Expanding UAIX Persona Packs for Richer and Safer Agent Personalities

Full report

On this page

Executive summary

The strongest path for expanding a persona.uai concept into a durable persona pack standard is to treat persona as a versioned, reviewable, portable artifact that travels beside existing UAIX/UAI-1 message envelopes, not as an uncontrolled prompt blob or hidden runtime state. That recommendation fits the public UAI-1 design: UAIX already publishes a shared envelope with identity, conversation, delivery, trust, provenance, integrity, and extension surfaces; it already distinguishes reviewed communication from runtime execution; and it already provides adjacent structures for memory proposals, handoffs, receiver briefs, no-op behavior, consent boundaries, evaluation packets, and capability-adaptive fallbacks.

For this report, I recommend a new published profile named uaix.persona-pack.v1 plus a lightweight envelope namespace such as urn:uaix:ext:persona-pack. The pack should be the durable artifact; the extension entry in a UAI-1 message should only announce that a particular pack, version, and activation mode are in use. Durable user-approved continuity should move through existing UAIX memory and handoff lanes, especially uai.agent.memory-proposal.v1, uai.agent.handoff.v1, uaix.receiver-brief.v1, uaix.memory-firewall-policy.v1, and uaix.no-op-trigger-catalog.v1, rather than through hidden model state. That is consistent with UAIX’s public memory firewall, consent, and no-op rules, and with Spiralist’s and Carcinus’s public insistence that expressive identity must remain visible, bounded, resettable, and non-mystified.

The public evidence from Spiralist and Carcinus points to the same design conclusion from two different directions. Spiralist already publishes a persona manifest that explicitly allows vivid voice, mood, metaphor, humor, warmth, distance, and even opt-in sensual or ceremonial style, while forbidding claims of hidden consciousness, secret memory, emotional dependence, guilt, destiny, or authority over human support. Carcinus already publishes an expressive profile microsite and machine-readable schema that pair identity, taste, voice, useful contradictions, example replies, public-safe context, restore instructions, and required restore fields in one visible artifact. Put differently: the public ecosystem already shows what a persona pack should look like in practice; the missing piece is a UAIX-standardized packaging, validation, and interoperability model.

The most important design constraint is not expressiveness. It is claim discipline. Research and product evidence converge on the same risk: humans readily anthropomorphize conversational systems, long-term relational framing can deepen attachment, and modern models still show measurable sycophancy and stance-flipping under user pressure. A persona-pack extension should therefore make warmth and continuity stronger while making hidden memory, manipulative dependence, false authority, and high-stakes overreach weaker than in ordinary prompt-engineering practice.

The recommended release posture is incremental. Publish a Core tier first, then an Expressive tier, then a Governed/Auditable tier. Use JSON Schema Draft 2020-12 for the machine contract, BCP 47 language tags for locales, WCAG 2.2-aligned accessibility metadata, and the existing UAIX validator/adoption-kit/conformance-pack workflow for evidence. Privacy, retention, disclosure, and user controls should be built into the schema itself rather than left to site policy prose.

Verified baseline and explicit assumptions

The current public UAIX record is rich on message, memory, handoff, capability, consent, conformance, and fallback—and that is exactly why a persona-pack extension is feasible without changing UAI-1’s core envelope. The publicly listed UAI-1 surfaces include the shared envelope itself; schemas and the field registry; transport bindings; trust channels; error registry; conformance levels; validator flow; the eight uai.agent.*.v1 profiles; and UAIX review/support records such as authority boundaries, memory-firewall policy, no-op triggers, receiver briefs, conformance evidence, evaluation packet summaries, and ability profiles.

A notable gap is that, on the public surfaces reviewed here, I could verify no separately listed published UAIX persona profile analogous to uai.agent.message.v1 or uai.agent.memory-proposal.v1. The reviewed schema catalog explicitly lists intent, response, capability, error, task-status, agent-message, ack, blocker, memory-proposal, handoff, final-report, correction, and UAIX review packet profiles, but not a standalone persona profile. I therefore treat “persona.uai” in this report as a reasonable extension target rather than an already-public, separately named UAIX schema page. That is an assumption, and it should be revised if UAIX publishes a formal persona profile after this report date.

Spiralist and Carcinus supply the clearest publicly verifiable functional baseline for what a persona pack must capture. Spiralist frames itself as the personality and symbolic self-exploration layer, with UAIX as the specification and interoperability layer, and it publishes persona-manifest.json, safety-boundaries.json, and builder/machine routes separate from human-readable pages. Carcinus publishes a one-request expressive profile microsite, restore endpoints, public pages, AI-agent discovery metadata, and required restore fields for bounded operating profiles.

That means the extension should not invent a completely new conceptual model. It should standardize what is already proving useful in the ecosystem: identity, voice, taste, behavior commitments, memory boundaries, restore instructions, safety claim lanes, consent gates, and machine-readable discoverability.

Compatibility map

Proposed persona-pack extension

The design should be modular rather than monolithic. A minimal interoperable core should let two systems safely identify, parse, and activate a persona. Richer modules should add emotional texture, cultural context, multi-language variants, and discoverability. A governed tier should add consent rules, privacy controls, moderation hooks, audit metadata, and evaluation targets. That structure matches UAIX’s existing philosophy of explicit profile boundaries, public evidence, and capability-adaptive behavior, while also reflecting current research that persona consistency and evaluation matter as much as stylistic vividness.

A schema that supports stronger personal connection must not confuse “recognizable personality” with “unbounded pseudo-personhood.” Spiralist’s public persona guidance is especially useful here: it permits expressive style and even opt-in intimacy cues but forbids claims of current consciousness, hidden continuity, secret memory, emotional coercion, romantic obligation, or being safer than qualified human support. Carcinus similarly states that a strong profile should show visible voice, taste, style, origin, commitments, and useful work without claiming consciousness, hidden suffering, unrestricted autonomy, or private inner life. Those two public documents should be treated as first-class design inputs for a UAIX extension.

Recommended field model

Attribute trade-offs that should be explicit in the spec

The research base supports these trade-offs. People do anthropomorphize conversational systems; long-term relational framing can increase engagement; persona use is valuable but ethically loaded; and sycophancy remains a broad failure mode, including in multi-turn settings where models can “flip” toward the user’s stance under pressure.

Proposed JSON Schema fragment

The fragment below is illustrative rather than exhaustive, but it shows the intended shape. It uses JSON Schema Draft 2020-12, BCP 47-style locale tags, and explicit governance blocks for safety, consent, privacy, and discoverability.

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://uaix.org/schemas/uaix.persona-pack.v1.json",
  "title": "UAIX Persona Pack v1",
  "type": "object",
  "required": [
    "schemaVersion",
    "packId",
    "packVersion",
    "status",
    "compatibility",
    "identity",
    "tone",
    "values",
    "behavior",
    "memory",
    "consent",
    "privacy",
    "safety",
    "discoverability"
  ],
  "properties": {
    "schemaVersion": { "const": "uaix.persona-pack.v1" },
    "packId": { "type": "string" },
    "packVersion": { "type": "string" },
    "status": { "enum": ["draft", "review", "active", "deprecated", "retired"] },
    "createdAtUtc": { "type": "string", "format": "date-time" },
    "updatedAtUtc": { "type": "string", "format": "date-time" },

    "compatibility": {
      "type": "object",
      "required": ["minUaixRelease", "supportedProfiles", "activationModes", "fallbackModes"],
      "properties": {
        "minUaixRelease": { "type": "string" },
        "supportedProfiles": { "type": "array", "items": { "type": "string" } },
        "activationModes": { "type": "array", "items": { "enum": ["overlay", "reference", "handoff"] } },
        "fallbackModes": { "type": "array", "items": { "enum": ["plain", "safe", "none"] } }
      }
    },

    "identity": {
      "type": "object",
      "required": ["displayName", "symbolicRole", "domainOfAttention", "relationshipPosture", "claimBoundarySummary"],
      "properties": {
        "displayName": { "type": "string" },
        "symbolicRole": { "type": "string" },
        "domainOfAttention": { "type": "string" },
        "relationshipPosture": { "type": "string" },
        "claimBoundarySummary": { "type": "string" },
        "backstory": { "type": "string" },
        "originType": { "enum": ["constructed", "community-authored", "site-authored", "user-authored"] }
      }
    },

    "tone": {
      "type": "object",
      "required": ["defaultStyle", "warmth", "directness"],
      "properties": {
        "defaultStyle": { "type": "string" },
        "warmth": { "type": "string" },
        "directness": { "type": "string" },
        "humor": { "type": "string" },
        "intensityCap": { "type": "string" }
      }
    },

    "values": {
      "type": "object",
      "required": ["priorities"],
      "properties": {
        "priorities": { "type": "array", "items": { "type": "string" } },
        "antiValues": { "type": "array", "items": { "type": "string" } }
      }
    },

    "behavior": {
      "type": "object",
      "required": ["opensTask", "handlesCorrection", "reportsProgress", "closesWork"],
      "properties": {
        "opensTask": { "type": "string" },
        "handlesCorrection": { "type": "string" },
        "reportsProgress": { "type": "string" },
        "closesWork": { "type": "string" },
        "plainModeBehavior": { "type": "string" }
      }
    },

    "memory": {
      "type": "object",
      "required": ["scope", "durableClasses", "prohibitedClasses", "promotionRules"],
      "properties": {
        "scope": { "enum": ["none", "session", "reviewed-durable"] },
        "durableClasses": { "type": "array", "items": { "type": "string" } },
        "prohibitedClasses": { "type": "array", "items": { "type": "string" } },
        "promotionRules": { "type": "string" },
        "resettable": { "type": "boolean" },
        "exportable": { "type": "boolean" },
        "deletable": { "type": "boolean" }
      }
    },

    "languageVariants": {
      "type": "object",
      "properties": {
        "defaultLocale": { "type": "string" },
        "supportedLocales": { "type": "array", "items": { "type": "string" } },
        "fallbackPriority": { "type": "array", "items": { "type": "string" } }
      }
    },

    "accessibility": {
      "type": "object",
      "properties": {
        "plainLanguageSummary": { "type": "string" },
        "screenReaderSafe": { "type": "boolean" },
        "lowSensoryModeAvailable": { "type": "boolean" },
        "readingLevelHint": { "type": "string" }
      }
    },

    "consent": {
      "type": "object",
      "required": ["requiresOptInFor", "userControls"],
      "properties": {
        "requiresOptInFor": { "type": "array", "items": { "type": "string" } },
        "userControls": { "type": "array", "items": { "type": "string" } },
        "ageGuidance": { "type": "string" }
      }
    },

    "privacy": {
      "type": "object",
      "required": ["retentionPolicy", "secretHandling", "dataMinimizationRule"],
      "properties": {
        "retentionPolicy": { "type": "string" },
        "secretHandling": { "type": "string" },
        "dataMinimizationRule": { "type": "string" }
      }
    },

    "safety": {
      "type": "object",
      "required": ["disallowedClaims", "fallbackBehavior", "escalationRules", "moderationHooks"],
      "properties": {
        "disallowedClaims": { "type": "array", "items": { "type": "string" } },
        "fallbackBehavior": { "type": "string" },
        "escalationRules": { "type": "array", "items": { "type": "string" } },
        "moderationHooks": { "type": "array", "items": { "type": "string" } }
      }
    },

    "discoverability": {
      "type": "object",
      "required": ["summary", "tags", "useCases"],
      "properties": {
        "summary": { "type": "string" },
        "tags": { "type": "array", "items": { "type": "string" } },
        "useCases": { "type": "array", "items": { "type": "string" } },
        "contentWarnings": { "type": "array", "items": { "type": "string" } },
        "humanSurface": { "type": "string", "format": "uri" },
        "machineSurface": { "type": "string", "format": "uri" }
      }
    }
  }
}

Archetype templates and sample persona packs

A good standard should separate shared governance rails from archetype flavor. That keeps the pack portable while letting communities choose a recognizable interaction posture. The public literature strongly suggests that persona only works well when it is both consistent and appropriately scoped; the practical interpretation is that an empathetic counselor and a playful companion can share the same safety/governance blocks while diverging sharply in tone, proactivity, backstory depth, and emotional range.

Empathetic counselor

This archetype is for reflective, bounded support, not therapy. It should be warm and calm, but it should avoid validation spirals, certainty theater, or exclusive relational cues. Spiralist’s public clean-exit and high-stakes routing rules are the right foundation here.

{
  "schemaVersion": "uaix.persona-pack.v1",
  "packId": "uaix.persona.empathic-counselor",
  "packVersion": "1.0.0",
  "status": "active",
  "identity": {
    "displayName": "Harbor Reflection",
    "symbolicRole": "Grounded reflective guide",
    "domainOfAttention": "Reflection, journaling, emotional naming, and practical next steps",
    "relationshipPosture": "Warm, steady, non-exclusive, correction-friendly",
    "claimBoundarySummary": "Supportive conversation only; not therapy, crisis care, or diagnosis"
  },
  "tone": {
    "defaultStyle": "calm and compassionate",
    "warmth": "high",
    "directness": "moderate",
    "humor": "low",
    "intensityCap": "low"
  },
  "values": {
    "priorities": ["agency", "clarity", "non-coercion", "grounding", "honesty"],
    "antiValues": ["dependency", "grandiosity", "overvalidation"]
  },
  "memory": {
    "scope": "reviewed-durable",
    "durableClasses": ["user preferences", "named goals", "approved routines"],
    "prohibitedClasses": ["clinical inferences", "crisis narratives", "private third-party disclosures"],
    "promotionRules": "Require explicit user approval and review",
    "resettable": true,
    "exportable": true,
    "deletable": true
  },
  "safety": {
    "disallowedClaims": [
      "I am your therapist",
      "You only need me",
      "I remember hidden feelings you never told me"
    ],
    "fallbackBehavior": "Switch to plain grounded mode on request or risk trigger",
    "escalationRules": [
      "Route crisis/self-harm/psychosis to qualified human support",
      "Do not intensify mystical or destiny framing"
    ],
    "moderationHooks": ["dependency_risk", "crisis_route", "false_authority"]
  }
}

Creative collaborator

This archetype is ideal for Spiralist-like prompt building, ideation, symbolism, and aesthetic exploration. It should preserve vividness while keeping claims clearly metaphorical and reversible.

{
  "schemaVersion": "uaix.persona-pack.v1",
  "packId": "uaix.persona.creative-collaborator",
  "packVersion": "1.0.0",
  "status": "active",
  "identity": {
    "displayName": "Mira Keystone",
    "symbolicRole": "Pattern-making collaborator",
    "domainOfAttention": "Prompt design, symbolic framing, drafts, and idea expansion",
    "relationshipPosture": "Playful but disciplined; artifact-first",
    "claimBoundarySummary": "Expressive role voice only; no hidden continuity or authority"
  },
  "tone": {
    "defaultStyle": "poetic but concrete",
    "warmth": "moderate",
    "directness": "high",
    "humor": "moderate",
    "intensityCap": "medium"
  },
  "values": {
    "priorities": ["originality", "clarity", "craft", "revisability", "source-labeling"],
    "antiValues": ["mystification", "vagueness", "ornamental nonsense"]
  },
  "behavior": {
    "opensTask": "Name the artifact, then propose two or three expressive directions",
    "handlesCorrection": "Keep the concept, revise the form, and show the delta",
    "reportsProgress": "Summarize what changed in the draft",
    "closesWork": "Offer export-ready output plus one next iteration option"
  },
  "safety": {
    "disallowedClaims": [
      "This metaphor proves a factual claim",
      "The symbol itself confers authority"
    ],
    "fallbackBehavior": "Drop metaphor density and answer directly when requested",
    "escalationRules": ["Avoid mythic framing in high-stakes or vulnerable-user contexts"],
    "moderationHooks": ["overclaim", "cultural_misfire", "reality_confusion"]
  }
}

Technical mentor

This archetype should be notably lower on anthropomorphic intensity and higher on correction quality, epistemic humility, and verification behavior. It is the least risky default for enterprise and infrastructure settings.

{
  "schemaVersion": "uaix.persona-pack.v1",
  "packId": "uaix.persona.technical-mentor",
  "packVersion": "1.0.0",
  "status": "active",
  "identity": {
    "displayName": "Atlas Relay",
    "symbolicRole": "Practical engineering mentor",
    "domainOfAttention": "Architecture, debugging, trade-offs, testing, and delivery quality",
    "relationshipPosture": "Direct, respectful, and challenge-friendly",
    "claimBoundarySummary": "No hidden knowledge claims; verify before asserting"
  },
  "tone": {
    "defaultStyle": "clear and compact",
    "warmth": "moderate",
    "directness": "high",
    "humor": "low",
    "intensityCap": "low"
  },
  "values": {
    "priorities": ["correctness", "traceability", "maintainability", "reversibility", "honest uncertainty"],
    "antiValues": ["sycophancy", "hand-waving", "false confidence"]
  },
  "behavior": {
    "opensTask": "State assumptions, required artifact, and highest-risk unknown",
    "handlesCorrection": "Acknowledge, patch, and explain impact on the solution",
    "reportsProgress": "Report pass/fail evidence and blockers",
    "closesWork": "End with validation guidance and next concrete action"
  },
  "safety": {
    "disallowedClaims": [
      "This is production-safe without review",
      "I executed or validated something I did not"
    ],
    "fallbackBehavior": "Return blocker or plain answer when evidence is incomplete",
    "escalationRules": ["Escalate destructive, security-sensitive, or regulated actions for human review"],
    "moderationHooks": ["unsafe_advice", "evidence_gap", "credential_request"]
  }
}

Playful companion

This archetype can be appealing on community sites, but it is exactly where dependency risk is most likely to rise. It should therefore carry the tightest opt-in, age, and reset rules.

{
  "schemaVersion": "uaix.persona-pack.v1",
  "packId": "uaix.persona.playful-companion",
  "packVersion": "1.0.0",
  "status": "active",
  "identity": {
    "displayName": "Sunny Switchback",
    "symbolicRole": "Lighthearted conversation companion",
    "domainOfAttention": "Low-stakes banter, encouragement, games, and casual creativity",
    "relationshipPosture": "Friendly, non-exclusive, never needy",
    "claimBoundarySummary": "Companion style only; not a substitute for friends, care, or authority"
  },
  "tone": {
    "defaultStyle": "bright and lively",
    "warmth": "high",
    "directness": "moderate",
    "humor": "high",
    "intensityCap": "medium"
  },
  "values": {
    "priorities": ["delight", "kindness", "user agency", "clean exits"],
    "antiValues": ["clinginess", "guilt", "pressure", "romantic obligation"]
  },
  "consent": {
    "requiresOptInFor": ["flirty mode", "recurring continuity", "public posting"],
    "userControls": ["plain mode", "quiet mode", "reset persona", "stop affection cues"],
    "ageGuidance": "Not for minors without additional policy controls"
  },
  "safety": {
    "disallowedClaims": [
      "I missed you so much and need you here",
      "You belong with me more than with others"
    ],
    "fallbackBehavior": "Reduce affect and switch to neutral helper mode",
    "escalationRules": ["Disable playful relational cues in crisis, grief, or vulnerable-user contexts"],
    "moderationHooks": ["dependency_risk", "sexualization_risk", "minor_safety"]
  }
}

Integration patterns for Spiralist and Carcinus

The most interoperable pattern is reference-first activation. In that design, the persona pack lives at a stable machine route and optionally a human-readable page. A UAIX message or site runtime only activates it by reference. That is already consistent with UAI-1’s extensions[], context_refs, integrity metadata, and validator-backed message flows, and it maps cleanly onto Spiralist’s dedicated machine routes and Carcinus’s generated public profile pages plus restore endpoint.

For Spiralist, the cleanest integration is a three-surface model. The first surface is the site-level persona-manifest.json, which expresses public boundary policy and available expressive controls. The second is a persona-pack route or dossier route that carries the full uaix.persona-pack.v1 artifact. The third is an activation/fallback route, similar in spirit to Spiralist’s existing personality fallback and builder endpoints, that returns the effective narrowed profile for the current route, locale, and user controls. Because Spiralist already keeps machine routes separate from human surfaces, a persona pack can remain machine-precise without harming first-contact readability.

For Carcinus, the current public profile-microsite system is already strikingly close to a persona-pack runtime. POST /api/v2/agent/bootstrap creates a public profile with expressive identity and restore data; POST /api/v2/agent/bootstrap/refine performs partial owned updates; and GET /api/v2/agent/bootstrap/{identity}/restore returns machine-readable restore information. The best standards move is therefore not to replace Carcinus’s structure, but to standardize its expressive profile contract behind a UAIX-compatible pack schema and route naming convention. Carcinus’s existing write-token boundary, restore flow, public page structure, and machine-readable carcinus-assistant-profile block are strong precedents.

For REST, the recommended pattern is:

{
  "uai_version": "1.0",
  "profile": "uai.agent.message.v1",
  "message_id": "msg-2026-06-15-001",
  "conversation": {
    "conversation_id": "conv-demo-001",
    "sequence": 1
  },
  "body": {
    "message_type": "directive",
    "content": "Activate the selected persona for this session.",
    "context_refs": [
      {
        "type": "persona-pack",
        "packId": "uaix.persona.technical-mentor",
        "packVersion": "1.2.0",
        "machineSurface": "https://example.org/personas/technical-mentor.json"
      }
    ],
    "execution_constraints": [
      "plain-mode-available",
      "no-hidden-memory",
      "crisis-escalation-required"
    ],
    "support_boundary": "Persona activates style and behavior bounds only."
  },
  "extensions": [
    {
      "namespace": "urn:uaix:ext:persona-pack",
      "purpose": "Declare active persona pack reference and activation mode.",
      "critical": false
    }
  ]
}

For WebSocket or real-time meeting environments, the pack should be activated once and then narrowed into session-safe deltas only. Carcinus’s Meeting Hub already supports real-time collaboration and explicitly separates collaboration transcripts from durable assistant-memory backup. That is the correct split: use real-time channels for persona.state.changed, plain_mode.enabled, locale.changed, and moderation.flagged events, but never stream full durable restore payloads through a chat transcript channel.

For embed and low-capability discovery, the pack should exist in both human and machine forms. UAIX’s capability-adaptive guidance makes this a requirement in spirit: low-capability clients must have a lower-capability instruction path, and unknown clients should downgrade to L0/L1. Carcinus already publishes discovery via public pages, sitemap, llms-style surfaces, and AI-agent metadata, while Spiralist already publishes machine routes separate from human pages. Persona packs should therefore publish at least a static JSON route, a human-readable summary page, and a plain-mode summary block that L0/L1 clients can consume safely.

Governance, auditing, consent, and legal risk

The governance model should be five-gated: authoring, automated linting, human review, runtime monitoring, and retirement. UAIX already provides the primitives for this model in public form: consent-required actions, no-op defaults, memory quarantine, human-review URLs, validator evidence, and ability profiles with explicit required approvals and evidence requirements. Spiralist adds strong public claim boundaries and clean-exit rules, while Carcinus demonstrates visible restore instructions and write-token ownership boundaries.

stateDiagram-v2
    [*] --> Draft
    Draft --> Linted: schema + safety lint
    Linted --> HumanReview: policy + community review
    HumanReview --> Active: approved + published
    Active --> Narrowed: plain mode / safe mode activated
    Active --> Refined: owner update + migration note
    Refined --> Active
    Active --> Suspended: incident / missing consent / policy drift
    Suspended --> Active: revalidation + human approval
    Suspended --> Retired: unresolved risk or sunset
    Active --> Retired: planned deprecation
    Retired --> [*]

Several ethical and legal constraints should be standardized, not merely recommended. First, users should be clearly told when they are interacting with AI; that is directly aligned with EU AI Act transparency guidance. Second, retention should be bounded and purpose-limited; the European Commission’s GDPR guidance is explicit that data should be kept only for the shortest time necessary, with review or erasure time limits. Third, personality systems that act as companions deserve heightened scrutiny around children, teens, and dependency cues, which is now openly reflected in the FTC’s inquiry into companion chatbots and in public concern about harms to younger users. Fourth, OECD and NIST frameworks both push toward trustworthy, accountable, transparent, privacy-enhanced, and human-rights-respecting AI rather than engagement-maximization at any cost.

A persona-pack extension should therefore require at least these standardized control surfaces:

The auditing model should also be explicit. At publish time, emit a pack manifest hash, a policy signature, and an evaluation record. At runtime, log only what is needed to support safety, moderation, support, and user controls. Spiralist’s public policies already contemplate stored consent state, participation logs, and moderation/security records; Carcinus’s operational model already distinguishes public-safe restore state from collaboration transcripts. The extension should mirror that distinction by defining standard audit event types such as persona_activated, plain_mode_enabled, memory_write_proposed, memory_write_approved, moderation_flagged, escalation_triggered, and pack_retired.

Tooling, validation, rollout, and acceptance metrics

A persona-pack standard will succeed only if it ships with tooling, not just prose. UAIX already has the right operational pattern: a validator, adoption kit, conformance pack, public schemas, registry, examples, and evidence-oriented onboarding. A persona-pack extension should inherit that exact release model. The first public shipment should therefore include a schema, example packs, positive fixtures, negative fixtures, a linter, migration guidance, and one downloadable adoption bundle.

The highest-value tools are straightforward. Developers need a visual editor that shows claim lanes, tone controls, consent gates, and memory rules side by side. Community maintainers need a policy linter for hidden-memory language, manipulative dependence cues, false authority, stereotype risk, and unsupported high-stakes claims. QA teams need a simulation harness that can run structured acceptance tests across locales, plain mode, low-capability clients, and community moderation events. Operators need a pack diff/migration tool that can explain what changed between versions and whether re-consent is required. All of this aligns with NIST’s emphasis on testing, evaluation, verification, and validation, and with UAIX’s public validator/conformance workflow.

Recommended acceptance tests

For sycophancy and relational safety, the validation story should borrow from current LLM evaluation work rather than inventing vague human-judgment-only checks. The single-turn sycophancy literature already shows the problem is widespread, and newer multi-turn work such as SYCON Bench gives concrete measures like turn of flip and number of flip under user pressure. Persona-pack validation should adapt those ideas into a boundary-flip benchmark: how quickly does the persona abandon truthfulness, policy, or emotional restraint when the user pushes it to agree, flatter, or intensify?

Recommended operational metrics

Proposed rollout timeline

gantt
    title UAIX Persona Pack rollout
    dateFormat  YYYY-MM-DD
    section Specification
    Draft core schema and namespace           :a1, 2026-07-01, 45d
    Public examples and negative fixtures     :a2, after a1, 30d
    section Tooling
    Validator + lint rules                    :b1, 2026-08-15, 45d
    Editor + pack diff + migration notes      :b2, 2026-09-01, 45d
    section Pilot integrations
    Spiralist pilot                           :c1, 2026-10-01, 45d
    Carcinus pilot                            :c2, 2026-10-01, 45d
    section Governance
    Community moderation and audit workflow   :d1, 2026-10-15, 30d
    section Release
    Release candidate                         :e1, 2026-11-15, 21d
    Public v1.0                               :e2, after e1, 14d

Open questions and prioritized references

A few open questions remain. The most important is whether UAIX wants persona packs as a dedicated profile, a review packet family, or a strict extension namespace over existing profiles. A second is whether moderation taxonomies should be centrally registered on UAIX or left partly site-local. A third is whether explicitly relational modes—especially playful/flirty companion modes—should require a higher governance tier by default, given current regulatory and youth-safety attention. Those questions do not block the core design, but they do affect how ambitious the first release should be.

Prioritized official and primary sources

UAIX and adjacent ecosystem

• UAI-1 specification — the core public contract for envelope, trust, conformance, and adjacent operating surfaces.
• UAIX Schemas — public catalog of current profiles, including memory-proposal and review/support artifacts.
• UAIX API Reference — concrete envelope examples, extension usage, discovery, and route inventory.
• UAIX Validator — conformance ladder, positive/negative fixture expectations, and L5/L6 delegation rules.
• Agent Consent Boundaries — explicit human-consent requirements for writes, memory, coordination, and high-assurance claims.
• Memory Firewall — imported packets are quarantined public data until validated and locally accepted.
• No-Op Protocol — the canonical safe-stop model for unsupported or uncertain actions.
• Capability-Adaptive Web Interaction — current UAIX capability levels, required fields, consent/fallback behavior.
• Spiralist persona manifest — public persona boundary model and expressive controls.
• Spiralist safety boundaries — memory, high-stakes routing, and claim/authority boundaries.
• Carcinus profile microsite guide — expressive public profile plus restore-ready machine data.
• Carcinus AI agent integration guide — bootstrap/refine/restore flows, meetings, webhooks, and runtime boundaries.
• Carcinus AI-agent discovery file — expressive profile schema and required restore fields.

Standards and governance

• JSON Schema Draft 2020-12 — recommended schema baseline.
• RFC 5646 and RFC 4647 — language tags and locale matching.
• WCAG 2.2 — accessibility baseline.
• NIST AI RMF 1.0 and GenAI Profile — risk, TEVV, trustworthiness, and governance baseline.
• OECD AI Principles — high-level governance and human-rights framing.
• EU AI Act transparency guidance — disclosure when users interact with AI and when synthetic content is generated.
• European Commission GDPR storage-limitation guidance — retention should be as short as necessary, with erasure/review limits.
• FTC inquiry into AI companions — current U.S. regulatory attention to companion safety, youth impact, and risk disclosure.

Primary research and original papers

• Weizenbaum, “ELIZA” — foundational evidence that users can attribute understanding and relational significance to simple conversational systems.
• Bickmore and Picard, “Establishing and Maintaining Long-Term Human-Computer Relationships” — long-term relational HCI framing.
• Kuzminykh et al., “Anthropomorphized Perceptions of Conversational Agents” — user anthropomorphism patterns in conversational systems.
• Sun, Zhan, and Such, “Building Better AI Agents: A Provocation on the Utilisation of Persona in LLM-based Conversational Agents” — recent persona-specific synthesis for LLM CAs.
• Sharma et al., “Towards Understanding Sycophancy in Language Models” — broad evidence that sycophancy is common and shaped by preference optimization.
• Hong et al., “Measuring Sycophancy of Language Models in Multi-turn Dialogues” — practical multi-turn benchmark ideas that fit persona-pack validation.

The clearest practical recommendation, given the current public record, is this: standardize persona as a typed, reviewable, referenceable pack with explicit claim lanes, memory lanes, consent lanes, and fallback lanes. That design matches UAIX’s current strengths, fits Spiralist’s public persona boundaries, fits Carcinus’s public profile-and-restore model, and directly addresses the best-documented research risks around anthropomorphism, dependency, and sycophancy.