AASSOCIATIONAL TRACK
Describe and stress-test
Estimate observed relationships across a declared universe of reasonable models, while preserving instability, contradictions and failed diagnostics.
Language ceiling · Association
Open discussions GAEF · METHODOLOGICAL BLUEPRINT · VERSION 1.0
Make the analysis
answerable.
客观不是消除所有判断，而是让判断受到规则、证据和审计的约束。
The GHEAS Auditable Evidence Framework governs the complete path from research question to conclusion. It is currently a pre-validation specification: a serious proposal to be implemented and tested, not a claim of proven superiority.
01 — WHY IT EXISTS
A correct calculation can still support the wrong conclusion.
Research credibility depends on measurement, design, model choice, falsification and interpretation together. GAEF treats these as one connected system.
BIDENTIFICATION TRACK
Earn causal interpretation
Require a defensible assignment mechanism, first stage, comparison group, timing checks, negative controls and interference assessment.
Language ceiling · Causal only after gates pass02 — NINE CONNECTED COMPONENTS
No single model carries the whole truth.
Each component constrains a different source of analytical discretion. The audit trail connects them so that later decisions cannot silently rewrite earlier ones.
Claim contract
Lock the population, exposure, outcome, comparison, time and estimand before results can influence them.
Provenance & measurement
Trace every source, transformation, unit, uncertainty and known measurement limitation.
Design map
Separate confounders, mediators, descendants and descriptive variables before modelling.
Two-track inference
Track A examines association; Track B requires a defensible identification strategy for causal language.
Analysis universe
Run the reasonable alternatives that a fair analyst could have selected—not only the preferred model.
Failure gates
Use placebos, negative controls, diagnostics and contradiction checks that can stop a claim.
Bias & uncertainty
Quantify sampling error, measurement error and plausible unmeasured bias where possible.
Claim licensing
Translate the evidence into a calibrated E0–E4 conclusion rather than a binary significant/not-significant label.
Executable audit trail
Preserve code, versions, decisions and bounded AI assistance so another analyst can reconstruct the path.
03 — VALIDATION PROGRAMME
The framework must survive attempts to break it.
Validation starts with software and known-truth simulations, then moves through adversarial conditions, public benchmarks and real-world transfer.
V1Software correctness
Unit, integration and reproducibility tests
PLANNEDV2Known truth
Simulations with controlled data-generating processes
PLANNEDV3Adversarial pressure
Missingness, measurement error, confounding and researcher degrees of freedom
PLANNEDV4Public benchmarks
Comparison against transparent, independently understood datasets
PLANNEDV5PM2.5 case
Reanalyse the preserved country-panel project under the full framework
PLANNEDV6Cross-domain transfer
Test whether the framework travels beyond air pollution and health
PLANNED04 — WORKING PAPER & OUTPUTSRead the specification.
Read the specification.
Inspect the earlier case.
The bilingual paper defines the framework. The PM2.5 dual-track files remain available as the historical implementation that inspired—but does not validate—the new method.
English working paper · PDFComplete GAEF v1 methodological blueprintOpen ↓中文工作论文 · PDFGAEF v1 方法学蓝图中文版Open ↓Dual-track case report · PDFThe earlier PM2.5 Track A / Track B implementationOpen ↓Registered case protocol · JSONQuestion, variable roles, estimands and decision rulesOpen ↓Case reconciliation · JSONMachine-readable evidence boundary and next-data requirementOpen ↓