected member support resources — Surfaces appropriate support resources to community members most affected by incidents.
• Community-wide acknowledgment — Supports community leaders in acknowledging harmful incidents transparently to the community without amplifying harmful content.
• Trust restoration pathways — Structured approaches to restoring community trust after incidents that shook it.

4.5 Culture Cultivation#

Culture cultivation is proactive rather than reactive: it reinforces positive norms during normal operation, not just during incidents. Communities with strong positive cultures are more resilient when conflicts arise.

• Community norms reinforcement — Proactively highlights and reinforces positive community culture norms during normal operation, not just during incidents.
• Positive pattern amplification — Amplifies and acknowledges constructive interactions and bridge-building behavior when detected.
• Community value surfacing — Surfaces shared community values during conflicts, reminding participants of what they have in common.
• Onboarding norm integration — Weaves community norms into the new member onboarding experience so new members understand the culture before encountering their first conflict.

5. Performer Protection (kuanyin-performer-protection)#

Performers, streamers, and creators face distinctive moderation challenges: they are public-facing, they interact with large numbers of strangers simultaneously, and they are subject to coordinated campaigns by hostile audiences. Standard community moderation is insufficient for their needs — they require specialized, performer-controlled protection systems.

5.1 Real-Time Shield#

• Real-time content and chat filtering — Live filtering of content and chat messages for live streams and performer-facing channels. Filters are applied before content reaches the performer, not just flagged after. Configurable sensitivity per performer.
• Configurable sensitivity thresholds — Performers configure their own protection sensitivity, trading off between over-filtering (blocking legitimate fans) and under-filtering (exposing themselves to harm).
• Word, phrase, and pattern blocking — Performer-defined and system-suggested word, phrase, and behavioral pattern blocking with regular expression support.

5.2 Parasocial Detection#

Parasocial relationships are one-sided emotional attachments where an audience member feels a personal connection with a performer who does not know them personally. While healthy parasocial appreciation is normal and expected in creator communities, pathological attachment can lead to boundary violations, stalking, and serious harm. The parasocial detection system identifies the transition from healthy fandom to concerning behavior.

• Unhealthy attachment detection — Identifies patterns indicating unhealthy parasocial attachment: excessive gifting, repeated attempts to establish personal contact, expressed belief in a personal relationship that does not exist.
• Boundary violation tracking — Tracks repeated attempts to violate performer-defined interaction boundaries across sessions.
• Obsessive behavior indicators — Flags escalating patterns of obsessive attention: following the performer across platforms, tracking their schedule, expressing entitlement over their behavior.
• Graduated response — Escalating interventions proportional to severity and trajectory: soft nudge (gentle redirect) → content filtering → shadow throttle → hard block.

5.3 Boundary Enforcement#

• Performer-defined rules — Allows performers to define and manage their own interaction boundaries in plain language without requiring moderation expertise.
• Boundary violation prevention — Prevents messages that violate performer-defined rules from reaching the performer before delivery, not just flagging them after.
• Topic and content rules — Blocks specific topics, request types, or content categories per the performer's stated preferences. A performer may block discussion of their personal life, certain request types, or references to past content they no longer want discussed.

5.4 NCII and Deepfake Protection#

NCII (Non-Consensual Intimate Imagery) is the distribution of intimate images of a person without their consent, a form of sexual abuse. Deepfakes create fake NCII using AI by inserting a person's face into intimate content they did not participate in. Both are serious harms with specific legal frameworks, and Kuanyin provides specialized detection and rapid response workflows for them.

• NCII detection — Non-consensual intimate imagery detection using image analysis models trained to identify intimate imagery and match it against performer-protected identities.
• Deepfake detection — AI-generated synthetic media detection, particularly face-swapped content replacing a performer's face in intimate material. Uses facial inconsistency detection and generation artifact recognition.
• Streamlined report and removal — Streamlined NCII/deepfake report and removal workflow ensuring rapid removal with minimal re-traumatization of the performer.
• Legal framework reference — References applicable legal frameworks (FOSTA-SESTA, UK Online Safety Act NCII provisions, Stop NCII tools) to support legal action where appropriate.

5.5 Performer Wellness#

• Emotional load monitoring — Tracks cumulative emotional load from moderation activity: total number of filtered harmful messages, severity distribution, and trend over time.
• Burnout risk detection — Identifies performers showing signs of harassment-related burnout: escalating protective measures, declining engagement, and changed behavioral patterns.
• Wellness check-ins — Periodic automated wellness check-ins for active performers with high moderation loads, offering resources and adjusting protection levels.
• Resource surfacing — Surfaces mental health resources, peer support communities for creators, and professional support at appropriate moments.

5.6 Performer Dashboard#

• Safety analytics dashboard — Comprehensive dashboard showing harassment volume trends over time, attack pattern summaries, boundary violation rates by category, a protective action log with decisions, and wellness indicators. Helps performers understand their safety landscape and make informed decisions about their interaction rules.

6. Rehabilitation Pathways (kuanyin-rehabilitation)#

Kuanyin's most distinctive feature is its restorative rather than purely punitive approach to users who have caused harm. Exclusion and punishment may protect the community in the short term but do not help the violating user become a better community member. Rehabilitation pathways offer structured opportunities for growth and reintegration — the long-term investment in community health that punitive systems cannot make.

6.1 Shadow Work Journeys#

Shadow work is a concept from Jungian psychology referring to the process of exploring and integrating the unconscious aspects of personality — including impulses, fears, and patterns of behavior that the conscious self does not acknowledge. Kuanyin adapts this framework for online conflict rehabilitation.

• Guided self-reflection journeys — Structured self-reflection experiences helping violators understand the impact of their behavior on others and themselves.
• Shadow work exercises — Carl Jung-inspired shadow integration exercises helping users recognize and work with the patterns driving harmful behavior, rather than simply suppressing them.
• Non-punitive framing — Journeys are presented explicitly as growth opportunities: "This is a chance to understand yourself better," not "you must do this or be banned."
• Progress milestones — Milestone achievements for completing reflection phases, providing a sense of forward momentum and accomplishment.

6.2 Empathy Training#

• Structured empathy-building exercises — Formal exercises building the capacity to understand and share the feelings of others.
• Perspective-taking scenarios — Concrete scenarios involving community members affected by similar behavior, making abstract impact concrete.
• Impact visualization — Helps users visualize the emotional impact of their behavior on real people.
• Graduated difficulty — Exercises begin with less emotionally challenging scenarios and progress toward more difficult ones as the user demonstrates readiness.

6.3 Restorative Circles#

A restorative circle is a facilitated conversation between the person who caused harm and those affected, focused on understanding, accountability, and repairing the relationship and community fabric. Restorative justice is an alternative to purely punitive approaches originally developed in criminal justice and here adapted for online community contexts. Participation by affected parties is always entirely voluntary.

• Facilitated conversations — Structured restorative conversations between the violating user and affected parties when both consent to participate.
• Consent-based participation — The affected party's participation is always entirely voluntary. No affected party is ever required to participate in a restorative process.
• AI facilitation option — Optional AI facilitation for initial sessions, providing structure and preventing escalation when human facilitation is not available.
• Closure protocols — Structured protocols for reaching mutually acceptable resolution: acknowledgment of harm, agreed changes in behavior, and explicit closure.

6.4 Accountability Tracking#

• Commitment tracking — Tracks accountability commitments made during restorative circles: specific behavioral changes the violating user agreed to make.
• Follow-through prompts — Reminder prompts at appropriate intervals to help users follow through on commitments.
• Accountability partner matching — Matches users undertaking rehabilitation with established community members who can serve as accountability partners and mentors.

6.5 Reintegration Pathways#

Reintegration is the structured return to full community participation after a rehabilitation journey. Rather than a binary excluded/included state, Kuanyin uses graduated reintegration: each phase requires demonstrated positive behavior before advancing to the next.

• Graduated reintegration — Phased return to community access: limited participation (read-only or specific channels only) → supervised participation (all actions visible to moderators) → full participation (normal community access).
• Milestone-based trust restoration — Each reintegration phase requires demonstrated positive behavior before advancing to the next.
• Community sponsor model — Established community members can voluntarily sponsor a returning user, vouching for them and supporting their reintegration with accountability.

7. Merit and Karma Systems (kuanyin-merit-karma)#

The karma system provides positive reinforcement for constructive behavior, creating an alternative to moderation-as-policing. Users who consistently contribute positively to the community earn concrete benefits that recognize and reward that contribution. This makes positive community behavior intrinsically visible in a way that purely punitive systems never do.

7.1 Merit Accumulation#

• Positive action tracking — Tracks merit-earning actions: helpful responses that received thanks, welcomed contributions, conflict de-escalation, community support, and effective mentorship.
• Merit decay — Gradual natural merit decay means scores must be maintained through ongoing contribution, not just past achievement. This prevents long-retired members from holding perpetual status based on historical work.
• Merit events — Specific configurable merit-earning events with configurable point values per platform context.

7.2 Privilege Tiers#

Merit accumulation unlocks privilege tiers with meaningful benefits. The merit-karma library's privilege-tiers module defines four tiers (PrivilegeTierLevel) with progressively expanded capabilities:

Three additional merit scales exist for different purposes. The database library's MeritLedger records a growth-metaphor merit tier progression — seed, sprout, sapling, tree, grove, forest, ecosystem — for ledger display. The foundation constants store an additional MeritTier scale of newcomer → diamond for point thresholds. The capability-gating tier list above (four tiers: basic through bodhisattva) is the one the privilege-tiers module enforces for access decisions.

7.3 Karma Visibility#

• Karma transparency controls — Users control the visibility of their karma score to others, respecting privacy preferences.
• Karma history — Full visible history of karma-affecting actions so users understand what increased or decreased their score.
• Karma explanation — Clear human-readable explanations of why karma increased or decreased for each event.
• Karma appeal — Formal appeal process for karma adjustments the user believes were unfair, with human review.

7.4 Achievement System#

Achievements recognize specific positive contributions, making extraordinary community behavior visible and celebrated:

• First meaningful contribution to the community
• Sustained positive engagement over time
• Effective community mentorship of new members
• Rehabilitation journey completion
• Successful conflict de-escalation
• Cross-community bridge-building between different groups

8. Dharma Analytics (kuanyin-dharma-analytics)#

Dharma (धर्म) in Buddhist philosophy refers to the natural law of things — the way things actually are. Dharma analytics applies this concept to community health: measuring the actual state of community wellbeing, not just surface metrics like engagement volume. A community can show high engagement while actually being in severe distress; dharma analytics is designed to see through that distinction.

8.1 Community Health Metrics#

• Aggregate health dashboards — Platform-wide and community-specific health score dashboards combining harm rates, positive engagement rates, rehabilitation outcomes, and temperature trends.
• Health trend visualization — Time-series visualization of community health trends enabling managers to see the long-term direction of community health.
• Comparative community health — Benchmark health metrics across different community spaces on the platform to identify best-practice communities and struggling ones.
• Health alert thresholds — Configurable alerts when health metrics cross configured warning or critical thresholds, enabling proactive intervention.

8.2 Individual Dharma Path#

• Per-user dharma path tracking — Each user's individual progress: current refuge stage, merit accumulation history, rehabilitation journey status, and growth milestones.
• Progress visualization — Visual representation of the user's journey through the Kuanyin system, making growth tangible and motivating.

8.3 Predictive Wellness#

• Community wellness deterioration prediction — Predictive models for community wellness deterioration based on early warning signal patterns observed in communities that subsequently declined.
• Early warning indicators — Leading behavioral indicators that precede measurable health deterioration: increasing negativity, decreasing positive interactions, rising conflict rates.
• Proactive intervention recommendations — Recommendations for community manager actions that have historically prevented deterioration in similar communities.

8.4 Wisdom Reports#

• Periodic wisdom reports — Generated wisdom reports for community managers: key behavioral trends, emerging conflicts, rehabilitation outcomes, community culture observations, and recommended actions.
• Executive summaries — High-level summaries of platform-wide moderation health for platform leadership.

8.5 Moderation Transparency#

Platform-level transparency about moderation operations is essential for community trust. Without transparency, users cannot evaluate whether the system is operating fairly or understand why actions were taken.

• Action transparency reports — Community-level reports on moderation volumes, action types, and outcomes without identifying individual users.
• False positive tracking — Systematic tracking and reporting of false positive moderation rates to drive continuous improvement.
• Appeal outcome reporting — Reporting on appeal volumes, outcomes, and resolution times.
• Bias monitoring — Continuous monitoring for demographic bias in moderation actions: are certain groups' content flagged at higher rates without corresponding harm? Active bias detection and correction.

9. Cross-Domain Ethics Integration (kuanyin-cross-domain)#

Kuanyin's ethics and safety layer extends across the entire Oshun platform through domain-specific integration modules. Each module adapts the core Kuanyin capabilities to the specific context, user population, and risk profile of each domain. A generic moderation API would force each domain to bend its model to fit moderation logic; domain-specific modules apply the same underlying capabilities with the domain's own vocabulary and calibration.

The following table shows each integration module, which domain it serves, and the specific adaptations it provides:

The platform-wide integration ensures that harm detected in one domain informs safety decisions in others: a user with a persistent harassment pattern in Yemaya will have that context available when they interact in Aphrodite, preventing domain-hopping to escape moderation history.

10. Data Foundation (kuanyin-database)#

All Kuanyin data is persisted with appropriate access controls, retention policies, and privacy safeguards. The 25 tables are organized around six major data concerns:

• Moderation event persistence — All harm detections, intervention decisions, and outcomes recorded with full context for audit trails and system improvement.
• Behavioral profile storage — User behavioral profiles and escalation trajectories stored with privacy-preserving design: profiles inform moderation decisions but are not exposed to other users.
• Rehabilitation journey tracking — Progress through shadow work, empathy training, restorative circles, and reintegration phases, with milestone records.
• Merit and karma ledger — Append-only karma and merit event ledger ensuring historical integrity. Past karma events cannot be altered retroactively.
• Community health time-series — Historical community health metrics for trend analysis and predictive modeling.
• Performer safety records — Performer-specific protection rules, violation logs, and wellness data with performer-controlled access.

11. AI/ML Models (kuanyin-ai-ml-models)#

The AI/ML models library provides typed model descriptors and inference records for every classification and prediction task in Kuanyin, decoupled from the feature libraries so that model implementations can be upgraded without touching feature code.

• Intent classification models — Fine-tuned classification models for distinguishing harmful from benign intent across different content types, communities, and cultural contexts.
• Emotion detection models — Multi-class emotion recognition models calibrated for online communication patterns rather than formal speech.
• Harm detection models — Content harm classifiers trained on domain-specific labeled data with continuous improvement from human review outcomes.
• Behavioral prediction models — Sequential behavior prediction models for escalation trajectory detection and recidivism prediction.
• Fairness and bias evaluation — Regular bias evaluation across demographic groups to detect and correct demographic disparities in model behavior.
• Cross-cultural models — Community-specific and culture-specific model adaptations for communities with distinct communication norms.

12. SDK and API (kuanyin-sdk-api)#

The sdk-api library is the stable public surface that consuming applications use to integrate Kuanyin. It has four modules: rest-api-endpoints, graphql-api, event-system, and typescript-sdk.

• REST API — 19 versioned /v1 endpoints covering intent, emotion, and harm analysis; intervention triggering and completion; trust score and merit queries; rehabilitation and restorative-circle lifecycle; community health; performer protection; dharma analytics; transparency reporting; and appeals.
• GraphQL API — A GraphQL surface over the same capabilities, for consuming applications that prefer graph queries over REST.
• Event system — Ten event categories (intervention, harm_detection, protection, rehabilitation, circle, merit, tier, community, analytics, system) with dotted event names such as intervention.triggered.gentle_nudge and harm.detected.harassment. The REST surface also defines seven WebSocket push-event types for real-time updates.
• TypeScript SDK — A fully typed client for consuming the Kuanyin API, with retry and jitter helpers for reliability.

13. UI Components (kuanyin-ui-components)#

The ui-components library provides typed configuration descriptors that specify UI component behavior — it does not ship React or DOM components itself. Platform teams consume these descriptors and render them in their own component framework, ensuring consistent friction patterns and wellness displays across all surfaces. Five module groups:

• Mindful friction components — Descriptors for pause-and-breathe interventions, Right Speech gate steps, and alternative-expression suggestions.
• Karma and merit displays — Descriptors for karma scores, merit badges, and tier progress.
• Rehabilitation journey UI — Descriptors for the shadow work journey, empathy training exercises, and reintegration progress tracking.
• Analytics dashboard components — Descriptors for community temperature, health scores, and moderation transparency reports.
• Performer dashboard components — Descriptors for the performer-facing safety analytics dashboard.

14. Restorative-Circle Safety Pre-Flight (kuanyin-concordia-restorative)#

The concordia-restorative library (Phase 179.7.3) is the Kuanyin side of the Concordia restorative-mediation integration. Before any restorative circle may open, this library enforces a structured safety pre-flight checklist — ensuring that the conditions for a safe, productive circle are in place before participants are brought together. Rushing into a circle without these preconditions risks re-traumatization and escalation rather than healing.

• Circle kinds — Seven kinds: community harm repair, moderation appeal, performer protection, creator protection, platform reintegration, school/youth program, and workplace restorative.
• Safety checks — Eleven Zod-validated checks covering harmed-party consent, responsible-party acknowledgment, absence of immediate danger, restraining-order status, power-imbalance assessment, facilitator training, language accessibility, a documented trauma-informed plan, child-safety review, guardian linkage, and active safety precautions.
• Required-checks map — Each circle kind has its own mandatory subset; child- and workplace-oriented circles add extra gates beyond the base requirements.
• Opening decision — canOpenCircle keeps a circle closed and returns the missing and failed checks until every required check has passed; helper functions build and update the checklist.

This library ships only the safety-checklist part of the restorative-mediation flow; the broader Concordia integration remains planned (see below).

15. Planned Features#

The following capabilities are not yet present in libs/kuanyin/:

• Live service deployment (planned) — The data layer is currently modelled as in-memory Map stores; a PostgreSQL-backed runtime and the API server implied by KUANYIN_API_PORT are not wired up in this tree.
• AI model evaluation framework (planned) — Systematic evaluation of intent classification, emotion detection, and harm detection model performance including precision/recall metrics, fairness evaluation across demographics, adversarial robustness tests, cross-cultural evaluation, and model drift detection.
• Full restorative mediation (planned, Phase 179) — Beyond the concordia-restorative safety pre-flight checklist (Section 14), the broader Concordia restorative-mediation flow — apology, restitution, reintegration, no-contact boundaries, content-takedown timelines, educational completion, monitoring windows, coercion detection, recurrence measurement, and safe escalation — remains planned.
• Multimodal harm analysis (planned, Phase 32.18.1) — Extending precognition beyond text and behavioral signals to image, video, and audio harm detection (visual threat and hate-symbol recognition, harmful audio and speech analysis, video-scene understanding), so every modality a platform carries passes through the same 74-category taxonomy and severity classification. Today only NCII image detection (Section 5.4) analyzes non-text media.
• Moderator and facilitator wellness (planned, Phase 32.19.1, critical) — Vicarious/secondary-trauma protection for the humans reviewing harmful content and facilitating circles: exposure budgeting and rotation, graded content blurring/desensitization controls, mandatory decompression windows, wellness check-ins, and burnout-trajectory monitoring — the same care Section 5.5 gives performers, applied to moderators and restorative-circle facilitators.
• Bridging-based ranking and toxicity-scoring integration (planned, Phase 32.19.4) — Integration with external toxicity scorers (Google Perspective API attributes) and bridging attributes that reward content which bridges divides between polarized groups, complementing the internal harm taxonomy with ecosystem-standard scores and pro-social ranking signals.
• Federated learning (planned, Phase 32.18.5) — Privacy-preserving federated training of harm-detection models across platforms and tenants, so models improve from distributed signals without centralizing raw user content.

V2 consumes @kuanyin/concordia-restorative through @v2/concordia-substrate for high-impact anti-cheat appeal restorative preflight. Kuanyin owns the moderation_appeal safety checklist and keeps the restorative circle closed until consent, no-immediate-danger, facilitator, and safety-precaution gates pass.

Interpretability and Research Safety Gates (Phases 177–178)#

Kuanyin is the safety-gate co-owner for the interpretability/continual-learning track (Phase 177) and the autonomous-research substrate (Phase 178). Nous emits interpretability safety reports (SAE side-channel signals, circuit-stability metrics, forgetting watchdogs) and autonomous-research proposals; Kuanyin consumes them as release and dual-use governance gates, applying the same harm-taxonomy and human-in-the-loop review it applies to content. Nous owns the models and the research loop; Kuanyin owns the safety verdict.