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

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

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

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:

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:

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

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:

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.