Overview Architecture Theory of Operation
● THEORY OF OPERATION
FOR TECHNICAL EVALUATORS

How ARKA AI Works

The Northstar Architecture: a closed-loop system that ingests enterprise signals, governs every decision through policy, and seals outcomes into cryptographic evidence.

THE GOLDEN DATA LOOP

The Northstar Architecture

Six interconnected zones form a continuous feedback loop. Data flows clockwise from Source through Intelligence to Proof, then feeds back to refine future decisions.

ZONE 1

The Source

Flexible ingestion with variable path normalization and semantic guards. Connects to your existing CRM, customer success, and data warehouse systems.

Signal → Normalize → Validate → Ingest
ZONE 2

The Ground Truth

Pydantic schema validation and Scenario Fixture Catalog ensure data integrity. Account360 snapshots are Tier-1 only.

Schema → Validate → Fixture → Lock
ZONE 3

The Seeding

Tier-1 Seed Runner loads Mission Blueprints and golden worker fleet from frozen snapshots. Deterministic and repeatable.

Blueprint → Seed → Workers → Fleet
ZONE 4

The Intelligence

Decision Fabric orchestrates Workers (Revenue Analyst, Churn Guard) through enterprise policy gates and the Trust Anchor.

Worker → Analyze → Policy → Propose
ZONE 5

The Realization

Durable Workflow Orchestration with HITL Governance Dashboard. Human approvals captured, execution proceeds under governance.

Approve → Execute → Measure → Record
ZONE 6

The Proof

Financial Outcome Ledger and Attribution Graph produce immutable evidence. Outcomes feed back into the loop.

Seal → Notarize → Attribute → Feedback

The Golden Data Loop: Outcomes from Zone 6 feed back into Zone 1 as refined signals, creating a continuous improvement cycle. Every iteration is governed, sealed, and auditable.

FORENSIC CHAIN

The 7-Event Canonical Evidence Chain

Every mission emits exactly 7 hash-chained events. Tampering with any event or its order invalidates the audit trail.

1
mission_started DETERMINISTIC · SEALED

Mission instance created, payload validated, immutable mission ID assigned.

2
context_validated DETERMINISTIC · SEALED

Data sources verified, Ground Truth confirmed sufficient. Context hash sealed.

3
intelligence_generated STABILIZED · SEALED

Worker produces Decision Proposal. LLM stabilized via temperature=0 and model pinning. Prompt hash captured.

4
governance_evaluated DETERMINISTIC · SEALED

Policy engine evaluates decision against rules (churn lock, confidence floor, thresholds). Policy result hash recorded.

5
approval_captured HUMAN-GATED · SEALED

Human approves or rejects. Sealed with who, when, and justification. Phase-1: Required for all decisions.

6
outcome_recorded DETERMINISTIC · SEALED

Financial Outcome Ledger entry created. Hash-chained to previous outcome. Idempotency-keyed to prevent double-billing.

7
mission_finalized DETERMINISTIC · SEALED

Mission marked complete. Evidence chain closed with final integrity root. Decision Lineage record created.

ANALYTICS & INTELLIGENCE

The Intelligence Layer

Insights derived exclusively from governed evidence. No ungoverned data enters the intelligence layer.

Decision Mining

  • Pattern Discovery: Recurring decision paths
  • Bottleneck Detection: Slow approval stages
  • Conformance Check: Actual vs. blueprint path
  • Optimization: Auto-approval candidates

Digital Twins

  • Process Replica: Model customer operations
  • What-If Simulation: Test policy changes safely
  • Coverage Metrics: Process model completeness
  • ODG Integration: Linked to decision graphs

Observed Decision Graph

  • Actual vs. Expected: Path divergence map
  • Worker Contribution: Who drove what
  • Confidence Map: Scores across the path
  • Process Truth: Bottleneck identification
OUTCOME NAVIGATION

From Governed Execution to Active Steering

Most AI platforms stop at execution and proof. ARKA AI goes further — actively steering KPIs toward contracted targets through deterministic drift detection, governed convergence, and continuous re-baseline.

DETECT

Drift Detection

Continuous measurement of KPI deviation from locked baselines. The platform detects when outcomes begin drifting from contracted targets — before humans notice.

drift = (current − baseline) / |baseline|
STEER

Deterministic Steering

The NavigatorKernel computes optimal parameter adjustments to steer KPIs back on track. Every adjustment is policy-bounded, deterministic, and requires governance approval.

KPI(t+1) = KPI(t) × (1+δ), δ bounded by policy
CONVERGE

Convergence Proof

Mathematical proof that the gap between current KPIs and contracted targets is monotonically decreasing. If convergence stalls, the platform surfaces it for governed re-baseline.

J(t₀) > J(t₁) > J(t₂) — monotonic proof

Outcome Navigators (ON): Domain-specific modules that apply this steering across verticals — ON: Revenue · ON: Compliance · ON: Legal · ON: Logistics

COMMERCIAL SPINE

Outcome Feasibility & Economics

The Commercial Spine is designed for outcome-aligned economics: billing infrastructure tied to realized, verified value — not AI usage or seat counts.

Outcome Contracts

Formal declarations of baseline, target, confidence band, and measurement period. Immutable once locked.

baseline_arr: [measured]
target_arr: [agreed]
confidence_band: [calibrated]
pricing: OUTCOME_ALIGNED

Confidence-Based Fee Grading

The confidence matrix enables fee structures that scale with decision quality. Built to support outcome-aligned economics.

High ConfidenceFull fee earned
Medium ConfidenceScaled fee
Low ConfidenceReduced fee
Below ThresholdNo fee charged

Evidence Bundles (Proof Packs)

  • • Baseline Snapshot (locked at mission start)
  • • Outcome Evidence (measured at mission end)
  • • Delta Calculation (verified change)
  • • Cryptographic Digital Seal (Vault-backed)
  • • Attribution Graph (DAG credit assignment)

Engagement Progression

1. OBSERVE

Read-only analysis · Discovery mode

2. EXECUTE

Governed mode · Outcome-aligned billing

3. SCALE

Multi-domain expansion · Enterprise governance

Key Architectural Principles

Determinism ≠ Outcome Correctness

Determinism means execution reproducibility. Given the same inputs, you get the same process. LLM outputs are stabilized, not guaranteed identical.

Fail-Closed, Not Fail-Open

In production, if any required service (OPA, Temporal, or Kafka) is unavailable, the platform refuses to operate or start. It does not fall back, degrade, or simulate operations without full integrity guards.

UI Is Non-Authoritative

The UI is a read-only shadow of the backend. All authority flows through the API and governance primitives.

Signal / Decision Boundary

Signals are ephemeral data with TTL. Decisions are governed and persistent. Signals must cross the boundary via the Decision Fabric before influencing outcomes.

Developer Guide

Platform Sovereign Lifecycle

How developers create and execute governed operations on ARKA AI.

STEP 1
Define Entities

Declare your Ground Truth schema via Pydantic or JSON Schema. Define the 'Account360' structure authorized for Worker access.

STEP 2
Deploy Workers
Select role-scoped Workers from the Fleet. Assign Authority Levels (Observe vs. Recommendation) to each worker identity.
STEP 3
Draft Mission

Instantiate a versioned Mission Blueprint. Bind the target outcome, success metrics, and human approval gates to the execution path.

STEP 4
Execute & Audit

Trigger execution via Signal Hub. Monitor the 7-event evidence chain and verify outcome realization on the Sovereign Ledger.

Explore API Documentation →

Ready to See It in Action?

Most enterprises begin with ARKA Advisors to design governance frameworks before platform deployment.

Consult with Advisors View Architecture →

New here? Start with the Platform Overview