giabaohuynhasu/vibe-researching

Vibe Researching: A Reproducibility Protocol, Agent Skill, and Epistemic Audit Engine for Humanities & Social-Science Quantitative Analysis

Bridging the gap between a qualitative humanities question and a fully reproducible, epistemically audited quantitative analysis

Author: Gia Bao Huynh
Independent Researcher, Ho Chi Minh City, Vietnam
Email: huynhbao@asu.edu · ORCID: 0009-0008-2372-5852
Research Stance: Challenging the unchecked power, epistemic asymmetries, and monopolized governance of non-state actors across frontier technologies.
Collaborator: Claude (Anthropic)

Persistent Identifiers & Repositories:

• 🏛️ Zenodo DOI: 10.5281/zenodo.22978433
• 🌐 GitHub: https://github.com/giabaohuynhasu/vibe-researching
• 🤗 Hugging Face Hub: https://huggingface.co/datasets/giabaohuynhasu/vibe-researching

---

📌 What This Is

Vibe Researching is an agent skill, reproducible-research framework, and automated epistemic audit system designed for researchers who work in humanities and social-science fields — philosophy, political theory, comparative jurisprudence, institutional analysis, and area studies — and want to move from a conceptual research question to a real quantitative analysis (fitting a model, testing a hypothesis, analyzing a dataset) without losing ownership of substantive intellectual decisions along the way.

The Problem It Solves

AI coding assistants can generate project scaffolding, data cleaning pipelines, econometric model calls, and diagnostic plots far faster than a non-programmer could build them by hand. However, that mechanical speed introduces an acute epistemic failure mode: the AI system silently resolves the researcher's substantive theoretical choices by default (e.g., implicitly selecting model functional forms, omitting confounders, establishing outlier exclusion cutoffs, or adopting arbitrary significance thresholds) without the researcher ever realizing a consequential decision was made.

The Mechanism

Vibe Researching enforces mandatory epistemic checkpoints that strictly isolate technical scaffolding generation from the researcher's theoretical sovereignty:

1. Mandatory Research Question Gate (Step 1): Collaborative completion of research_question.md before any code is written or data is touched.
2. Strict Raw-to-Processed Separation (Step 2): Immutable data/raw/ storage with reproducible scripted pipelines.
3. Ordered-Script Workflow (Step 3): Sequential execution (01_clean_data.py, 02_fit_model.py, etc.) making analytical provenance transparent.
4. Substantive Decision Cards (Step 4): Explicit documentation of every choice capable of shifting empirical conclusions, detailing alternatives and counterfactuals.
5. Contemporaneous Research Log (Step 5): Real-time session logging recording actual deliberation, not retrospective rationalization.
6. The Anti-Default Hard Rule (Step 6): AI systems are strictly prohibited from silently resolving substantive choices; they must pause and surface 2–3 defensible alternatives with empirical trade-offs.
7. Living Documentation (Step 7): Continuous maintenance of README.md and CITATION.cff.
8. The Third-Order Epistemic Audit (Step 8): Automated, stateless verification across Popperian falsifiability (Order 1), Mertonian external grounding and escape-route closure (Order 2), and Lakatosian research trajectory progressiveness (Order 3).

---

📂 Repository Structure

vibe-researching/
├── SKILL.md                          # The core agent skill definition (8 steps)
├── audit/                            # Third-Order Audit engine (Huynh 2026)
│   ├── sandbox.py                    # Core research sandbox & multi-order audit engine
│   ├── vibe_audit.py                 # Bridge & CLI for auditing Vibe Researching projects
│   ├── SPEC.md                       # Formal audit specification & mapping standards
│   ├── example_corpus.json           # Reference test corpus
│   ├── test_sandbox.py               # Engine unit tests (16 tests)
│   └── test_vibe_audit.py            # Integration tests for project audits
├── vibe-researching-skill/           # Installable skill distribution package
│   ├── SKILL.md                      # Skill instructions (with template references)
│   └── templates/                    # Blank standardized templates
│       ├── research_question.md      # Prerequisite question-framing gate
│       ├── decision_card.md          # Substantive choice record template
│       ├── research_log_entry.md     # Contemporaneous session log template
│       ├── third_order_audit_report.md # Epistemic audit report template
│       ├── README.md                 # Project root template
│       └── CITATION.cff              # Citation metadata template
├── demo-project/                     # Complete end-to-end worked demonstration
│   ├── research_question.md          # Completed research question
│   ├── audit/                        # Automated third-order audit artifacts
│   │   ├── third_order_audit_report.md
│   │   └── audit_object.json
│   ├── analysis/
│   │   ├── scripts/                  # Numbered reproducible pipeline (01–05)
│   │   │   ├── 01_generate_data.py
│   │   │   ├── 02_clean_data.py
│   │   │   ├── 03_fit_model.py
│   │   │   ├── 04_robustness_check.py
│   │   │   └── 05_make_figure.py
│   │   └── models/                   # Formal decision cards
│   │       └── decision_leverage_points.md
│   ├── data/
│   │   ├── raw/                      # Untouched raw observation data
│   │   └── processed/                # Pipeline-generated clean data
│   ├── figures/                      # High-resolution publication plots
│   │   └── income_lifespan_fit.png
│   ├── research_log/                 # Real-time session documentation
│   │   └── log.md
│   ├── README.md
│   └── CITATION.cff
├── CITATION.cff                      # Master citation metadata
├── LICENSE                           # MIT License
└── README.md                         # Project documentation

---

🔬 The 8-Step Protocol

Step	Phase	Function & Enforcement
0	Recognize	Activates when a researcher from a qualitative/humanities discipline initiates a project without an existing codebase.
1	Research Question Gate	Halts technical execution until research_question.md is agreed upon: defining Question, Theoretical Claim, Model Justification, Identifying Assumptions, Falsification Criteria, and Reproducibility Specifications.
2	Scaffold Project	Establishes standard directory layout. Enforces strict write-protection on data/raw/.
3	Numbered Pipeline	Enforces chronological script naming (01_, 02_, 03_), narrating the complete analytical trajectory.
4	Decision Cards	Requires a stand-alone decision record in analysis/models/ for any substantive analytical fork.
5	Research Log	Records real-time decisions, rejected alternatives, and open questions during the working session.
6	The Hard Rule	Mandatory pause on any fork that could alter conclusions. AI must present options with trade-offs; silent picking is forbidden.
7	Living Closeout	Synchronizes status, citations, and outputs into README.md and CITATION.cff at every natural checkpoint.
8	The Third-Order Audit	Programmatic verification: Order 1 (Popperian falsifiability), Order 2 (Mertonian grounding / closed escape routes), and Order 3 (Lakatosian Constraint Ratio across session revisions).

---

🛡️ The Third-Order Epistemic Audit (Huynh 2026)

The Third-Order Audit engine (audit/vibe_audit.py, built upon audit/sandbox.py) provides rigorous, automated verification of a project's epistemic integrity across three philosophical orders:

1. Order 1 (Karl Popper) -- Object-Level Falsifiability

• Principle: A scientific hypothesis must specify empirical observations that would refute it.
• Audit Verification: Inspects research_question.md to verify that an explicit falsification condition ($F_1 \dots F_k$) is formally articulated (rejecting placeholders, tautologies, or unfalsifiable formulations).

2. Order 2 (Robert K. Merton / The Price of a Promise) -- Mechanism-Level Grounding & Closed Escape Routes

• Principle: Falsification criteria must point outward to independent external reality rather than collapsing into circular, model-internal jargon.
• Audit Verification:
  • Confirms external anchors in data/raw/ and authoritative citations in references/.
  • Verifies that analytical escape routes are closed by confirming that decision cards in analysis/models/ document counterfactual options and explain the exact sensitivity of conclusions.
  • Monitors for Third Break boundary risk: flags hypotheses that rely on model-internal queries without empirical validation.

3. Order 3 (Imre Lakatos / The Artifact Problem) -- Programme-Level Trajectory

• Principle: Distinguishes progressive research programmes from degenerating ones across time. Progressive programmes respond to empirical anomalies by narrowing their claims or withdrawing refuted assertions; degenerating programmes preserve initial dogmas through ad-hoc protective belt additions.
• Audit Verification: Evaluates multi-session entries in research_log/log.md, classifying each revision into:
  • NARROWED: Scope or claim strength reduced under empirical tension.
  • WITHDRAWN: Claim explicitly abandoned.
  • REAFFIRMED: Claim retained unchanged despite testing.
  • EXTENDED: New hypothesis or expansion.
• The Lakatosian Constraint Ratio: $$\text{Constraint Ratio} = \frac{\text{Narrowed} + \text{Withdrawn}}{\text{Total Revisions}}$$
  • $\text{Ratio} > 0$: Progressive programme (empirical evidence actively constrains claims).
  • Repeated REAFFIRMED with $\text{Ratio} = 0$: Degenerating Risk (confirmation bias / protective insulation).

The Stateless Auditor Advantage

Human researchers naturally experience escalation of commitment (Staw 1976) and motivated reasoning (Kunda 1990) to defend decisions they previously invested effort into. A fresh, stateless AI instance executing the Third-Order Audit has zero personal attachment or sunk cost in prior conclusions, enabling unbiased structural verification.

Running the Audit

# Audit any Vibe Researching project folder
python audit/vibe_audit.py demo-project

# Run the complete test suite
python -m unittest audit/test_sandbox.py audit/test_vibe_audit.py

Sample audit output:

# Third-Order Epistemic Audit Report: demo-project

## 📊 Audit Scorecard

| Level | Order Name | Status | Evaluated Metric / Criterion |
| :--- | :--- | :---: | :--- |
| **Order 1** | **Object-Level Falsification** | **✅ PASS** | Explicit falsification condition stated in `research_question.md` |
| **Order 2** | **External Grounding & Escape Routes** | **✅ PASS** | Outward linkage to independent data/records; closed analytical escape routes |
| **Order 3** | **Programme-Level Lakatosian Trajectory** | **ℹ️ BASELINE** | Constraint Ratio = `0.000` (0 narrowed/withdrawn of 1 total revisions) — Initial baseline established |

---

🎯 Demo Walkthrough

The repository includes a complete empirical demonstration (demo-project/) analyzing household income vs. life expectancy:

• Empirical Question: Does household income predict longevity, and what is the magnitude of association?
• Model: OLS regression of life expectancy on log-transformed household income.
• Surfaced Decision (Step 6): Influence diagnostics identified 12 high-leverage observations at extreme income brackets.
• Deliberation & Resolution: Rather than arbitrarily dropping points or silently applying robust errors, the researcher chose a tripartite comparison (baseline OLS, HC3 heteroskedasticity-robust SEs, and leverage-excluded OLS).
• Result: Demonstrated empirical stability across all three specifications (slopes: 2.696 vs. 2.696 vs. 2.624; <2.7% divergence), establishing robust conclusions without obscured data deletion.
• Epistemic Audit (Step 8): Automated evaluation in demo-project/audit/ confirms Order 1 pass (explicit slope threshold for falsification), Order 2 pass (anchored to raw observation records and decision card sensitivity tests), and Order 3 baseline tracking.

Executing the Demonstration

cd demo-project
pip install pandas numpy statsmodels matplotlib
python analysis/scripts/01_generate_data.py
python analysis/scripts/02_clean_data.py
python analysis/scripts/03_fit_model.py
python analysis/scripts/04_robustness_check.py
python analysis/scripts/05_make_figure.py
cd ..
python audit/vibe_audit.py demo-project

---

🏛️ Intellectual Context & Authorship

This framework and audit architecture were developed by Gia Bao Huynh (huynhbao@asu.edu, ORCID: 0009-0008-2372-5852) during the empirical investigations of The Floor That Does Not Rise and The Third-Order Audit research programs.

Conducting complex quantitative modelling alongside generative AI assistants revealed that speed of synthesis often masks epistemic capture: technical co-pilots routinely make foundational theoretical compromises under the guise of syntax completion. Vibe Researching codifies a protocol of epistemic sovereignty, ensuring non-programmer scholars retain total ownership, methodological clarity, and transparent accountability over their scientific claims.

---

📜 Citation

@software{huynh2026viberesearching,
  author       = {Huynh, Gia Bao},
  title        = {Vibe Researching: A Reproducibility Protocol, Agent Skill, and Epistemic Audit Engine for Humanities and Social-Science Quantitative Analysis},
  month        = sep,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {2.1.0},
  doi          = {10.5281/zenodo.22978433},
  url          = {https://doi.org/10.5281/zenodo.22978433}
}

---

⚖️ License

Distributed under the MIT License.