PoC Benchmark Lab

Validation

RFI accuracy, latency, throughput and source-coverage questions should be answered with measured PoC evidence, not marketing claims.

Prototype principle

The page intentionally starts with “Not measured”. Run simulation only to demonstrate the workflow; actual provider results must replace these values during PoC.

Capability	Slice	Metric	Dataset / test design	Result	Status
Document classification	Per document type	Accuracy / macro-F1	Labeled NIB/NPWP/Akta/SIUP set	—	Not measured
Field OCR	Per field	Exact match / CER	Indonesian legal-document corpus	—	Not measured
Tamper detection	Attack classes	Precision/Recall, FAR/FRR	Recapture/edit/photo-of-photo test set	—	Not measured
Website LOB	Category taxonomy	Macro-F1	Indonesian merchant websites	—	Not measured
Prohibited product	High-risk categories	Recall at chosen precision	Curated positive/negative sites	—	Not measured
Google sentiment	Indonesian/slang	Macro-F1	Human-labeled reviews	—	Not measured
Review manipulation	Authenticity labels	PR-AUC / precision	Synthetic + verified patterns	—	Not measured
Adverse-media entity	Common names	Precision@threshold	Labeled entity matches	—	Not measured
Adverse issue classification	Fraud/corruption/etc.	Macro-F1	Local/international articles	—	Not measured
Social sentiment	Indonesian	Macro-F1	Business-related social content	—	Not measured
Image safety	Each category	Precision/Recall	Adult/violence/racy/medical/spoof	—	Not measured
End-to-end latency	Full KYB	p50/p95/p99	Concurrency profile	—	Not measured

Performance test

p50/p95/p99 latency per module.
Throughput at 1/10/50/100 concurrent cases.
Batch volume and backpressure.
Degraded connector behavior.

Accuracy test

Per-class and per-field metrics.
False positive/negative analysis.
Indonesian language slices.
Common-name entity-resolution slices.

Operational acceptance

Source freshness and availability.
Explainability completeness.
Deletion/correction propagation.
Audit reconstruction test.