Validation
Measure what matters.
Benchmarks, Golden Standard and acceptance before scale.
CHINH PHONG does not treat OCR speed as the final metric. Every project agrees a set of quality and capacity metrics tied to accepted output.
Metrics
What we measure
| Metric | How to read it |
|---|---|
| R = Accepted Output / Day / Production Unit | The number of items or files accepted for handover per day on one production unit. |
| Critical Field Accuracy | Accuracy on the critical fields. |
| Auto-Pass Rate | The share of records the system confirms as passing against locked rules. |
| Review Rate | The share of records that need a human check. |
| Rework Rate | The share of records that must be processed again. |
| Source Issue Rate | The share that cannot be completed because source data is missing or damaged. |
| Traceability Coverage | The share of data that can be traced back to its source when required. |
| Stability | The ability to run continuously across the agreed benchmark window. |
Metric definitions are agreed per project before benchmarking; no public numbers are published without approval.
Golden Standard
The Golden Standard locks
The Golden Standard is the technical contract between AI capability and business requirements.
Representative dataset
The dataset used to define, test and re-test the standard.
Schema
The output structure the data must fit.
Field definitions
What each field means, including type, format and allowed values.
Source priority
Which source wins when documents disagree.
Critical fields
The fields where accuracy is mandatory before acceptance.
Conflict rules
How contradictions between sources are resolved.
Missing-source rules
What happens when a required source is absent.
Pass/fail
The conditions that make a record or batch acceptable.
Sampling method
How records are selected for review and re-sampling.
Acceptance owner
The role that signs off acceptance before scaling.
The rule that governs scale
No Golden Standard, no industrial scale.
Benchmark & acceptance
How acceptance works
A benchmark runs on the representative dataset against the frozen Golden Standard inside an agreed window. Results cover accuracy on critical fields, auto-pass, review, rework and source-issue rates, traceability coverage and stability. Exceptions are reviewed, reworked and re-sampled. Acceptance is signed by the agreed owner before scaling.
How to start
Start with one representative dataset
Choose one representative dataset, lock the Golden Standard and run the benchmark before scaling production.
Related