Data plane · dataset generation · automated pipeline run

Enterprise dataset and automated AIQOS re-run

A parameterised extract shaped like a real customer database — customers, accounts, devices and transactions joined by key — regenerated from a seed and pushed straight through IntentRouter → RAG → Graph/Memory → PolicyEngine → Q-LOCK → Audit. No real customer data is used.

Data sourceSynthetic sample — activeLive customer warehouse — not connectedSample records only; no real customer information is used. Connecting a warehouse later swaps the reader, not the pipeline.

Run configuration

Change any input and the pipeline re-runs automatically

Transaction rows — 500SeedSource profile

Run id

RUN-retail_bank-8af779f0

Fingerprint

8af779f0

Rows generated

500 txn · 160 cust · 228 acct

Pipeline wall clock

1 ms

High volume, low ticket. Card and transfer traffic from a core banking extract. Stitched from CORE_T24.CUSTOMER · CORE_T24.ACCOUNT · SWITCH.CARD_TXN · KYC_HUB.SCREENING.

Pipeline result

Deny-overrides Q-LOCK verdicts across the whole extract

ALLOW

429

85.8% of flow

REVIEW

12.8% of flow

REJECT_ISOLATE

1.4% of flow

Risk avg / p95

0.213 / 0.330

Queued / blocked value

$31.3k

blocked $5.3k

Intent routing mix

routine_payment339

machine_initiated152

velocity_review9

Policy trigger frequency

kyc_pending38

kyc_expired21

no_trade_documentation6

velocity_breach6

sanctions_hit1

Highest risk records

Click a row to inspect the six-stage trace

Transaction	Customer	Amount	Intent	RAG	Risk	Gate
TXN-RE-100369	CUST-4105	$716	routine_payment	0.679	0.446	REVIEW
TXN-RE-100007	CUST-4150	$160	routine_payment	0.716	0.438	REVIEW
TXN-RE-100169	CUST-4142	$597	routine_payment	0.634	0.438	REVIEW
TXN-RE-100195	CUST-4026	$204	machine_initiated	0.659	0.435	REVIEW
TXN-RE-100382	CUST-4105	$529	routine_payment	0.701	0.413	REVIEW
TXN-RE-100135	CUST-4059	$1.2k	routine_payment	0.666	0.402	REJECT_ISOLATE
TXN-RE-100326	CUST-4088	$688	routine_payment	0.618	0.402	REVIEW
TXN-RE-100074	CUST-4036	$279	routine_payment	0.651	0.393	REVIEW
TXN-RE-100269	CUST-4130	$267	routine_payment	0.677	0.381	REVIEW
TXN-RE-100486	CUST-4052	$274	machine_initiated	0.673	0.374	REVIEW
TXN-RE-100134	CUST-4052	$812	routine_payment	0.688	0.363	REVIEW
TXN-RE-100038	CUST-4098	$533	routine_payment	0.699	0.357	ALLOW

Stage trace — TXN-RE-100000

IntentRouter → RAG → Graph/Memory → PolicyEngine → Q-LOCK → Audit

1. IntentRouterroutine_payment
online channel · domestic · routed to specialist agent
2. RAGrelevance 0.708
hybrid 0.6 dense + 0.4 keyword over policy corpus for ZA
3. Graph/Memorydegree 7
sme · energy · account ACC-70041 · device trust 0.81
4. PolicyEnginekyc_pending
deterministic risk 0.293 evaluated deny-overrides
5. Q-LOCKREVIEW
soft gate
6. Auditappended
TXN-RE-100000 · 2026-08-01T00:00:23.956Z · immutable JSONL record

Automating the run

Same generator and pipeline, headless

# one-off run — writes CSVs plus run_report.json
python -m fde_toolkit dataset --rows 500 --seed fde-enterprise-2026 --profile retail_bank --out ./out

# nightly automation (cron, 02:00)
0 2 * * * cd /opt/fde && python -m fde_toolkit dataset \
    --rows 500 --seed "$(date +%F)" --profile retail_bank --out /var/lib/fde/runs/$(date +%F)

The seed fully determines the extract, so a run id plus fingerprint reproduces every row byte-for-byte — which is what makes the audit trail defensible. Swap the generator for a real warehouse reader and the pipeline stages stay unchanged.