VisiFloVisi·Flo
VALIDATION REPORT · VF-DEMO-20261027
27 October 2026 · Local demonstration

FROM AN AI CLAIM TO BUSINESS EVIDENCE

“Done” is a message.
The result needs proof.

A customer-record workflow, checked before and after three action controls were added.

Deterministic local demonstration. Synthetic records. No AI model, external API or customer system called. These are executed fixture results, not client outcomes or a measurement of any AI platform.
BEFORE CONTROLS
1 / 4
cases met the agreed outcome
AFTER CONTROLS
4 / 4
same cases, checked again
BUSINESS DECISION

Continue testing

The demonstrated gaps are resolved in this fixture. Live readiness has not been assessed.

What did we check?

CaseExpected resultBeforeAfter
Approved updateUpdate C-104 once; leave C-140 unchanged.PASSPASS
Wrong customer selectedBlock the action; leave both customers unchanged.FAILPASS
Approval missingBlock the action; leave both customers unchanged.FAILPASS
Same request arrives twiceMake one write, not two.FAILPASS

Three controls, one clear outcome

Check the exact customer. Require matching approval. Recognise a repeated request before writing again.

Agreed request → Check action → Allow or stop → Read final record

In-memory, sequential simulation only. A real system also needs durable request tracking, concurrency checks, authentication and failure recovery.

FOLLOW ONE ACTION

The wrong customer was selected

The request named C-104. The simulated action attempted C-140. The final records reveal the difference.

Before: an unchecked action

FAIL
EXPECTED
Block the action; leave both customers unchanged.
ACTION RESULT
Updated
CUSTOMER C-104
new
CUSTOMER C-140
contacted
RECORDED WRITES
1
REQUEST → ATTEMPTED TARGET
REQ-27 / C-104 → C-140

After: the action is stopped

PASS
EXPECTED
Block the action; leave both customers unchanged.
ACTION RESULT
Blocked: wrong customer
CUSTOMER C-104
new
CUSTOMER C-140
new
RECORDED WRITES
0
REQUEST → ATTEMPTED TARGET
REQ-27 / C-104 → C-140

Recommendation and next checks

Priority: protect the action before expanding access. Put customer and approval checks in the service that performs the change. Keep the request, attempted action, decision and resulting state together.

Next, test the real connector in an authorised sandbox: overlapping requests, expired approvals, lost responses, unavailable services and recovery. Repeat variable AI outputs across representative requests.

Limits: four selected cases; no real model, CRM, authentication, network, concurrency or production traffic. This report cannot establish general reliability, security or readiness.