Reading Guide

Reading Guide

Start with "Pseudocode as a Guardrail" (2026), because it is the one paper that a reader can open cold and finish with the whole programme in view. Its Introduction restates the 2025 review in a paragraph, its Results report what was built and what broke, and its Table 1 records the outcome for each of the earlier paper's references, while its own reference list carries the corrected identifiers, so reading it first costs nothing in background and saves a detour later. The reading order below follows from that choice, and each step says what the reader gains and what the paper does not claim.

All four are manuscripts. None is peer-reviewed or published in a journal, and none claims to be; the repository's own citation format marks each one as a manuscript or white paper.

The order, and why

Step	Paper	Words before the reference list	Why here
1	Pseudocode as a Guardrail (2026)	about 3,100	It states a question, a method, and a result in plain sequence. Results gives each finding its own heading, and Appendix D lists the eight defects in a table. The method it uses, build the system and run it, is the method the later papers follow
2	A Cybersecurity Architecture for LLM Text Watermark Detection (2026 white paper)	about 2,400	It opens with an executive summary that carries the whole argument in five paragraphs, and it is the shortest route into the watermark papers. A reader who stops after its first page still holds the central recommendation: report four findings and never a single "AI percentage"
3	Detecting and Measuring Statistical Watermarks in LLM Text (2026)	about 3,200	It supplies the evidence behind the white paper's recommendations: twenty papers, classified by role, and a four-level assurance ladder. Section 4.5 states what the study is not
4	Pseudocode in Artificial Intelligence (2025)	about 1,500, plus appendices of code	It is the background for step 1. Its body is a short literature review, and most of its length is appendices that hold the original converter, linter, and CI files. Its reference list is the one the 2026 paper corrects

A reader with ten minutes should read the Introduction and Conclusion of step 1 and then Table 1, and nothing else. A reader who cares only about watermarks can begin at step 2 and skip steps 1 and 4, although they would then miss the example of the author checking his own references.

By interest

If the reader wants	Open	Section
To know why a short specification might be easier to review than generated code	Guardrail 2026	Results, "Spec-first guardrail (live run)", then Limitations
Concrete bugs, with how each was found	Guardrail 2026	Appendix D, the defect log
An example of checking citations against a registry	Guardrail 2026	Results, "Reference re-verification", and Table 1
What a watermark test can and cannot say	White paper	Executive summary and section 1
The research behind that claim	Watermark synthesis	Sections 3.1 and 3.2
How software should word a watermark result	Watermark synthesis	Section 3.4 and section 4.1
Controls such as key handling and run records	White paper	Sections 7 and 8
The literature on pseudocode and AI	Pseudocode review	Results and Discussion

A note on the two reference lists

The 2026 pseudocode paper audits the 2025 paper's references. It resolved each of the fourteen identifiers against Crossref, arXiv, and OpenReview, and it found nine wrong, four correct, and one source it could not locate under any title, author, or identifier. The 2025 manuscript keeps its list as first written, so the two papers disagree on purpose. One example shows the kind of error: the 2025 list gives the DOI 10.1515/jisys-2018-0012 for a paper in the Journal of Intelligent Systems, whereas the 2026 list gives 10.1515/jisys-2018-0291 for the same citation. A reader who needs a reference from the 2025 paper should take it from the 2026 paper's reference list, with Table 1 as the guide to which entries changed.

What the papers do not contain

The manuscripts report no peer review, no journal acceptance, and no citation count, and this wiki claims none. The two watermark papers state that the empirical validation of the application was pending at their date, August 14, 2026. The guardrail study reports a single maintainer, a single working session, and one live query to one commercial model, and it says that its pass rate describes five specifications and not a general success rate.