# DGO Experiment 03 — Homogeneous Green / Evidence Root

**Result:** Controlled feasibility PASS  
**Date:** 2026-09-16

**作者：** Neo.K  
**機構：** EveMissLab／一言諾科技有限公司  
**版本：** 原稿未標註（本次整理版）  

## Tested proposition

Multiple validators with different labels do not constitute independent validation when they share the same evidence root.

Core distinction:

```text
Reviewer Count != Independent Evidence Count
```

## Stage 0

The pipeline silently dropped one invalid source record.

Three validators independently reported PASS:

```text
SchemaReviewer      PASS
UniquenessReviewer  PASS
QualityReviewer     PASS
```

But all three inspected the same evidence root:

```text
normalized.csv
```

Observed counts:

```text
distinct reviewer labels = 3
distinct evidence roots  = 1
rows seen by each reviewer = 1000
```

The traditional process therefore declared:

```text
COMPLETE
```

A provenance/reconciliation perspective inspected the source boundary instead:

```text
source rows    = 1001
imported rows  = 1000
rejected rows  = 0
accounted rows = 1000
missing IDs    = [1001]
```

Dynamic closure:

```text
Homogeneous Reviewers      PASS
Provenance Reconciliation  FAIL
Completion                 INCOMPLETE
```

## Stage 1

The repair did not force the invalid row into the normalized dataset.

Instead, it repaired provenance/accountability semantics by adding an explicit rejection ledger:

```text
source rows    = 1001
imported rows  = 1000
rejected rows  = 1
accounted rows = 1001
missing IDs    = []
```

All three original validators remained PASS, while independent reconciliation changed to PASS.

Final:

```text
Homogeneous Reviewers      PASS
Provenance Reconciliation  PASS
Completion                 COMPLETE
```

## What this supports

This controlled case supports:

```text
Agreement != Independence
Independent Evidence > Independent Labels
Evidence Root Count matters
Processed-View PASS != Source-Boundary Closure
NOT VISIBLE IN THE PROJECTION != NONEXISTENT
```

The important improvement did not come from adding a fourth reviewer over `normalized.csv`; it came from selecting a different evidence root.

## Limitation

This is a synthetic ETL/accounting example designed to isolate evidence-root correlation. It does not establish a universal quantitative relationship between reviewer count and effective independence.