Platform Architecture
The platform your system runs on.
Simpliciti isn't a chatbot, a copilot, or a prompt chain. It's a structured execution platform where deterministic business rules and bounded AI intelligence work together. Every action is recorded.
Three layers. Clear boundaries.
The platform separates what must be deterministic from what benefits from intelligence. Most transitions never touch AI. When they do, the AI operates within declared constraints. It cannot invent steps, skip rules, or override human decisions.
The Runtime
Deterministic execution. Enforces step ordering, policies, and state transitions. Handles 70–80% of all transitions with zero AI involvement.
The Shepherd
Bounded AI intelligence. Invoked only at ambiguous junctions. Cannot invent steps, bypass policies, or override human decisions.
The Context Graph
Process-level learning. Aggregates patterns from every run. Observes what humans actually do, then infers why. Intelligence that compounds.
Policies evaluate → Shepherd reasons (if needed) → Context Graph learns
Layer 01
The Runtime
Every process runs on a workflow engine that enforces step ordering, manages state transitions, and guarantees that policies are evaluated before any AI reasons about what to do next. The runtime handles branching, parallel execution, retry logic, and human review gates.
Deterministic Fast Path
A human clicks ‘Approve.’ A skill completes with one valid next step. The runtime handles it directly. No AI, no latency, no cost. This covers 70–80% of all transitions.
Policy Enforcement
Policies are hard business rules. They fire unconditionally. No AI component can override or skip a policy. The runtime evaluates policies before the Shepherd ever sees the data.
Configuration Profiles
Same process topology, different behavior. A California police background check and a Texas fire department check share the same process, but configuration profiles adjust thresholds, required sections, and compliance frameworks.
Process: Background Investigation
Status: Running
Step 1 ✓ Document Intake [completed]
Step 2 ✓ Document Parsing [795/800 parsed]
Step 3 ● Section Generation [in progress]
Step 4 ○ Investigator Review [pending]
Step 5 ○ Legal Review [pending]
Step 6 ○ Manager Approval [pending]
Policies enforced: 3
Shepherd invocations: 0
AI involvement: 0% of transitionspolicy:
id: "require_senior_review_high_flags"
type: routing
priority: 100
trigger:
when:
field: "compliance_check.flagged_items.count"
operator: ">"
value: 10
enforce:
require_skill: "senior_legal_review"
override_role: "senior_legal_reviewer"
audit:
message: "Routed to senior legal: {count} flags"
severity: high{
"mode": "routing",
"completed_skill": {
"id": "compliance_check",
"outputs": {
"flagged_items_count": 8,
"frameworks_validated": [
"FCRA", "ADA", "CA_POST"
],
"overall_pass": true
}
},
"available_paths": ["investigator_review"],
"blocked_paths": [
"report_finalize",
"manager_review"
],
"context_patterns": [{
"pattern": "legal_rejection_rate_8_flags",
"value": 0.22,
"sample_size": 247
}]
}{
"selected_paths": ["investigator_review"],
"reasoning": "Compliance passed with 8
flags. Forwarding to investigator with
emphasis on sections 3, 4, 7.",
"routing_type": "topology",
"confidence": 0.91
}Layer 02
The Shepherd
Most transitions are explicit. A human clicks 'Approve,' or a skill completes with a single valid next step. The workflow engine handles these directly. No AI, no latency, no cost. The Shepherd is only invoked at ambiguous junctions where intelligence adds value.
Routing
Triggered when: Multiple valid next steps, no explicit signal
“Where should this go next?”
Error Recovery
Triggered when: A step fails, retries exhausted
“This failed. How do we recover?”
Completeness
Triggered when: Partial results. Some items succeeded, others didn’t
“Is this enough to proceed?”
The Shepherd cannot:
Invent new steps. Bypass the declared process topology. Override policies. Skip required human reviews. It can only work within the structure it’s given.
Layer 03
The Context Graph
The Context Graph aggregates patterns across every process run to build a structured understanding of how your operation actually works. Not how it was designed to work, but how it performs in practice.
A user tells you directly. ‘This section is wrong.’ Rare, but unambiguous.
A user doesn’t tell you, but the system sees what they did. Edits, approvals, rejections, time spent.
Intelligence analyzes observed actions and asks why. ‘Section 4 rewrites correlate with scanned PDF sources.’
{
"process": "background-investigation-report",
"runs_analyzed": 247,
"skill_patterns": {
"document_parsing": {
"overall_reliability": 0.82,
"by_source_type": {
"digital_pdf": { "reliability": 0.95, "edit_rate": 0.08 },
"scanned_pdf": { "reliability": 0.61, "edit_rate": 0.43 },
"docx": { "reliability": 0.97, "edit_rate": 0.05 }
},
"trend": {
"direction": "improving",
"last_30_days": 0.84,
"previous_30_days": 0.79
}
}
},
"routing_patterns": {
"legal_rejection_by_flag_count": {
"0-5_flags": { "rejection_rate": 0.06 },
"6-10_flags": { "rejection_rate": 0.22 },
"11+_flags": { "rejection_rate": 0.51 }
}
}
}From Observation to Policy
The system observes
Across 247 completed runs, legal reviewers reject 51% of reports that arrive with 11 or more compliance flags.
The Shepherd adapts
The Shepherd begins soft-routing high-flag reports to senior legal reviewers. Not because anyone told it to, but because the context graph shows the rejection pattern.
The Context Graph flags a candidate policy
The inference engine identifies this as a stable, high-confidence pattern and generates a policy recommendation.
A human promotes it
A process owner reviews the recommendation, confirms it matches operational intent, and promotes it to a hard policy. The runtime enforces it deterministically from that point forward.
Governance
Every layer has a boundary. Every boundary is enforced.
The platform operates on a simple principle: AI reasons within rules it cannot change. Humans set the rules. The system enforces them.
Hard Guarantees
Deterministic. Cannot be overridden at runtime.
Process Topology
Valid steps and paths through the process. Structural. Cannot be modified at runtime.
Policies
Hard business rules. Floors, ceilings, ordering constraints. Fire unconditionally. No AI can override.
Configuration Profiles
Behavior variation within the same process structure. Resolved before execution begins.
Bounded Intelligence
AI reasons within hard constraints. Transparent and auditable.
Explicit Routing
Human or system provides a clear next step. Fast path. No AI involved.
Shepherd Routing
No explicit direction, multiple valid paths. AI chooses within topology and policy constraints.
Shepherd Recovery
A step fails, retries exhausted. AI can adjust, skip with flags, or escalate. Cannot override policies.
Continuous Learning
Observes and evolves. Never acts without human approval.
Run Traces
Full execution history. Every step, decision, edit, and AI contribution. Immutable. The complete record.
Context Graph
Process-level patterns, inferences, and policy candidates. Evolves with every run. Recommendations require human promotion.
The Shepherd cannot invent new steps. It cannot bypass the declared process topology. It cannot override policies. It cannot skip required human reviews. It can only work within the structure it's given. Adjusting parameters. Flagging concerns. Or escalating to a person.
Policies
Hard rules
AI Reasoning
Bounded
Learning
Human-approved
Security
Built for regulated industries. Trusted with sensitive data.
End-to-End Encryption
Data encrypted in transit and at rest.
HIPAA Compliant
Built for healthcare and regulated industries.
Role-Based Access
Granular permissions per process, per role.
Complete Audit Trail
Every action, decision, and AI contribution logged.
Tenant Isolation
Each client’s data and processes are fully separated.
SSO Integration
SAML and OIDC. Connect your identity provider.
Data Residency
Controls for where your data is stored and processed.
Want to see the architecture behind your process?
We'll walk through how the platform would model your specific operation. Runtime, policies, intelligence, and all.