ArcFides: The always on gatekeeper that ensures you can trust your AI

6 Ways ArcFides Works

Embedded continuous oversight, not just quarterly audit findings
Real-time drift and bias detection, not post-deployment discoveries
Hallucination detection and identification
Injection prompt testing to ensure filter integrity
Objective retraining validation before deployment, not after incidents
Measurable performance metrics, not compliance documentation

ArcFides works with proven tools, monitors continuously, validates in real-time, and provides independent assessment of model behtavior as it evolves. This approach reduces assessment overhead significantly while vastly improving oversight frequency and quality.

Unlike traditional audits that assess systems at fixed intervals, ArcFides transforms periodic compliance reviews into continuous authorization. Federal cybersecurity is evolving from point-in-time assessments to ongoing governance frameworks that integrate zero trust and real-time monitoring with authorization controls—ArcFides applies this same evolution to AI systems.

High Level Architecture – Major Components

Tool agnostic – adaptable to existing systems and toolsets

Technical Architecture – Components & Tool Options

Representative federal-grade tools per layer · Agency-deployed within existing FedRAMP-Moderate boundary

In-Boundary telemetry (raw)

AI System / Model

Anthropic Claude

Anthropic Claude
Azure OpenAI Gov
AWS Bedrock / Vertex
Mistral / Llama (Gov)
Ollama (air-gap)

Validated / hashed metrics

ArcFides Control Plane · FedRAMP-Moderate-Compatible
Embedded, Federated, or Hybrid deployment

Telemetry Collector

Infra observability

Prometheus · CloudWatch
Datadog

LLM-native observability

Langfuse · Arize Phoenix
LangSmith · OpenLLMetry

Evaluation Engine

Data-quality validation

Great Expectations
MLflow Evaluate

LLM behavioral eval

DeepEval · Ragas
TruLens · Promptfoo
NeMo Guardrails

Rules & Alerting

Microsoft Sentinel (SIEM)
Splunk ES / SOAR
ServiceNow FedRAMP
PagerDuty Gov · Opsgenie
Slack / Teams Gov hooks

Evidence Vault (WORM)

AWS S3 GovCloud (Object Lock)
Azure Blob Gov · Conf. Ledger
GCP Storage Gov · MinIO
Immudb (crypto ledger)
HashiCorp Vault (keys)

Assurance Score Engine · Compute · Orchestration · Output APIs

Compute

ArcFides Score Svc
Python
MLflow Tracking

Orchestration

GitHub Actions
GitLab GovCloud
Azure DevOps Gov
Jenkins

Output APIs

Power BI Dataset
Splunk HEC
JSON Score API

External output — no PII

Output & Distribution

Visualization & Output

BI dashboards

Power BI Gov · Tableau Gov
Grafana · Kibana

GRC / IRM

ServiceNow GRC
RSA Archer

SIEM consumption

Splunk Dashboard Studio
Microsoft Sentinel

Distribution

SharePoint Gov
Assurance Packets (PDF)
Agency RM System

End-to-End Data Flow

Eight stages · Six EAF record types · Continuous re-baselining loop

01 AI System generates inference telemetry	02 Telemetry Collector captures traces + infra metrics	03 Evaluation Engine validates behavior in real time	04 Evidence Vault WORM-sealed record of every check
05 Rules & Alerts flags drift, hallucination, injection	06 Assurance Score composite trust score · T0–T3 tier	07 Visualization executive + ops dashboards	08 Assurance Packet signed quarterly deliverable to oversight

rebaseline trigger · drift = governance event, not retrain

EAF Record Types

The data contract carried across every stage. Evidence Vault stores all six.

Metric Telemetry + Validation

Incident Validation on Breach

Assurance Score Score Engine

Change Governed Artifact Promote

Risk Item POA&M

Attestation Quarterly Packet

USE CASE: Civilian Federal (AWS GovCloud)

Bedrock + Claude · Object-Lock + Vault evidence · Power BI Gov + Splunk Cloud Gov dual-output

Civilian Enclave · AWS GovCloud · FedRAMP-Moderate · GitHub Actions Federal Orchestration

In-Boundary telemetry (raw): AI SYSTEM / MODEL

Bedrock + Claude

AWS Bedrock (GovCloud) hosting Anthropic Claude 4.x. Emits inference logs, latency, token, and tool-use telemetry.

Validated / hashed metrics TELEMETRY COLLECTOR

Langfuse + Prometheus

Langfuse captures LLM behavioral traces; Prometheus scrapes infra metrics on a 60-second interval.

Validated / hashed metrics EVALUATION ENGINE

DeepEval + Great Expectations

DeepEval runs behavioral suites (hallucination, refusal, injection). GE validates metric data quality on snapshots.

Validated / hashed metrics RULES & ALERTING

PagerDuty + Splunk SOAR

PagerDuty Gov routes S1–S3 incidents to on-call. Splunk SOAR runs automated containment playbooks on validated incidents.

EVIDENCE VAULT (WORM)

S3 GovCloud + Vault

Object Lock enforces write-once. HashiCorp Vault signs each dataset hash. 3-year retention. All six EAF record types stored here.

ASSURANCE SCORE ENGINE

GitHub Actions + Python

GitHub Actions Federal pipeline reads S3 records, computes T0–T3 composite score, pushes to both Power BI and Splunk Cloud Gov.

External output — no PII VISUALIZATION

Power BI + Splunk Cloud

Dual output: Power BI Gov for executives and program PMs; Splunk Cloud Gov for SOC analysts. Signed packets distributed via SharePoint Gov.

What Makes ArcFides Different

Three thesis distinctions that separate ArcFides from generic ML observability

Early-Detection
Posture

Continuous assurance, not gatekeeping

ArcFides judges whether an authorized system continues to behave within its envelope—not whether it should exist. Authorization is a one-time decision; assurance is the ongoing answer to "is it still safe?"

Behavioral Drift > Population Drift

Watch what the system does, not just what it sees

Most observability tools track input distributions. ArcFides tracks behavior—hallucination, refusal, injection susceptibility, calibration—because that's what breaks Federal trust.

Rebaseline = Re-Authorization

Drift triggers governance, not a retrain

When behavioral drift exceeds thresholds, ArcFides doesn't fire a retrain job. It fires a governance event with a complete audit trail—re-authorization is the workflow, not a model update.

West 4th designs the customized framework, builds the implementation, and runs ongoing assurance.

Why a Framework, Not Just Tools?

Federal agencies already own observability tools. Tools produce data. Frameworks produce assurance

What are the alternatives?	What none of them provide — ArcFides
Langfuse · Phoenix · LangSmith · Datadog capture LLM traces. Engineering observability.	01 Policy-mapped tier badges A drift z-score becomes a T0–T3 action only through ArcFides tier map. Tools produce numbers; the framework produces decisions aligned to OMB / NIST / FedRAMP.
Evidently · Ragas · DeepEval · TruLens score drift and behavior. Data-science evaluators.	02 WORM evidence vault Cryptographically hashed, tamper-evident, audit-grade records. Tools store data in their own DB. The framework produces evidence an IG can pull two years from now.
Agent Artemis · Kov.AI · Vibrantix vendor "AI governance" products. Proprietary schema, proprietary scoring, not framework-defined.	03 The Assurance Packet Quarterly PDF + bundle: Exec Summary, 4-Color rating, Metrics snapshot, Incident Log, POA&M, Change Summary, Attestation. Tools produce dashboards. The framework produces a deliverable to oversight.
In-house build possible, but every agency reinvents the schema, the tier map, the evidence chain.	04 Six-record evidence schema Assurance Score, Metric, Incident, Change, Risk Item, Attestation Artifact. A canonical framework that makes evidence cross-comparable across systems and agencies.
	05 Independence by architecture ArcFides reads from production, never writes back. Structural separation that makes assurance independent. Tools are wired into the request path or called from app code; the framework produces audit independence.

ArcFides – Engineering Summary

Continuous Assurance · Cryptographic Traceability · FedRAMP-Ready

01 Five-Layer Control Plane Telemetry → Validation → Evidence Vault → Assurance Score Engine → Visualization	02 Event-Driven CI / CD GitLab / GitHub Actions / Azure DevOps Gov—triggers on every ingest, refresh, and score recalculation	03 Cryptographic Integrity Every output dataset hash—BI dataset, SIEM index, GRC export—stored in the Evidence Vault for end-to-end lineage; six EAF record types signed at write
04 Five Operational Dashboards Executive Command Center plus four specialized ops dashboards · renders to the agency’s chosen surface (Power BI, Tableau, Splunk, Grafana, Kibana)	05 Three Deployment Modes Embedded, Federated, or Hybrid—configurable per FedRAMP tier and agency boundary	06 Rebaseline as Governance Event Drift fires re-authorization workflow, not a retrain—ATO-aligned, audit-trail-complete
Comprehensive Framework Alignment OMB M-25-21 / M-25-22 · lineage from OMB M-24-10 / M-24-18 · NIST AI RMF 1.0 · FedRAMP Rev 5 · ISO 42001 · DoD RAI

ArcFides in Practice

A reference implementation of the ArcFides Control Plane is available for review. Streamlit-Based Assurance Console · Ollama / Mistral Inference · Tracing / LLM-as-Judge · WORM Evidence on Disk

Request a Demo

The always on gatekeeper that ensures you can trust your AI.

“Is it still safe?” — answered continuously.

Continuous monitoring, continuous assurance.

Open-source tools, government-approved.