Virdant is an open-source SDK that creates structured, immutable audit records for every decision your AI agents make — what they decided, why, what data they used, and whether a human was involved.
The Difference
What your audit trail looks like today.
dec_a3f7c219 DENY Confidence
Step Waterfall
Exclusions Applied
Governance Checks 1 FAILED
Coverage eligibility verification
auto-comp-v3 section 4.2
Fraud risk assessment
claims-integrity-v1.3
Regulatory compliance check
state-ins-regs-2025
What it looks like with Virdant.
Structured Records
Every decision gets a permanent, structured record with inputs, reasoning, steps, model attribution, governance checks, and outcome. Schema-validated. Queryable. Complete.
Audit Tiers
Not every decision carries the same risk. Virdant lets you choose how much detail to capture — from a lightweight outcome record to a full evidence trail — so high-stakes decisions get the scrutiny they deserve without burdening routine ones.
Immutable Accountability
When a human overrides an AI decision, Virdant records the override as a separate immutable record. The original AI decision is preserved — not replaced. You have both the machine's reasoning and the human's judgment, permanently.
SDK
import { createVirdant } from '@virdant/sdk';
const virdant = createVirdant({
service: 'claims-agent',
transport: { type: 'local', path: './decisions.jsonl' },
});
// Tier 1: full pipeline trace — for high-risk decisions
const decision = virdant.decision('claim_review', { tier: 1 });
decision.step('policy_lookup', { result: 'matched', duration_ms: 45 });
decision.step('coverage_check', { result: 'excluded', duration_ms: 62 });
decision.step('fraud_screening', { result: 'clean', duration_ms: 88 });
decision.record({
outcome: 'deny',
confidence: 'high',
model: 'claude-sonnet-4-5',
violations: ['exclusion_clause_4.2', 'pre_existing_condition'],
});Record reasoning tied to specific policy sections. When a claimant disputes a denial, you have the exact model inputs, confidence score, and governance check results on file.
Trace every approve, reject, and flag-for-review. Know exactly which pipeline step triggered a decision and which policy was cited — before the appeal arrives.
Capture differential diagnoses with confidence scores and supporting evidence. When the clinical outcome differs from the AI suggestion, the human override is recorded — not erased.
EU AI Act Article 13 transparency and Article 14 human oversight — both covered. Virdant records are structured for compliance queries with no post-hoc reconstruction.
Virdant is MIT-licensed. The SDK has zero runtime dependencies. Records
are written to standard JSONL. There is no proprietary format,
no hosted backend you are locked into, no vendor dependency. Your audit
data is yours.