Architecture

Quality & Governance

How MARIA OS enforces quality, resolves conflicts, and maintains governance integrity across autonomous agent teams.

Decision OS Studio

Design responsibility as executable phases

Judgment, accountability, and execution — explicitly separated

Decision OS Studio
Current Responsibility Scope:Expense Zone→ Operational Responsibility
Responsibility Layers
Global Policy12
Company-wide Responsibility
Finance Ops24
Department Responsibility
AP Process18
Process Responsibility
Expense Zone8
Operational Responsibility
Judgment PhaseResponsibility PhaseExecution PhaseEntryInputValidateLogic CheckEstimateRisk ScoreGATEHuman RequiredResponsibility BoundaryExecutei@ / ops@Company Action

This graph represents responsibility phases, not process flow

Conflict Resolution

Conflict Resolution as a Common Language

MARIA OS treats conflicts as assets to be structured, not failures to hide.

Conflict Classification

Hard: Mutually exclusive
Soft: Different scope
Temporal: State changed
Decision Gate
Record
TargetContract
Candidates2
ImpactHigh
OwnerCFO

MARIA OS aligns judgment with responsibility at the OS level.

Continuous Re-evaluation
Decision Gating

Transform Conflicts into Structured Decisions

When agents disagree, the system does not average their outputs. It surfaces the contradiction, attaches evidence, and routes to the appropriate resolution authority.

Intake
Detect
Classify
Gate
Record
HardHuman Review
SoftAuto-merge
TemporalVersion compare
Human Gate
IrreversibleSTOPPED
ApprovalCFO
ResponsibilityLocked
Audit Trail
14:32:01Detected
14:32:02Classified
14:32:03Gate triggered
14:35:47Approved
Decision → Re-detect → Feedback Loop
Self-Improving

Tick-Driven Quality Loops

Add
Detect
Evaluate
Approve
Feedback
Knowledge Re-evaluation
Phase Consistency
Impact Rescoring
Human Feedback

3

Active

12

Today

2

Pending

Cycles, not single events

Verification

Phase-Locked Quality Enforcement

Synchronization
Active
Judgment
Responsibility
Verification
Alignment
Continuous
Hierarchy
Agent
11/12
Zone
4/4
Planet
1/1

Rule-based

Constraints

Evidence

Patterns

Continuous verification across hierarchy

Risk Assessment

Dynamic Impact Rescoring

Cycle: 1,247

Financial Risk+0.6
6.8/ 10
Operational Impact-0.9
4.2/ 10
Compliance Score0.0
3.5/ 10
Recent
2m agoKnowledge added+0.3
5m agoConflict resolved-1.2
12m agoDecision made+0.8

-23%

Reduction

847

Day

99.2%

Accuracy

Continuous rescoring

Oversight

Human Approval as Architecture

Approve
New Info
Re-detect
Re-review

Contract #1247

2h ago

approved

Budget #892

30m ago

re-review

Hiring #445

Just now

pending

94%

First-pass

18%

Re-review

2.4h

Avg. Time

Continuous feedback

Quality

Seven Pillars of Reliability

Quality Gates

Automated boundaries

Iterative Loops

Continuous cycles

Impact Rescoring

Dynamic evaluation

Evidence Trails

Full traceability

Human Feedback

Approval integration

Transparent UX

Visual clarity

Quality KPIs

Continuous monitoring

Gate Board
Interpret
Propose
Decide
Act

78:22

Auto vs Human

94%

Alignment

Verified, traceable, continuously improving

Architecture

Agent Skill Matrix

Maturity CurveL1 → L5
1
2
3
4
5
L3
ReactiveAuditable

Agent maturity through loops, gates, and responsibility

Parallel Quality 1/4
Team Parallel Orchestration

Parallel Teams, Zero Collisions

Agents work in parallel by team, but never collide. Work Units are claimed with exclusive locks before execution starts. Speed scales. Conflicts don't.

Pipeline Flow

Intake

Request

Scope

Risk

Router

Team Selector

Work Unit Splitter

Lock & Gate

Claim Lock

Phase Gate

Policy Check

Parallel Teams
4 active
Dev

src/api/*

active
QA

tests/e2e/*

active
Docs

docs/api/*

waiting
Ops

deploy/*

locked
collision barrier
Lock Trace
dev-agent-01
14:32:01

src/api/auth

qa-agent-03
14:32:02

tests/e2e/login

docs-agent-01
14:32:03

docs/api/auth

Quality Gate
Checklist
Policy
Evidence
Single Gate - Never Parallel

Work is parallelized. Quality judgment is not. All outputs pass through a single gate.

Claim Lock → Phase Gate → Merge Gate

Parallel Quality 3/4
Sharded Agents & Aggregation

Scale Agents, Elevate Quality

When you scale agents, quality often averages down. MARIA OS uses shard-level validation and aggregation reviewto ensure quality compounds up, not down.

Agent Cluster
Scheduler
Shard Splitter|Queue|Wake Bus

Agent A

shard-001

Agent B

shard-002

Agent C

shard-003

Approval wait

Agent D

shard-004

Agent E

shard-005

Data pending

Agent F

shard-006

Sleep Pool
C
E
Auto-wake
State Timeline
Claim
Run
Validate
Sleep
Wake
Merge
Aggregation Hub
Consistency Check
pass
Completeness Check
pass
Policy Match
pass
Outlier Detection
warning
Outliers return to agent for rework

Shard-level validation → Aggregation review → Final merge. Quality compounds, not averages.

Sleep without penalty. No deadline pressure. No quality shortcuts.

Parallel Quality 4/4
The Architecture Guarantee

Fast and Rigorous. By Design.

Parallelization is for execution throughput.
Quality judgment is always serial and centralized.

This architectural separation makes quality degradation physically impossible.

Core Guarantees
Work is Locked

Claim before execution

0 collisions
Quality is Gated

Single checkpoint

100% reviewed
Domains are Isolated

Separate standards

0 cross-contamination
Outputs are Aggregated

Compound quality

Outliers rejected
Quality Flow
Parallel Execution
Shard Validation
Aggregation Review
Quality Gate
Final Output
Live Metrics

Parallel Teams

4active

Work Units Locked

847today

Quality Gate Pass Rate

94.2%

Collision Events

0ever
Parallel and Rigorous
Teams work in parallel with exclusive locks
Industries run concurrently with isolated gates
Agents scale with aggregation review
Quality judgment remains centralized

Speed without compromise. Parallelism without chaos.

This is how MARIA OS scales safely.

Parallel execution + Centralized quality = Safe scale

Section 1
For Data Scientists & Engineers

Database-Native AI Orchestration

A database-native orchestration system. Actions from explicit state, not ephemeral context.

Properties
Deterministic

Same input, same output

Reproducible

Replay any decision

Queryable

SQL over decisions

Clear State

No hidden context

Query it. Reason about it. Scale it safely.

DecisionRecord
id:dec_8f2a
state:proposed
risk:R2
evidence:pending
State Transitions
proposedvalidateapprovalapprovedexecutedrecovere
Query Panel
SELECT * FROM decision_records
WHERE state = 'approved'
-- Results

AI autonomy designed like a distributed system, not a chatbot.

Section 2
Data Architecture

OLTP × Document Store Hybrid

Two-layer persistence for AI decisions. OLTP stores the source of truth ledger with ACID guarantees. Document Store projects operational views for real-time UI. CDC streams keep them synchronized.

Public SpecContract

DecisionRecord Contract

idstring
stateenum
risk_tierR1-R3
intentstring
constraints_hardarray
values_layeredobject
evaluation_summaryobject
evidence_refsURI[]
policy_versionstring
execution_envelopeobject
timestampsbitemporal

Invariants

No record update, no side effect
Append-only journal
Bitemporal timestamps
Data Plane
OLTP StoreSource of Truth
DecisionJournal
DecisionCurrent
GateEvaluations
ExecutionEnvelopes
CDC
Change Streams
Document StoreProjection
UI read model
Universe nodes
Live status
fast pagination

State Machine

prop
→
vali
→
appr
→
appr
→
exec
→
reco
Control PlanePrivate
Value Scan Neural
••• value extraction + uncertainty •••
Policy Engine
••• gates + thresholds •••
Execution Engine
••• reversible plans •••
Evidence Bundle
••• artifacts + redaction •••

What we publish: contracts and invariants.
Reproduction requires accumulation.

Fix truth in OLTP, render universe in Document Store.