● AGENTIC DEVELOPMENT LIFECYCLE

How Arka AI Works

Every Arka AI deployment follows the ADLC — 8 structured steps from discovering your operational intent to productizing proven missions across your PE portfolio. Governed at every step. Evidenced at every cycle. Fail-closed always.

1. Discover
2. Blueprint
3. Compose
4. Simulate
5. Execute
6. Evidence
7. Optimize
8. Productize

From intent to outcome — continuous, governed, replayable

The Lifecycle

8-Step Sovereign Mission Lifecycle

Every mission — from billing reconciliation to tax filing — follows the same governance pipeline. No worker executes until it has cleared Simulation, co-signed a Baseline, and entered Shadow mode. Autonomy is earned, not assumed.

1
Discover Mission
Understand the operational problem — define the intent
Arka Outcome Engineer Led

The Discover phase is led by your Arka Outcome Engineer. Together we map the operational problem, define what success looks like in measurable terms, and produce the Baseline Measurement Package — the co-signed document that anchors every subsequent Capacity Created measurement.

No Outcome Contract activates until the Baseline is co-signed. You define the intent; we build to it.

Deliverables
  • ✦ Baseline Measurement Package (co-signed)
  • ✦ ROI analysis with Capacity Created projections
  • ✦ Stakeholder alignment map
  • ✦ Compliance & data residency requirements documented
  • ✦ Mission scope definition (2 services for Pilot)
2
Define Blueprint
Author the mission — workflow DAG, HITL rules, evidence requirements

The Blueprint is a declarative YAML definition of the mission. It defines which capabilities are invoked, how the workflow DAG is structured, which steps require HITL approval, and what evidence events must be produced at each stage.

Blueprints are version-controlled and peer-reviewed. A mission cannot move to Compose without a signed-off Blueprint.

Blueprint Defines
  • ✦ Mission metadata (name, version, tenant, domain)
  • ✦ Capability bindings (which workers, which APIs)
  • ✦ Workflow DAG (task order, dependencies)
  • ✦ HITL escalation rules (which decisions require human)
  • ✦ Evidence event requirements per step
  • ✦ SLA constraints and budget limits
3
Compose Runtime
Bind governance policies — compile the governed mission runtime

Compose binds the Blueprint to OPA governance policies, validates worker qualifications against the Trust Progression State Machine, and compiles the mission runtime with the Sovereign LLM Router configured. This step is deterministic — no runtime changes after Compose is sealed.

Compose Produces
  • ✦ OPA policy bundle (compiled, Rego)
  • ✦ Worker qualification assertions
  • ✦ Sovereign LLM Router configuration
  • ✦ Runtime manifest (ECDSA-SHA256 signed)
  • ✦ ActionGateway 9-gate policy bindings
4
Simulate & Validate
Synthetic data testing, red team, replay — no production impact

Before any mission touches production data, it goes through a full Simulation phase. Synthetic datasets designed to cover edge cases, drift conditions, and adversarial inputs. No production data is used. No production systems are touched.

A mission that fails Simulation does not proceed — the gap is documented, the Blueprint revised, and the cycle repeats.

Validation Gates
  • ✦ Synthetic data test suite (edge cases, drift)
  • ✦ Red team & adversarial input testing
  • ✦ Deterministic replay testing
  • ✦ OPA policy enforcement validation
  • ✦ HITL escalation path verification
  • ✦ Evidence event completeness check
Trust Progression State Machine

Autonomy Is Earned, Not Assumed

Before a worker executes autonomously on production data, it must progress through nine trust states. Shadow mode is mandatory: workers execute against real data, but outputs are discarded and compared to human baselines before any autonomous execution is approved.

Draft
Mission defined, not yet compiled
Qualifying
Simulation in progress; must pass all gates
Shadow ★
Executes on production data; outputs discarded & compared to human baseline before approval
Qualified
Shadow outputs match baseline; ready for autonomous
Autonomous
Executes & commits; all 9 gates active every cycle
Active
In production, meeting SLAs, producing evidence
Audited
Completed cycle; full audit record sealed
Archived
Retired or superseded; record immutable
Quarantined ✕
Any gate failure → immediate halt, no partial execution
Draft
Qualifying
Shadow ★
Qualified
Autonomous
Active
Audited
Any gate fail → Quarantined
5
Execute & Govern
Temporal orchestration, 9 Gates, HITL approvals, autonomous execution (bounded)

Live execution runs through the ActionGateway — 9 governance gates evaluated synchronously before any state-changing action. Every mutation request is identity-verified, role-checked, OPA-validated, budget-constrained, SLA-bounded, evidence-complete, HITL-approved, audited, and ECDSA-sealed before execution.

If any gate fails, execution drops to QUARANTINED. The worker does not retry, does not partial-execute, and does not proceed.

ActionGateway — 9 Gates
1Identity
2Role
3Policy (OPA/Rego)
4Budget
5SLA
6Evidence
7Approval (HITL)
8Audit (pre-commit)
9Seal (ECDSA-SHA256)
6
Evidence & Forensics
Cryptographic sealing, immutable ledger, full replayability

Every mission cycle produces a sealed Proof Pack — ECDSA-SHA256 signed, written to the immutable Evidence Ledger. The Proof Pack captures full input data, all actions taken, every HITL approval, elapsed time, and the final sealed output. Every Proof Pack is independently replayable bit-for-bit.

The Evidence Package — delivered to the client per cycle — is your structured record of what was done, when, by whom, and what result was produced. It is the basis for the Quarterly Value Review.

7 Canonical Evidence Events
1mission_spawned
2intelligence_generated
3policy_validated
4approval_captured
5mutation_executed
6proof_pack_sealed ← ECDSA signed
7financial_realized
7
Optimize & Qualify
Quarterly Value Review, Capacity Created measurement, continuous improvement
Quarterly Value Review

The Quarterly Value Review is the commercial conversion gate at the end of the 4-month Pilot. Your Arka Outcome Engineer presents Capacity Created measurements against the co-signed Baseline. Both parties review the Evidence Ledger record. Conversion is evidence-based — not a sales call.

Ongoing optimization runs continuously: SLA monitoring, drift detection, root cause analysis, and worker score updates feed back into the Trust Progression — workers that underperform are demoted to Qualifying or Quarantined.

Capacity Created Metrics
  • → Filing prep effort (hrs/filing) vs. baseline
  • → Reconciliation cycle time (days) vs. baseline
  • → Manual hours avoided per cycle
  • → Exception identification accuracy (%)
  • → Combined hours avoided (monthly)
  • → Worker error rate vs. human baseline
8
Productize & Scale
Mission Library, portfolio adapter model, PE rollout
Portfolio Multiplier

Proven missions are packaged into the Mission Library — versioned, certified blueprints with production-validated evidence records. Each new portfolio company gets the same Mission Library with a new adapter layer. No rebuild. The governance pipeline, evidence architecture, and worker qualifications are already proven.

This is the PE Portfolio Multiplier: the more acquisitions, the stronger the platform — shared intelligence, shared compliance, shared evidence standard.

Scale Mechanics
  • ✦ Mission Library: versioned, certified, reusable
  • ✦ New PortCo = same missions + adapter (~10 days)
  • ✦ Multi-tenant: data isolation, per-tenant policy
  • ✦ PE Portfolio Dashboard: consolidated FP&A view
  • ✦ Expansion by change order — no re-contract
  • ✦ Compound intelligence across every acquisition
Architectural Principle

Zero-LLM Control Plane

Every governance decision in Arka AI — identity, role, policy, budget, SLA, evidence, HITL, audit, seal — is made by deterministic OPA/Rego. No large language model is in the governance path.

LLMs operate only at the user-facing intelligence layer — generating analysis, drafting content, identifying patterns. They never control access to systems, approve financial transactions, or make policy decisions. The result: a worker that cannot hallucinate on a tax filing, and governance outcomes that are auditable, repeatable, and bit-for-bit replayable.

Governance Layer (Zero-LLM)

OPA/Rego policy engine, ActionGateway 9 gates, identity/role/budget/SLA enforcement, HITL orchestration, ECDSA sealing. Deterministic. Auditable. No inference.

Intelligence Layer (LLM — bounded)

Sovereign LLM Router, domain-specific model selection, intelligence generation, pattern analysis. Operates within policy-enforced boundaries. Never controls execution or approvals.

Evidence Layer (Immutable)

Dual-ledger architecture. Evidence Ledger + Financial Ledger. Hash-chained, append-only, OSCAL-native. 99.99% read availability. Full forensic replayability on every Proof Pack.

Foundation Layer

Data Readiness Certification

Before any governed mission executes, Arka certifies your data into readiness tiers. Only Gold-tier data — scored, validated, and confidence-cleared — can drive autonomous action. A mission operating on uncertified data fails the evidence gate before any action is taken.

Bronze — Ingested

Raw source data is ingested, cleaned, deduplicated, and schema-validated. Bronze tier establishes that data exists and is structurally sound — but it is not yet trustworthy enough to act on.

✓ Deduplication
✓ Schema validation
✓ Source lineage tagged
Silver — Enriched

Bronze data is joined, enriched with derived signals, and normalized into the feature sets mission workers expect. Silver is computationally ready but not yet scored for mission action.

✓ Cross-system joins
✓ Derived signal computation
✓ Feature normalization
Gold — Mission-Ready ★

Silver data is scored, classified, and confidence-thresholded. Gold tier passes the evidence provenance check (Gate 6) and is cleared for autonomous mission action. Below threshold: routes to HITL.

✓ Confidence-scored (≥threshold)
✓ Suppression overlay checked
✓ Evidence Gate cleared

Evidence Gate enforced: A mission worker that attempts to act on Bronze or Silver-tier data fails Gate 6 (Evidence provenance) before any action executes. Data readiness is not advisory — it is a hard governance gate.

Complex Revenue Programs

Sequential Sub-Mission Architecture

Complex revenue programs require more than one mission. Arka structures multi-phase programs as governed sub-mission chains — each phase gates on the previous phase's sealed Proof Pack before it can spawn. No phase skipping. No governance gap between phases.

A
Phase A — Segmentation & Scoring

Accounts are scored and segmented by propensity, risk, and commercial profile. Phase A seals a Proof Pack. Phase B cannot spawn until Phase A's evidence record is validated at the entry gate.

B
Phase B — Adoption & Engagement

Targeted engagement actions execute against the scored segment from Phase A. Usage gap closures, activation nudges, and onboarding follow-ups — each action governed by the 9-gate pipeline. Phase B seals its own Proof Pack before Phase C can begin.

C
Phase C — Conversion & Commercial Action

Revenue motions execute: renewal outreach, expansion offers, contract triggers. High-value decisions route to HITL. All actions produce sealed evidence before the phase closes. Phase C cannot proceed without Phase B's evidence record at its entry gate.

D
Phase D — Retention & Capacity Created

Post-conversion health monitoring, churn signal tracking, and outcome measurement. The combined Proof Pack chain — Phase A through D — forms the complete evidence record for the Quarterly Value Review and Capacity Created measurement.

Each phase:
Proof Pack sealed
Entry gate validated
Next phase spawns
Evidence chain complete

Ready to map your operational intent?

Your Arka Outcome Engineer starts with a 30-minute discovery call. We identify where Capacity Created is achievable, scope a 4-month Pilot, and co-sign your Baseline — before any autonomous execution begins.