# Dynamic Global Observation Controlled Benchmark v0.1

**Date:** 2026-09-16  
**Status:** Three controlled feasibility experiments completed

**作者：** Neo.K  
**機構：** EveMissLab／一言諾科技有限公司  
**版本：** v0.1  

## Result matrix

| Experiment | Hidden failure family | Traditional result | Dynamic result before repair | Final |
|---|---|---|---|---|
| 01 | Mobile visual + cross-boundary purchase intent | False COMPLETE | Correctly INCOMPLETE | COMPLETE after two targeted repairs |
| 02 | Deployment receipt vs actual target runtime | False COMPLETE | Correctly INCOMPLETE | COMPLETE after promotion-state repair |
| 03 | Homogeneous reviewers sharing one evidence root | False COMPLETE | Correctly INCOMPLETE | COMPLETE after provenance/rejection-ledger repair |

## Most informative intermediate states

### Experiment 01

```text
Local PASS
Visual PASS
Transaction Boundary PASS
Macro Witness FAIL
```

This showed that locally repaired components can still fail the actual user-intent relation.

### Experiment 02

```text
Local Artifact PASS
Deployment Receipt PASS
Environment Closure FAIL
Macro Witness FAIL
```

This showed that action/provider receipt cannot replace fresh target-world observation.

### Experiment 03

```text
3 reviewer labels
1 evidence root
all reviewer verdicts PASS
source reconciliation FAIL
```

This showed that more reviewer labels do not automatically create independent evidence.

## Cross-experiment pattern

All three traditional false completions shared the same structural mistake:

```text
a valid local/proxy signal
was promoted to a claim whose scope was larger than the evidence
```

The dynamic method prevented promotion by keeping closure dimensions separate until the missing world relation was actually observed.

## What is now supported

The experiments provide controlled feasibility evidence for:

```text
Green != Done
Local Closure != Macro Closure
Receipt != Outcome
Agreement != Independence
Perspective selection can be intent/evidence conditioned
Targeted reopening can repair only affected closure dimensions
Macro Witness is operationally distinct from local tests
```

## What remains unproven

```text
large-repository scalability
router precision/recall on blind tasks
token/wall-clock superiority over strong baselines
learned-router advantage over rules
performance under ambiguous real user requirements
multi-agent correlation with real heterogeneous models
```

## Readiness decision

The theory is now strong enough to justify a small **Perspective Closure Runtime MVP**.

It is still too early to freeze a large workflow Skill.

Recommended sequence:

```text
Controlled Benchmarks v0.1  [DONE]
        ↓
Perspective Closure Runtime MVP
        ↓
Mixed hidden-failure benchmark suite
        ↓
Cost / precision / recall measurements
        ↓
Thin internal Skill
```

The Skill should expose almost none of the internal ceremony. Its job should be to invoke the runtime silently and return finished work plus an evidence-aware completion state.