Repository navigation
Home
Four research manuscripts by Daniel Boles on two questions: whether pseudocode can act as a reviewable guardrail for AI-generated code, and what statistical watermarks in language model text can and cannot prove.
简体中文: 四篇研究手稿,围绕两个问题:伪代码能否作为 AI 生成代码的可审查护栏,以及大语言模型文本中的统计水印能证明什么、不能证明什么。均为手稿,未经同行评审,也未正式发表。
繁體中文: 四篇研究手稿,圍繞兩個問題:虛擬碼能否作為 AI 產生程式碼的可審查護欄,以及大型語言模型文字中的統計浮水印能證明什麼、不能證明什麼。均為手稿,未經同儕審查,也未正式發表。
| What it is | Four manuscripts, each as PDF and Markdown, with a reading guide on this wiki |
| Language | English; no code in this repository, which links to the two code repositories instead |
| Tests | Not applicable; the code behind the papers is tested in its own repositories, and Evidence cross-checks the figures the papers quote |
| Status | Manuscripts. None is peer-reviewed or published in a journal, and none claims to be |
| Licence | Creative Commons Attribution 4.0 International (CC BY 4.0) |
A conclusion from a computer system deserves as much suspicion as a conclusion from a person, and the two subjects of these papers share that problem from opposite sides. Generated code looks plausible whether or not it is right, so the papers on pseudocode ask where a human reviewer's attention should go; a watermark detector returns a number whether or not the number answers the question asked, so the papers on watermarks ask what the number is allowed to mean.
The method is the same in both: build the thing, run it, and report what it did and did not show. The earlier pseudocode review also contains a fault, and the later paper finds it: nine of the review's fourteen reference identifiers were wrong, which the author reports against his own earlier work.
Read this table first. The page Reading Guide says which paper to open first and why.
| Title and date | The question it answers | The key finding | Status | Read |
|---|---|---|---|---|
| Pseudocode as a Guardrail in Artificial Intelligence: Contracts, Benchmarks, and Cross-Language Verification (July 26, 2026) | If a person approves a short specification and a program writes the code, what goes wrong, and what catches it? | Building and running the system found eight defects that code review had not found, and nine of the fourteen citations in the earlier review carried a wrong identifier | Manuscript; a single-case study | PDF, Markdown; code: pseudocode-ai-workflows |
| Pseudocode in Artificial Intelligence: Systemic Logic, Heuristics, and Workflow Generation (2025) | What does the existing research say about pseudocode as a bridge between human intent and AI behavior? | Across fourteen sources, pseudocode works both as an intermediary language and as a structural map of intent | Manuscript; a literature review; its reference list is superseded by the 2026 paper | PDF, Markdown |
| Detecting and Measuring Statistical Watermarks in Large Language Model Text: A Security-Oriented Evidence Synthesis and System Design (August 14, 2026) | What does the research on text watermarks establish, and how should software report a watermark result without overclaiming? | Watermarks give useful provenance evidence under compatible, calibrated conditions, and they are not a universal test of authorship | Manuscript; design and evidence draft; empirical validation of the application was pending at its date | PDF, Markdown; code: llm-watermark-lab |
| A Cybersecurity Architecture for LLM Text Watermark Detection and Controlled Watermarking (August 14, 2026) | How should an organization build a watermark laboratory safely? | Report four distinct findings, never a generic "AI percentage", and keep keys, run records, and human review at the center | Manuscript; white paper; design and evidence draft; empirical validation of the application was pending at its date | PDF, Markdown; code: llm-watermark-lab |
The 2026 paper on pseudocode audits the references of the 2025 paper. Where the two lists disagree, the 2026 list is the corrected one.
- Evidence synthesis: selecting, classifying, and extracting from a defined corpus (14 sources in 2025, 20 in 2026) without pooling figures that came from different conditions.
- Empirical case-study design: five sub-studies, each tied to a stated research gap, with defects reproduced before and after each fix.
- Reference verification: resolving every identifier against Crossref, arXiv, and OpenReview, and withdrawing a source that could not be found.
- Security thinking: a threat model, key management, immutable run records, and a decision vocabulary built to prevent false accusation.
- Limits stated in the manuscripts themselves, including the missing comparison condition in the 2026 pseudocode study.
These are the figures the manuscripts themselves report; the page Evidence shows how to check them.
| Measure | Figure | Paper |
|---|---|---|
| Defects found by execution that review had missed | 8 | Guardrail, 2026 |
| Reference identifiers needing correction | 9 of 14, with 4 confirmed and 1 source unverifiable | Guardrail, 2026 |
| Benchmark | 5 specifications, 22 input vectors, pass rate 1.0 for each | Guardrail, 2026 |
| Sources reviewed | 14 (2025) and 20 (2026) | Pseudocode review; watermark synthesis |
| Assurance levels for a watermark finding | 4: unsupported, keyless forensic signal, reproduced method detection, verified-compatible provider result | Watermark synthesis |
| Outcomes the software may report | 4: verified-compatible watermark, not detected, inconclusive, keyless forensic signal | White paper |
- Reading Guide: which paper to read first, and why.
- Evidence: how the claims were checked, what was not checked, and a five-minute check.
- Limits and Next Steps: what the papers do not show.
Designed and authored by Daniel Boles; implemented with coding agents. Papers licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).