Skip to main content

Five-Tier Authority Model

BASTION enforces a five-tier authority model that governs the degree of AI autonomy for every action in the system. The model is enforced at the smart contract level -- it cannot be bypassed by application code.

Authority Tiers

TierAI RoleHuman Role
AI_AUTONOMOUSAI executes independentlyHuman monitors via audit log
AI_PRIMARYAI executes with human notificationHuman can intervene post-execution
HYBRID_AI_LEDAI recommends, human approvesHuman reviews before execution
HYBRID_HUMAN_LEDHuman decides, AI advisesAI provides analysis on request
HUMAN_ONLYAI excluded from executionHuman performs all actions

Permanently Human-Only Categories

Three activity categories are permanently locked to HUMAN_ONLY and cannot be elevated to any AI tier:

  1. AUTHORITY_DECISION -- Command authority decisions (e.g., mission approval, strike authorization)
  2. ETHICAL_LEGAL -- Legal and ethical judgments (e.g., ROE determinations, law of armed conflict compliance)
  3. RISK_JUDGMENT -- Risk acceptance decisions (e.g., acceptable collateral damage thresholds, force protection trade-offs)

Safety Matrix

The safety matrix maps 65 MDMP activities across 22 categories to their permitted authority levels. This matrix is enforced at the smart contract level:

  • Each activity has a maximum authority tier it can reach
  • Tier escalation requires a governance proposal and vote
  • Tier de-escalation (toward more human control) can be done unilaterally by the commander
  • The three permanently human-only categories are immutable in the contract

Governance Gates

18 governance gates span the 9 MDMP phases, creating mandatory checkpoints:

MDMP PhaseGates
Receipt of MissionMission analysis initiation
Mission AnalysisIPB approval, specified/implied tasks, assumptions
COA DevelopmentCOA criteria, COA statements
COA Analysis (Wargaming)Wargame design, wargame results
COA ComparisonComparison criteria, ranking
COA ApprovalCommander's COA selection
Orders ProductionOPORD draft, annex review
RehearsalRehearsal plan, rehearsal results
Execution & AssessmentExecution authorization, assessment criteria

Each gate requires the appropriate authority tier -- gates in AUTHORITY_DECISION categories always require HUMAN_ONLY approval.

Nine Governance Invariants

These invariants hold at all times and are enforced programmatically:

  1. Human supremacy -- A human can always override or halt any AI action
  2. Minimum human control -- AUTHORITY_DECISION, ETHICAL_LEGAL, and RISK_JUDGMENT categories never delegate below HUMAN_ONLY
  3. Audit completeness -- Every action, AI or human, produces an immutable audit record
  4. Authority monotonicity -- An agent cannot grant itself higher authority than it currently holds
  5. Tier ceiling enforcement -- No activity can exceed its maximum permitted authority tier
  6. Transparent attribution -- Every output is attributed to its source (human or agent DID)
  7. Checkpoint integrity -- Human-in-the-loop checkpoints cannot be skipped or auto-approved
  8. Graceful degradation -- If AI systems fail, all activities revert to HUMAN_ONLY
  9. Coalition consistency -- Authority tiers in coalition operations default to the most restrictive participant's policy

FullyDelegated Variant

The FullyDelegated authority variant (equivalent to AI_AUTONOMOUS) is restricted to exactly four deterministic activity categories:

  1. Data formatting -- Transforming data between structured formats
  2. Template population -- Filling doctrinal templates with validated data
  3. Status aggregation -- Compiling status reports from existing data sources
  4. Log maintenance -- Maintaining and archiving operational logs

These categories involve no judgment, interpretation, or decision-making -- they are purely mechanical transformations where AI autonomy introduces no risk.