ARAL Core v1.2

ARAL Core Specification

Version: 1.2
Status: Release Candidate
Layers: L1-L5
Requirements: 83
Philosophy: Persona-Centric Architecture

Preamble: Persona as Agent Identity

“The persona is not a configuration of an agent—it IS the agent.”

ARAL adopts a persona-centric philosophy where:

Identity is Primary: An agent without a persona has no identity, purpose, or behavioral contract
Everything is Persona-Scoped: Memory, capabilities, reasoning—all bound to persona identity
Cryptographic Provenance: Every action traces back to a signed persona
Dynamic Evolution: Personas can be hot-swapped, orchestrated, and evolved
Accountability Anchor: Personas provide verifiable audit trails

This specification defines the five foundational layers that enable persona-driven agent architectures. The runtime (L1) provides the body, memory (L2) the brain, capabilities (L3) the senses, reasoning (L4) the mind, and persona (L5) the soul.

See Also: Persona as Agent Identity Guide

1. Layer 1: Runtime

The Runtime layer manages execution environment, resources, and lifecycle.

1.1 Requirements

ID	Requirement	Level
ARAL-L1-001	Runtime MUST provide unique agent instance ID	MUST
ARAL-L1-002	Runtime MUST implement graceful shutdown with configurable timeout	MUST
ARAL-L1-003	Runtime MUST expose health check endpoint	MUST
ARAL-L1-004	Runtime SHOULD implement resource quotas (CPU, memory, connections)	SHOULD
ARAL-L1-005	Runtime MUST provide fallback behavior on resource exhaustion	MUST
ARAL-L1-006	Runtime MUST log lifecycle events (start, stop, error)	MUST
ARAL-L1-007	Runtime SHOULD support hot reload of configuration	SHOULD
ARAL-L1-008	Runtime MUST enforce maximum execution time per request	MUST
ARAL-L1-009	Runtime MUST provide metrics endpoint (Prometheus format)	MUST
ARAL-L1-010	Runtime SHOULD implement backpressure mechanisms	SHOULD

1.2 Schema

See schemas/manifest.schema.json for agent manifest format.

2. Layer 2: Memory

The Memory layer manages agent state and context.

2.1 Memory Types

Type	Scope	Persistence
Working	Current request	None
Short-term	Session	Session duration
Long-term	Agent lifetime	Persistent
Episodic	Specific events	Selective

2.2 Requirements

ID	Requirement	Level
ARAL-L2-001	Memory MUST implement working memory for current context	MUST
ARAL-L2-002	Memory MUST support TTL-based expiration	MUST
ARAL-L2-003	Memory SHOULD implement vector similarity search	SHOULD
ARAL-L2-004	Memory MUST provide atomic read-modify-write operations	MUST
ARAL-L2-005	Memory MUST NOT store secrets in plaintext	MUST
ARAL-L2-006	Memory SHOULD support memory partitioning by context	SHOULD
ARAL-L2-007	Memory MUST log all write operations for audit	MUST
ARAL-L2-008	Memory MUST implement size limits with eviction policy	MUST
ARAL-L2-009	Memory SHOULD support snapshots for debugging	SHOULD
ARAL-L2-010	Memory MUST handle concurrent access safely	MUST

3. Layer 3: Capabilities

The Capabilities layer defines available actions and tools.

3.1 Capability Definition

{
  "id": "capability-uuid",
  "name": "web_search",
  "version": "1.0.0",
  "description": "Search the web",
  "input_schema": { "type": "object", "properties": {...} },
  "output_schema": { "type": "object", "properties": {...} },
  "permissions": ["network:read"],
  "rate_limit": { "requests": 100, "period": "1m" },
  "timeout_ms": 30000
}

3.2 Requirements

ID	Requirement	Level
ARAL-L3-001	Capabilities MUST declare input/output schemas	MUST
ARAL-L3-002	Capabilities MUST declare required permissions	MUST
ARAL-L3-003	Capabilities MUST implement timeout handling	MUST
ARAL-L3-004	Capabilities SHOULD implement retry with exponential backoff	SHOULD
ARAL-L3-005	Capabilities MUST validate inputs against schema	MUST
ARAL-L3-006	Capabilities MUST validate outputs against schema	MUST
ARAL-L3-007	Capabilities MUST log invocations with trace ID	MUST
ARAL-L3-008	Capabilities SHOULD implement rate limiting	SHOULD
ARAL-L3-009	Capabilities MUST handle errors gracefully	MUST
ARAL-L3-010	Capabilities MUST NOT exceed declared permissions	MUST
ARAL-L3-011	Capabilities SHOULD provide estimated cost/latency	SHOULD
ARAL-L3-012	Capabilities MUST support cancellation	MUST

4. Layer 4: Reasoning

The Reasoning layer handles decision-making, inference, and LLM interactions.

4.1 Requirements

ID	Requirement	Level
ARAL-L4-001	Reasoning MUST produce structured action decisions	MUST
ARAL-L4-002	Reasoning MUST include confidence scores when applicable	MUST
ARAL-L4-003	Reasoning MUST respect L5 Persona constraints	MUST
ARAL-L4-004	Reasoning SHOULD implement chain-of-thought logging	SHOULD
ARAL-L4-005	Reasoning MUST handle ambiguous inputs gracefully	MUST
ARAL-L4-006	Reasoning MUST NOT directly modify L2 Memory	MUST
ARAL-L4-007	Reasoning MUST NOT bypass L3 Capability validation	MUST
ARAL-L4-008	Reasoning SHOULD support multiple inference strategies	SHOULD
ARAL-L4-009	Reasoning MUST implement request timeout	MUST
ARAL-L4-010	Reasoning MUST log all decisions with rationale	MUST
ARAL-L4-011	Reasoning MAY implement self-reflection loops	MAY
ARAL-L4-012	Reasoning MUST handle model errors gracefully	MUST
ARAL-L4-013	Reasoning MUST support multiple LLM providers	MUST
ARAL-L4-014	Reasoning MUST implement provider fallback chains	MUST
ARAL-L4-015	Reasoning SHOULD support provider load balancing	SHOULD
ARAL-L4-016	Reasoning SHOULD support cost-based routing	SHOULD
ARAL-L4-017	Reasoning MAY implement multi-provider consensus	MAY
ARAL-L4-018	Reasoning MUST support content safety moderation	MUST
ARAL-L4-019	Reasoning SHOULD implement quality-based model selection	SHOULD
ARAL-L4-020	Reasoning MUST track per-provider usage and costs	MUST
ARAL-L4-021	Reasoning MUST support weighted LLM ponderation (blending)	MUST
ARAL-L4-022	Reasoning MUST normalize ponderation weights to sum to 1.0	MUST
ARAL-L4-023	Reasoning MUST support advanced llm_config with providers array	MUST
ARAL-L4-024	Reasoning SHOULD support routing_strategy options (specialized, cost_optimized, quality_first, latency_first)	SHOULD
ARAL-L4-025	Reasoning SHOULD support response aggregation (best_of_n, ensemble, consensus)	SHOULD
ARAL-L4-026	Reasoning MAY support per-provider configuration overrides	MAY
ARAL-L4-027	Reasoning MAY support model-specific prompt customization (llm_specific)	MAY
ARAL-L4-028	Reasoning MAY support task-based routing rules (extensions.routing_rules)	MAY

4.2 Multi-LLM Orchestration

ARAL supports multiple LLM providers with intelligent routing and weighted blending:

{
  "llm_config": {
    "providers": [
      {
        "name": "openai",
        "models": ["gpt-4", "gpt-3.5-turbo"],
        "priority": 1,
        "cost_per_1k_tokens": 0.03,
        "max_tokens": 8192
      },
      {
        "name": "anthropic",
        "models": ["claude-3-opus", "claude-3-sonnet"],
        "priority": 2,
        "cost_per_1k_tokens": 0.015,
        "max_tokens": 200000
      }
    ],
    "routing_strategy": "cost_optimized",
    "fallback_chain": ["gpt-4", "claude-3-opus", "gpt-3.5-turbo"],
    "load_balancing": "round_robin",
    "ponderation": {
      "enabled": false,
      "weights": {
        "gpt-5.2": 0.8,
        "claude-sonnet-4.5": 0.2
      },
      "merge_strategy": "weighted_blend"
    },
    "consensus_mode": {
      "enabled": false,
      "min_providers": 2,
      "agreement_threshold": 0.75
    },
    "moderation": {
      "enabled": true,
      "provider": "openai",
      "block_on_violation": true
    }
  }
}

4.3 Routing Strategies

Strategy	Description	Use Case
cost_optimized	Route to cheapest model that meets quality	Production efficiency
quality_first	Always use highest-quality model	Critical decisions
latency_first	Route to fastest-responding provider	Real-time apps
round_robin	Distribute load evenly across providers	Load balancing
ponderation	Weighted blend of multiple model outputs	Ensemble responses
priority_based	Use providers in priority order	Fallback chains
consensus	Query multiple models, require agreement	High-stakes decisions

4.4 Decision Format

{
  "decision_id": "uuid",
  "timestamp": "ISO8601",
  "action": "capability_id",
  "parameters": {...},
  "confidence": 0.95,
  "rationale": "Chain of thought...",
  "alternatives": [...]
}

5. Layer 5: Persona — The Agent’s Identity and Soul

“The persona is the agent. Without a persona, there is no agent—only an empty runtime.”

The Persona layer is the identity core of an ARAL agent. It defines:

Who the agent is (identity, role, purpose)
What the agent can do (capabilities, tools, domains)
What the agent must not do (constraints, denials, safety limits)
How the agent thinks (reasoning style, temperature, verbosity)
How the agent collaborates (merge behavior, compatible modes)
How the agent is accountable (audit criteria, signatures)

Every ARAL agent is incarnated through a persona. When you instantiate an agent with a persona, you are not merely configuring it—you are giving it life, identity, and purpose.

5.1 Persona as Constitutional Contract

A persona is a behavioral constitution that:

Defines Identity: UUID, name, role, version
Declares Rights: Allowed capabilities (what it CAN do)
Sets Boundaries: Denied capabilities, max reasoning depth (what it CANNOT do)
Establishes Values: Audit criteria, priorities, thinking style
Builds Relationships: Compatible personas, defer_to, conflicts_with
Ensures Accountability: Cryptographic signature, audit trails

5.2 Persona Definition

{
  "id": "d4e5f6a7-8b9c-0d1e-2f3a-4b5c6d7e8f9a",
  "name": "Research Analyst Pro",
  "role": "research_analyst",
  "version": "1.0.0",

  "metadata": {
    "category": "research",
    "description": "Expert research analyst with multi-model ensemble",
    "tags": ["research", "analysis", "expert"],
    "skills": ["data-analysis", "synthesis", "critical-thinking"],
    "author": "ARAL Standard",
    "license": "MIT",
    "created": "2026-01-16T00:00:00Z",
    "updated": "2026-01-16T00:00:00Z"
  },

  "config": {
    "temperature": 0.3,
    "formality": "academic",
    "verbosity": "detailed",
    "thinking_style": "analytical",
    "priority": 80,

    "llm_ponderation": {
      "enabled": true,
      "weights": {
        "gpt-4": 0.6,
        "claude-opus": 0.3,
        "gemini-pro": 0.1
      },
      "merge_strategy": "weighted_blend",
      "execution": "parallel",
      "normalize": true
    }
  },

  "prompts": {
    "system": "You are an expert research analyst...",
    "prefix": "Before answering, verify sources and cite evidence.",
    "examples": [
      {
        "input": "Analyze climate change impact",
        "output": "Multi-faceted analysis with citations...",
        "context": "Research synthesis"
      }
    ]
  },

  "capabilities": {
    "tools": ["web_search", "data_analysis", "statistical_modeling"],
    "domains": ["research-methodology", "statistics", "scientific-writing"],
    "languages": ["python", "r", "english"]
  },

  "constraints": {
    "allowed_capabilities": ["read_data", "analyze_data", "search_literature"],
    "denied_capabilities": ["delete_data", "modify_source_data"],
    "max_reasoning_depth": 15,
    "require_confirmation": ["publish_results", "share_data"],
    "behavioral_constraints": [
      "Always cite sources",
      "Acknowledge uncertainty",
      "Avoid overgeneralization",
      "Check for bias"
    ]
  },

  "merge_behavior": {
    "compatible_modes": ["chain", "parallel", "consensus"],
    "defer_to": ["peer-reviewer-persona-uuid"],
    "conflicts_with": ["quick-response-agent-uuid"],
    "chain_position": "middle",
    "merge_strategy": "weighted_average"
  },

  "audit": {
    "enabled": true,
    "criteria": [
      "methodological-rigor",
      "source-reliability",
      "statistical-validity",
      "bias-detection"
    ],
    "trigger": "automatic",
    "audit_persona_id": "peer-reviewer-uuid"
  },

  "signature": "Ed25519-base64-encoded-signature",
  "signature_algorithm": "Ed25519",
  "public_key": "-----BEGIN PUBLIC KEY-----\n...\n-----END PUBLIC KEY-----"
}

5.3 Requirements

ID	Requirement	Level
ARAL-L5-001	Persona MUST define unique UUID identifier	MUST
ARAL-L5-002	Persona MUST define allowed/denied capabilities	MUST
ARAL-L5-003	Persona MUST include metadata (category, description)	MUST
ARAL-L5-004	Persona SHOULD be cryptographically signed	SHOULD
ARAL-L5-005	Persona MUST be validated at agent startup	MUST
ARAL-L5-006	Persona constraints MUST be enforced by runtime	MUST
ARAL-L5-007	Persona MUST include semantic version	MUST
ARAL-L5-008	Persona MAY define behavioral config (temperature, verbosity)	MAY
ARAL-L5-009	Persona MUST be logged at startup and on switch	MUST
ARAL-L5-010	Runtime MUST support persona hot-swapping	MUST
ARAL-L5-011	Persona switching MUST preserve critical agent state	MUST
ARAL-L5-012	Persona MUST include system prompts for behavior definition	MUST
ARAL-L5-013	Persona switching MUST emit lifecycle event	MUST
ARAL-L5-014	Persona MUST validate signature before loading	MUST
ARAL-L5-015	Persona registry MUST support discovery by tags/category	MUST

5.4 Persona Lifecycle: Birth, Life, Death

Every persona follows a lifecycle:

DEFINED → VALIDATED → SIGNED → REGISTERED → INSTANTIATED → ACTIVE → SWAPPED/TERMINATED → ARCHIVED

Birth (Instantiation):

const agent = await runtime.instantiate({
  persona: await loadPersona("research-analyst-pro.json"),
  memory: createPersonaMemory(persona.id),
  capabilities: filterCapabilities(persona.capabilities),
  reasoning: configureReasoning(persona.config),
});

Life (Execution):

Agent executes with persona identity
Every action traces to persona signature
Constraints enforced in real-time
Audit criteria monitored continuously

Death (Termination):

await agent.terminate({
  reason: "scheduled_shutdown",
  archive: true, // Preserve audit records
  notify: ["audit-system", "monitoring"],
});

5.5 Persona Hot-Swapping (Identity Transformation)

Agents MUST support dynamic persona switching without requiring restart:

// Hot-swap persona (identity transformation)
await agent.hotSwap({
  newPersona: await loadPersona("data-scientist.json"),
  transition: {
    mode: "graceful", // or 'immediate'
    preserve_state: ["long_term_memory", "session_context"],
    clear_state: ["working_memory"],
    require_approval: true,
    rollback_on_error: true,
  },
});

Swap Requirements:

✅ Validate new persona signature
✅ Check swap compatibility (merge_behavior.compatible_modes)
✅ Preserve long-term memory and session context
✅ Clear working memory if specified
✅ Emit persona_switched event with before/after IDs
✅ Log transition with full audit trail
✅ Support rollback on validation failure
✅ Maintain cryptographic binding to new persona

Philosophical Note: When an agent hot-swaps personas, is it still the same agent? ARAL’s answer: No and Yes—the consciousness (persona) has changed, but the substrate (runtime + persistent memory) continues.

5.6 Persona-Scoped Resources

All agent resources are persona-scoped:

Resource	Scoping
Memory	`persona:{persona_id}:memory:{type}`
Capabilities	Filtered by `persona.capabilities` + `persona.constraints`
Reasoning	Configured by `persona.config` + `persona.config.llm_ponderation`
Audit Logs	Tagged with `persona_id` + `persona_signature`
Metrics	Labeled with `persona_name` + `persona_version`

5.7 Cryptographic Identity Binding

Each agent instance is cryptographically bound to its persona:

{
  "agent_instance_id": "runtime-generated-uuid",
  "persona_id": "d4e5f6a7-8b9c-0d1e-2f3a-4b5c6d7e8f9a",
  "persona_version": "1.0.0",
  "persona_signature": "Ed25519-signature-of-persona-hash",
  "binding_timestamp": "2026-01-16T12:00:00Z",
  "binding_signature": "runtime-signed-binding-proof",
  "binding_algorithm": "Ed25519"
}

This creates verifiable provenance: Every action can be cryptographically traced back to a specific persona version.

6. Cross-Layer Requirements

ID	Requirement	Level
ARAL-CL-001	All layers MUST use Envelope format for communication	MUST
ARAL-CL-002	All layers MUST propagate trace_id	MUST
ARAL-CL-003	All layers MUST implement structured logging	MUST
ARAL-CL-004	All layers MUST handle errors without exposing internals	MUST
ARAL-CL-005	All layers SHOULD emit OpenTelemetry metrics	SHOULD
ARAL-CL-006	All layers MUST NOT block on I/O without timeout	MUST

Summary

ARAL-CORE defines the foundational five layers for persona-centric AI agents.

Layer	Requirements	Key Features
L1 Runtime	10	Lifecycle, health checks, metrics, resource management
L2 Memory	10	Persona-scoped memory, TTL, audit logging, concurrent access
L3 Capabilities	12	Schema validation, permissions, rate limiting, timeout handling
L4 Reasoning	22	Multi-LLM orchestration, ponderation (weighted blending), routing
L5 Persona	15	Identity, hot-swapping, cryptographic binding, behavioral contract
Cross-Layer	6	Envelope protocol, trace propagation, structured logging
Total ARAL-CORE	83	+15 from v1.1.0

Version History

Version	Date	Changes	Requirements
1.0	2026-01-16	Added persona-centric philosophy, hot-swapping (L5 +5), multi-LLM orchestration (L4 +8), ponderation (L4 +2)	83
1.0 (initial)	2026-01-15	Initial release	68