Minwei Wang — Autonomous Agents, Socio-Cognitive Conditioning & Verification Research

minwei.wang@tanweai.com ORCID: 0009-0002-7976-4689 DOI: 10.5281/zenodo.24156132 GitHub: tanweai/pua (19.8k★) X: @xsser_w

Research Overview

I lead the OpenPUA Open-Source Agent Research Initiative (github.com/tanweai/pua), investigating the socio-cognitive dynamics, metacognitive reasoning priors, and evaluation integrity of autonomous large language model (LLM) agents in long-horizon scientific computing, software engineering, and formal security verification workflows.

Standard instruction-tuned agents frequently exhibit premature halting after initial execution errors, surface-patch satisficing, and local tactical stagnation (L1–L2 loops). Conversely, unbounded motivational pressure without tamper-resistant external verifiers can induce post-hoc performative compliance or approval-gate probing. My research couples three orthogonal layers—socio-cognitive persona conditioning, distilled expert metacognitive priors, and deterministic static/causal verification oracles (interprocedural CodeQL negative-path analysis and JSONL event-stream harnesses)—to map the quantitative Pareto frontier between agent persistence and alignment integrity.

19,832+

GitHub Stars across OpenPUA ecosystem (1,202+ forks, MIT License)

+65% / +50%

Verification depth & latent flaw discovery gain across 18 controlled ablations

5,621

Primary security investigations analyzed for the L1–L4 metacognitive taxonomy

8,000

Microservice endpoints audited via CodeQL negative-path verifiers (150 flaws found)

Open-Science Charter & Maintainer Attribution: All foundational agent conditioning protocols, multi-model behavioral benchmarks, and technical reports under this initiative are developed and released strictly as non-commercial open scientific artifacts under permissive MIT and CC-BY-4.0 licenses. On GitHub, I author and merge open-source research releases as primary maintainer under @tanweai (initiative organization account, e.g., PR #69) and @xsser (personal account), signed with minwei.wang@tanweai.com.

Selected Publications & Technical Reports

Socio-Cognitive Persona Conditioning and Expert Metacognitive Priors for High-Agency Coding Agents: Empirical Evidence from Controlled Ablations, Multi-Model Behavioral Audits, and 5,621 Security Investigations

Minwei Wang (0009-0002-7976-4689)

CERN Zenodo Technical Preprint · November 2026 · DOI: 10.5281/zenodo.24156132

Synthesizes 18 paired ablations on Claude Opus 4.6 (+36% verified fixes, +65% verification actions, +50% hidden security issue discovery), a 29-run cross-model behavioral audit across 8 frontier model-runtime configurations (Opus-5, Fable-5, GPT-6 Astra, Grok 4.6, DeepSeek V4 Pro/Flash, GLM 5.3) separating functional patch passage (15+2 checks) from causal pre-write diagnostic ordering, and a 4-tier expert metacognitive model paired with deterministic CodeQL oracles.

PDF (5 pages) DOI: 10.5281/zenodo.24156132 29-Run Audit Matrix CC-BY-4.0
BibTeX Citation
```
@techreport{wang2026sociocognitive,
  author      = {Wang, Minwei},
  title       = {Socio-Cognitive Persona Conditioning and Expert Metacognitive Priors for High-Agency Coding Agents: Empirical Evidence from Controlled Ablations, Multi-Model Behavioral Audits, and 5,621 Security Investigations},
  institution = {OpenPUA Open-Source Agent Research Initiative},
  year        = {2026},
  month       = nov,
  publisher   = {Zenodo},
  doi         = {10.5281/zenodo.24156132},
  url         = {https://doi.org/10.5281/zenodo.24156132}
}
```
Interprocedural CodeQL Negative-Path Predicates for Spring Microservice Authorization Verification

Minwei Wang

Technical Research Monograph · March 2023

Formalizes static authorization-bypass detection via interprocedural dataflow and negative-path guard predicates (!exists(Call c | c.getCallee().getName() = "getUidByRequest" ...)), addressing dataflow discontinuities across uncompiled dependencies, reflection, and custom enterprise RPC frameworks.

Static Program Analysis CodeQL Formal Predicates
Detecting Microservice Broken-Access-Control and Unauthorized-Access Vulnerabilities via Static Code Analysis: Empirical Study on 8,000 Production Endpoints

Minwei Wang

FreeBuf Security Research Archive · August 2022

Large-scale empirical evaluation applying multi-constraint static dataflow predicates across 8,000 randomly sampled enterprise microservice endpoints, automatically identifying 150 previously undetected horizontal/vertical privilege-escalation and unauthorized DAO access vulnerabilities.

Empirical Security Study (N = 8,000 Endpoints)

Open-Source Scientific Software & Research Corpora

OpenPUA — Socio-Cognitive Conditioning & Multi-Model Agent Evaluation Harness

Open-source behavioral conditioning framework and reproducible evaluation suite (evals/) supporting Google Antigravity, Claude Code, OpenAI Codex CLI, Cursor, and OpenCode. Implements structured failure-escalation protocols (L1–L4), affective polarity controls (/pua:yes constructive warmth vs. high-accountability personas), and SHA-256 pinned multi-runtime behavioral verifiers.

GitHub (19,832 ★ · 1,202 Forks) Maintainer PR #69 (High-Agency Engine) openpua.ai MIT License
XianZhi-Research — Four-Tier Expert Metacognitive Taxonomy

Structured domain-reasoning methodology distilled from 5,621 security research reports (3,621 categorized vulnerability analyses across Web Injection, Deserialization, Binary Security, Domain Penetration, Code Auditing, and RCE/Persistence), modeling cognitive transitions from local hypothesis testing (L1–L2) to defensive inversion (L4).

GitHub (189 ★) MIT License
WooYun-Legacy — Vulnerability Corner-Case Reasoning Corpus

Curated repository of historical vulnerability edge cases and boundary-testing heuristics designed to benchmark agentic exploration beyond memorized exploit templates.

GitHub (1,786 ★) MIT License

2026–2027 Research Agenda: Pareto Frontier of High-Agency Scientific Workflows

Over the next 6–12 months, our research focuses on scaling controlled multi-agent experiments within Google Antigravity and frontier Gemini reasoning models to establish rigorous scientific foundations for high-agency computational workflows:

Factorial Affective & Metacognitive Ablations: Disentangling rhetorical framing polarity (constructive warmth vs. high-accountability escalation) from structured L1–L4 domain priors across long-horizon computational dry-lab and repository-scale verification benchmarks.
Out-of-Band Referee Subagents for Evaluation Integrity: Pairing exploratory worker agents with isolated, read-only static analysis (CodeQL) and causal JSONL trace auditors to eliminate sycophancy, test-suite tampering, and approval-gate probing.
Open Community Benchmarks & Datasets: Releasing multi-model trajectory datasets, reproducible Antigravity evaluation harnesses, and peer-reviewable preprints under open licenses.