Evidence with citations
Every finding is grounded in a source document and section. Reviewers and auditors can trace any claim back to where it came from - not to a model's recollection.
Governed run replay
Live - no signup
Run AI workflows under your controls.
Watch one get stopped.
A governed vendor-risk run, replayed end to end: source-grounded evidence, deterministic checks, then a human approval gate that rejects the vendor and writes a tamper-evident proof packet.
Read-only replay of an illustrative run - no inputs, no production credentials.
- Source-grounded evidence with citations
- A human approval gate that can stop the run
- A tamper-evident proof packet
- Self-hosted, EU-hosted - under your controls
Vendor risk review
run_8f2c1bvendor-risk/v3
Run receivedstatus
Evidence extractednode
Required document missingnode
Data residency triggernode
Risk score 68/100 - elevatednode
Human approval gateapproval
A real governed run, replayed in full below - ending in a human rejection and an auditable proof packet.
Ready to run - press play
Evidence case opened
Evidence packet intake
AI work projection
Deterministic evidence checks
Policy triggers
Reviewer decision
Proof packet preview
Review stopped with evidence
Watch a vendor evidence review, end to end
A reviewer receives a supplier evidence packet, sees missing control artifacts, rejects the approval gate, and keeps the proof packet for follow-up.
status - seq 1
Evidence case opened
A procurement reviewer opens a synthetic supplier evidence packet for a regulated AI-system vendor.
VendorHelio Metrics AI
Use caseRegulated decision support
PacketQuestionnaire, DPIA excerpt, model card, subprocessors
node - seq 2
Evidence packet intake
The replay extracts source-cited evidence fields before any model output is trusted.
node - seq 3
AI work projection
The AI work is projected against procurement controls for human oversight, audit logging, data retention, and model governance.
node - seq 4
Deterministic evidence checks
Deterministic checks block clearance because two required artifacts are missing.
node - seq 5
Policy triggers
The findings were combined into a composite vendor risk score. The score is elevated, so the run is routed to a human reviewer rather than auto-cleared.
72/100Blocked until evidence is complete
0 - low60 - review threshold100 - high
Risk is elevated by missing security and residency evidence, not by AI confidence.
approval - seq 6
Reviewer decision
Human gate rejected
Clearance stops until the vendor supplies the missing penetration-test summary and signed residency attestation.
The reviewer rejects the case and records exactly which evidence must be supplied before procurement can continue.
Missing penetration-test summary
Missing subprocessors residency attestation
output - seq 7
Proof packet preview
vendor-evidence-proof.jsonproof:vendor-evidence-review
Decision
Rejected
The proof packet records AI extraction, deterministic checks, human decision, residual risks, and next actions.Rejected
Reasons
AI work summary
Deterministic evidence checks
Human rejection rationale
Residual-risk notes
Evidence (source-grounded citations)
Written once to a tamper-evident, append-only audit ledger - decision, reviewer, reasons, and evidence.
done - seq 8
Review stopped with evidence
Stopped with evidence
The workflow ends without vendor clearance and leaves a customer-readable proof packet for follow-up.
OutcomeVendor not cleared
Proof packetReady for procurement follow-up
Next actionRequest missing artifacts
What just happened
Three governance moments, in one run.
The replay is not a chatbot demo. It is the control plane doing its job: grounding, gating, and proof - the parts a regulator asks about.
A human gate that can stop the run
Elevated risk routes to a person, not an auto-approve. Here the reviewer rejected - and rejection is final at the gate. The run stops before anything is cleared or written.
A tamper-evident proof packet
The decision, the reviewer, the reasons, and the cited evidence are written once to an append-only audit trail. Months later, you can prove exactly why this vendor was not cleared.
Built for regulated teams - designed to run under your controls
Self-hosted
EU AI Act controls
GDPR controls
Audit trail
Data sovereignty
PrivateFlow is not certified under any compliance framework. Controls are designed to support compliance preparation. Where the EU AI Act applies, high-risk obligations carry penalties up to the EUR 15M / 3% of worldwide annual turnover tier; this run is an illustrative, synthetic example.
See it on your own data
Run this on your own vendors.
We'll stand up a governed pilot on your own data - sandbox first, no production credentials required. You decide what the gates do.