ACGP-1011: Implementation Profiles¶

Status: Draft
Last Updated: 2026-01-08
Spec ID: ACGP-1011
Normative Keywords: MUST, SHOULD, MAY (per RFC 2119)

Abstract¶

This document defines standard implementation profiles for the Agentic Cognitive Governance Protocol (ACGP). Profiles provide pre-configured subsets of ACGP features optimized for specific use cases, reducing cognitive load for implementers and ensuring appropriate governance for different deployment scenarios.

Table of Contents¶

Introduction
Profile Definitions
Minimal Profile
Production-Lite Profile
Regulated Profile
Safety-Critical Profile
Profile Selection Guide
Conformance Requirements
References

1. Introduction¶

ACGP is a comprehensive protocol with many optional features. Not all deployments need all features. Implementation profiles define canonical subsets that:

Reduce complexity: Clear guidance on what to implement
Ensure appropriateness: Match governance depth to risk level
Enable interoperability: Common feature sets for similar use cases
Support progressive adoption: Clear upgrade path between profiles

1.1 Requirements Language¶

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

1.2 Relationship to Conformance Levels¶

Implementation profiles are orthogonal to conformance levels (ACGP-1009):

Conformance levels (Minimal/Standard/Complete) define implementation completeness
Implementation profiles define feature subsets optimized for use cases

A Minimal conformance implementation MAY use any profile. A Standard conformance implementation MUST support at least the Production-Lite profile features.

2. Profile Definitions¶

2.1 Profile Overview¶

Profile	Target Use Case	ACL Tiers	Eval Tiers	Trust Debt	Governance Contracts
Minimal	POC, development	0-1	0-1	No	No
Production-Lite	General production	0-3	0-2	Yes	Optional
Regulated	Finance, healthcare	0-5	0-3	Yes	Required
Safety-Critical	Autonomous systems	0-5	0-3	Yes + Enhanced	Required + Consensus

2.2 Feature Matrix¶

Feature	Minimal	Production-Lite	Regulated	Safety-Critical
Core Protocol
Interventions (all 6)	OK, Block only	All	All	All
CTQ Calculation	Simplified	Full	Full + Audit	Full + Audit
Tripwires	Basic	All categories	All categories	All + Real-time
Trust Debt	—	Basic	Full + Recovery	Enhanced
Blueprint Inheritance	—
ReflectionDB	In-memory	Local	Distributed	Replicated
Registry	—	Optional	Required	Required + MPA
Governance Contracts	—	Optional	Required	Required
HITL (Eval-3)	—	—		Required
Distributed Consensus	—	—	—	Required
HSM Integration	—	—	Optional	Required

3. Minimal Profile¶

3.1 Overview¶

The Minimal profile is designed for learning, development, and batch processing. It provides basic governance without production overhead.

profile: minimal
version: "1.0"

target_use_cases:
  - Learning ACGP concepts
  - Development and testing
  - Batch jobs and background processing
  - Non-latency-sensitive workloads

not_suitable_for:
  - Production interactive applications
  - User-facing systems
  - Latency-sensitive workloads
  - Systems requiring SLAs

3.2 Required Components¶

minimal_profile:
  acl_tiers: [0, 1]

  interventions:
    required: [ok, block]
    optional: [nudge, flag, escalate, halt]

  evaluation:
    eval_tiers: [0, 1]
    ctq_calculation: simplified    # May use static scores
    tripwires: basic               # Standard severity only

  storage:
    type: in_memory
    retention: session             # No persistence required

  security:
    transport: use_existing        # TLS optional for dev
    signatures: optional

  governance_contracts: disabled
  trust_debt: disabled

3.3 Simplified CTQ Calculation¶

Minimal profile MAY use simplified CTQ:

def simplified_ctq(trace):
    """
    Minimal profile CTQ calculation.
    May return static scores or use simple heuristics.
    """
    # Option 1: Static score based on ACL tier
    if trace.acl_tier in ["ACL-0", "ACL-1"]:
        return 0.8  # Assume low-risk agents are OK

    # Option 2: Simple heuristic
    if trace.reasoning and len(trace.reasoning) > 50:
        return 0.75

    return 0.5  # Default moderate score

3.4 Example Blueprint (Minimal)¶

id: dev/test-agent@1.0
version: "1.0.0"
description: "Development testing agent"
profile: minimal

scope:
  agent_tier: [ACL-0, ACL-1]

# Minimal CTQ - use defaults
ctq:
  profile: "minimal-defaults"

# Only essential tripwires
tripwires:
  - id: prevent_production_access
    condition: NOT environment == "production"
    on_fail:
      decision: block
      reason: "Cannot access production from dev agent"

scoring:
  thresholds:
    ok: 0.40
    block: 0.80

4. Production-Lite Profile¶

4.1 Overview¶

The Production-Lite profile is designed for general production deployments with balanced governance.

profile: production-lite
version: "1.0"

target_use_cases:
  - Production web applications
  - Customer service agents
  - Content generation systems
  - General enterprise automation

suitable_for:
  - Moderate risk tolerance
  - Interactive user-facing systems
  - Standard compliance requirements

4.2 Required Components¶

production_lite_profile:
  acl_tiers: [0, 1, 2, 3]

  interventions:
    required: [ok, nudge, flag, escalate, block, halt]

  evaluation:
    eval_tiers: [0, 1, 2]
    eval_tier_2: async              # LLM evaluation is async
    ctq_calculation: full
    tripwires: all_categories       # Standard, Critical, Severe

  trust_debt:
    enabled: true
    mode: basic                     # Accumulation + decay only
    recovery: false                 # No active recovery

  storage:
    type: local_persistent          # SQLite, PostgreSQL
    retention: 90_days

  security:
    transport: tls_1_3
    signatures: required_acl_3_plus

  governance_contracts: optional

  monitoring:
    metrics: required
    alerting: recommended

4.3 Evaluation Profile¶

Production-Lite uses a specific evaluation configuration:

evaluation_profile:
  name: "production-lite-eval"

  tier_0:
    mode: synchronous
    latency_budget_ms: 50
    checks:
      - schema_validation
      - auth_verification
      - critical_tripwires

  tier_1:
    mode: synchronous
    latency_budget_ms: 200
    checks:
      - policy_evaluation
      - rate_limiting
      - pattern_matching

  tier_2:
    mode: asynchronous              # Runs post-action
    latency_budget_ms: 5000
    checks:
      - llm_reasoning_quality
      - anomaly_detection
    on_failure: log_and_flag        # Don't block on async failure

  tier_3:
    mode: disabled                  # No HITL in production-lite

5. Regulated Profile¶

5.1 Overview¶

The Regulated profile is designed for financial services, healthcare, and other regulated industries.

profile: regulated
version: "1.0"

target_use_cases:
  - Financial services
  - Healthcare applications
  - Legal document processing
  - Insurance claims handling
  - Regulatory compliance systems

requirements:
  - Full audit trail
  - Human review capability
  - Compliance reporting
  - Data retention (7+ years)

5.2 Required Components¶

regulated_profile:
  acl_tiers: [0, 1, 2, 3, 4, 5]

  interventions:
    required: [ok, nudge, flag, escalate, block, halt]

  evaluation:
    eval_tiers: [0, 1, 2, 3]        # Full tiers including HITL
    ctq_calculation: full_audited   # All decisions logged
    tripwires: all_categories

  trust_debt:
    enabled: true
    mode: full                      # All features
    recovery: true                  # Active recovery enabled
    severity_weights: required

  storage:
    type: distributed
    retention: 7_years              # Compliance retention
    integrity: merkle_tree          # Cryptographic audit trail
    export: required                # Compliance exports

  security:
    transport: tls_1_3_mutual
    signatures: required_all        # Sign all messages
    hsm: optional

  governance_contracts: required

  human_in_the_loop:
    enabled: true
    timeout: 24_hours
    escalation_path: required

  registry:
    enabled: true
    min_trust_score: 0.7

  monitoring:
    metrics: required
    alerting: required
    siem_integration: recommended

5.3 Compliance Features¶

compliance:
  audit_trail:
    complete: true                  # Every trace, eval, intervention
    tamper_evident: true            # Merkle tree integrity
    exportable: true                # Standard formats (JSON, Parquet)

  reporting:
    real_time_dashboard: required
    periodic_reports: daily
    compliance_attestation: monthly

  retention:
    min_days: 2555                  # 7 years
    immutable: true
    geo_redundant: recommended

6. Safety-Critical Profile¶

6.1 Overview¶

The Safety-Critical profile is for autonomous systems where failures have significant consequences.

profile: safety-critical
version: "1.0"

target_use_cases:
  - Autonomous vehicles
  - Industrial automation
  - Critical infrastructure
  - Medical devices
  - High-value financial transactions

requirements:
  - Defense in depth
  - Distributed consensus
  - Hardware security
  - Real-time monitoring
  - Fail-safe defaults

6.2 Required Components¶

safety_critical_profile:
  acl_tiers: [0, 1, 2, 3, 4, 5]

  interventions:
    required: [ok, nudge, flag, escalate, block, halt]
    default_on_failure: halt        # Fail safe

  evaluation:
    eval_tiers: [0, 1, 2, 3]
    mode: synchronous_all           # All tiers sync for critical
    ctq_calculation: full_audited
    tripwires: all_real_time        # Real-time evaluation

  trust_debt:
    enabled: true
    mode: enhanced
    recovery: true
    severity_weights: required
    real_time_tracking: true

  storage:
    type: replicated                # Multi-region
    retention: 7_years
    integrity: merkle_tree
    sync: strong_consistency

  security:
    transport: tls_1_3_mutual
    signatures: required_all
    hsm: required                   # Hardware security module
    zero_trust: required

  governance_contracts: required

  distributed_consensus:
    enabled: true
    min_stewards: 3
    quorum: 2                       # 2/3 agreement required
    byzantine_fault_tolerant: true

  human_in_the_loop:
    enabled: true
    mandatory_for: [critical_risk, first_time_high_value]
    timeout: 5_minutes              # Shorter for critical

  fail_safe:
    default_decision: halt
    on_steward_failure: deny
    on_consensus_failure: halt

6.3 Distributed Consensus¶

Safety-critical actions require multiple steward agreement:

class DistributedConsensus:
    """
    Byzantine fault-tolerant consensus for safety-critical.
    """

    def evaluate(self, trace):
        # Query multiple stewards
        decisions = []
        for steward in self.steward_pool:
            try:
                decision = steward.evaluate(trace, timeout=100)
                decisions.append(decision)
            except TimeoutError:
                decisions.append(Decision.ABSTAIN)

        # Require 2/3 agreement
        quorum = len(self.steward_pool) * 2 // 3

        if count(decisions, Decision.HALT) >= 1:
            return Decision.HALT  # Any halt is halt

        if count(decisions, Decision.BLOCK) >= quorum:
            return Decision.BLOCK

        if count(decisions, Decision.OK) >= quorum:
            return Decision.OK

        # No consensus - fail safe
        return Decision.HALT

7. Profile Selection Guide¶

7.1 Decision Tree¶

Start
  |
  +- Is this for production? --- No --► Minimal Profile
  |
  | Yes
  |
  +- Is this regulated industry? --- Yes --► Regulated Profile
  |   (finance, healthcare, legal)
  |
  | No
  |
  +- Are failures safety-critical? --- Yes --► Safety-Critical Profile
  |   (autonomous, infrastructure)
  |
  | No
  |
  +-► Production-Lite Profile

7.2 Profile Comparison¶

Criteria	Minimal	Prod-Lite	Regulated	Safety-Critical
Setup Time	Minutes	Hours	Days	Weeks
Latency Overhead	<10ms	<50ms	<100ms	<500ms
Cost (monthly)	$0	$100-500	$1K-5K	$10K+
Compliance Ready	No	Partial	Yes	Yes+
Audit Trail	None	Basic	Full	Full + Tamper-proof
Recovery Options	None	Decay only	Full	Enhanced

7.3 Upgrade Path¶

Minimal --► Production-Lite --► Regulated --► Safety-Critical
                |                    |
                |                    +-- Add: HITL, full registry,
                |                           compliance features
                |
                +-- Add: Trust debt, all tripwires,
                        governance contracts, persistent storage

8. Conformance Requirements¶

8.1 Profile Conformance¶

Implementations claiming a profile MUST:

Implement ALL required features for that profile
Meet latency requirements for specified eval tiers
Provide storage with required retention
Support all specified intervention types
Pass profile-specific conformance tests

8.2 Profile Declaration¶

Implementations MUST declare their supported profiles:

{
  "implementation": "acme-steward",
  "version": "2.1.0",
  "profiles_supported": [
    {"name": "minimal", "conformant": true},
    {"name": "production-lite", "conformant": true},
    {"name": "regulated", "conformant": true, "note": "Requires HSM addon"},
    {"name": "safety-critical", "conformant": false}
  ]
}

8.3 Profile Validation¶

Each profile has a conformance test suite:

# Run profile conformance tests
acgp-conformance --profile=production-lite --endpoint=https://steward.example.com

# Expected output
Profile: production-lite
Tests: 47/47 passed
Latency: Tier-0: 23ms, Tier-1: 145ms, Tier-2: 2340ms (async)
Trust Debt: Enabled, decay verified
Status: CONFORMANT

9. References¶

Normative References¶

ACGP-1000: Core Protocol Specification
ACGP-1009: Conformance Requirements
ACGP-1010: Governance Contracts

Informative References¶

ACGP-1002: Architecture Specification
ACGP-1004: Blueprint Specification
ACGP-1007: Security Considerations

End of ACGP-1011