Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

PriMera Scientific Engineering Volume 9 Issue 5 November 2026 DOI: 10.56831/PSEN-09-304 ISSN: 2834-2550

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application Type: Research Article Received: October 22, 2026 Published: October 30, 2026

Citation: Christian Mancas., et al. “Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application". PriMera Scientific Engineering 9.5 (2026): 30-48. Copyright: © 2026 Christian Mancas., et al. This is an open-access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Christian Mancas1* and Cristina Mancas2 1

Mathematics and Computer Science Department, Ovidius University at Constanta, Romania

2

Gymnasium School No. 86, Bucharest, Romania

*Corresponding Author: Christian Mancas, Ovidius University, Bd. Mamaia 124, Constanta, CT, Romania.

Abstract This research paper proposes a broader definition of vibe coding, extending the original one

with non-developers and other natural languages than English, introduces the notion of vibe debugging, as well as a sound, complete, and optimal algorithmic methodology for testing and

vibe debugging vibe-coded software applications. This methodology is illustrated by applying it to a vibe-coded real-world database application of medium complexity generated with Google AI Studio Pro. We prove that even non-developers can obtain a final bug-free and functional sat-

isfactory application version, even in other languages than English. We also prove that, although Google AI Studio Pro has sophisticated built-in engines for both language analysis, software development, and automated testing, it is still making lot of mistakes, so that testing and vibe debugging of its generated code is a must.

Keywords: vibe coding; vibe debugging; Google AI Studio; relational database design; software application testing and debugging

Abbreviations AI = Artificial Intelligence.

app(s) = software application(s). db(s) = database(s). e.g. = for example.

GUI(s) = graphical user interface(s). i.e. = that is.

LLM(s) = large language model(s). RDM = Relational Data Model.

spec doc = specification documentation.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

31

Introduction We presented in [1] a case study on generating through vibe coding with Google AI Studio Pro (powered by the Gemini 3.8 engine)

[2] a medium-complex useful database (db) software application (app) [3]. We proved in it that this can be done by a non-developer as

well and, moreover, that it can be fully done without using English, probably in almost any other natural language, for sure in Romanian as well.

We thoroughly tested the resulting app [3] and presented in detail both its fine features, its bugs, and our enhancement ideas. As

we wrote in the conclusion and further work section of [1], we then structured our reply, submitted it to Google AI Studio, obtained a

corresponding new app version, tested it as well, and then recursively repeated this process until we were satisfied with the app or/ and sure that we cannot obtain from Google AI Studio everything we would like.

After our experience described in [1], we understand by vibe coding a programming paradigm where both non-developers and de-

velopers develop apps primarily (and even solely) by interacting with code-generating large language models (LLMs), using prompts in plain English or any other natural language accepted by these LLMs.

After our experience described in the present paper, we understand by vibe debugging an app debugging paradigm having same

characteristics as vibe coding.

This article describes in detail these steps, as well as the new features of the final app. In the sequel, we also provide a sound, com-

plete, and optimal pseudocode algorithmic methodology for app testing and vibe debugging.

The whole conversation with Google AI Studio Pro was conducted in Romanian. We present here their translations to English done

by Google Translate, in which we however preserved the Romanian names for the db tables, columns, and app forms, queries, and reports.

Just like for the initial app specification documentation (spec doc), our prompts were mainly written by the second author of this

paper; the first one added only the technical specs from the Relational Data Model (RDM) theory [4-6] and MS Access technology [7]. Related Work

Software testing is internationally standardized [8]. Lot of books and articles were published on software testing and debugging

(e.g., [9-12]). Typically, in the beginnings, these tasks were done by developers. Then, non-developers were mostly employed for man-

ual testing apps, who then passed their findings to developers for debugging. Since a quarter of a century already, automated software

testing tools are available (e.g., [13-15]), easing this task for both non-developers and developers. Their technologies or similar ones

are today embedded in code-generating LLMs, blending them with AI-driven test synthesis, closing feedback loops via static analysis, execution engines, and multi-agent validation (e.g., [16, 17]). Google AI Studio performs highly sophisticated automatic testing of its generated code with its own Code Execution mechanism [18].

Materials and Methods

Figure 1 shows the reply we fetched to Google AI Studio Pro (in the ExcursiiScoala Project [3]). Please note that it is structured into

4 groups: the first one with thanks for the app’s best features, the second one with suggested meta-axioms to be followed for any db app, the third one with signaling app’s bugs, and a final one with desired app enhancements.

Figure 2 shows the corresponding answers provided by Google AI Studio when it finished building the second app version according

to our requests and suggestions from Figure 1, after 10 minutes of work (i.e., 3 minutes more to correct and enhance than to build the first version).

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

32

Figure 1: Our first reply to Google AI Studio.

Figure 1: (Continued).

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

33

Figure 1: (Continued).

Figure 1: (Continued).

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

34

Figure 1: (Continued).

Figure 1: (Continued).

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

35

Figure 1: (Continued). We then republished the new app version [3] (which took some 3 minutes), to enable several users to parallelly test it.

During testing of these app’s updates, we found the issues shown in Figure 3 that we submitted to Google AI Studio; this time, it first

took to the Gemini 3.8 Flash engine one hour and 3 seconds of work, after which it stopped and displayed the message “The request timed out. Please try again”.; fortunately, after hitting the Retry button, it displayed “There was an unexpected error. Finish what you were doing”. and after only other 39 seconds finished its tasks. Its corresponding reply is shown in Figure 4.

Figure 2: Google AI Studio reply to our prompts from Figure 1.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

36

Figure 2: (Continued). We then iterated through this process with republishing the new app and retesting it, beginning with the issues shown in Figure 3

and finishing with the new features added by Google AI Studio (as per Figure 4). The bugs we found in this process were then submit-

ted to Google AI Studio (see Figure 5): after 1h18’ (again, with the process having been canceled for time out after 1h and immediately resumed) it replied as shown in Figure 6, and we could start testing and vibe debugging the app’s 3rd version.

Figure 3: Our second reply to Google AI Studio.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

37

Figure 4: Google AI Studio reply to our prompts from Figure 3.

Figure 4: (Continued). PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

38

Figure 5: Our third reply to Google AI Studio.

Figure 6: Google AI Studio reply to our prompts from Figure 5. PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

39

Figure 6: (Continued). Finally, as all tests passed ok in the third iteration, we retested the whole app. Fortunately, as no regressions or new bugs were found,

we could share the third and final app version to all interested colleagues of the second author of this paper.

The methodology we designed and followed in this iterative process to test and vibe-debug apps is shown in Figure 7. Obviously,

“Google AI Studio” may be replaced with the name of any other similar AI tool, without impacting algorithm ATVBM’s power and usefulness.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

40

Figure 7: The pseudocode algorithm ATVBM. It is easy to formally prove that the algorithm from Figure 7 is both sound, complete, and optimal, as, first of all, it never loops in-

finitely (i.e., at some point, for any app and vibe debugging user, no more iterations are needed, be it because the app is finally fine, or you lowered your expectations -i.e., the remaining bugs are accepted as features-, or you give up completely and forget about that

app or about using that AI tool, at least for the time being, waiting for the next, hopefully better version), then because it accepts any vibe-codded app, outputs only fully validated ones, and it is doing it in the minimum possible number of steps.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

41

Figure 7: (Continued). The first author of this paper tested and vibe-debugged only the Navigation Pane, the underlying db structure, the SQL SELECT

statements underneath queries and reports, as well as the db backup and restore functionalities. He also designed the testing and

vibe-debugging algorithm in Figure 7. The second author tested and vibe-debugged the rest of the functional behavior of the app, which accounted for some 80% of the team effort.

Testing and vibe debugging this app, as prescribed by the algorithm from Figure 7, took us, in total, 56+ work hours: 24+ hours for

the first iteration, 8+ for the second one, 4+ for the third one, and 20+ for the final one. Google AI Studio Pro worked for it, in total, 3h19’: 10’ for the first iteration, 1h42’ for the second, 1h18’ for the third, and 9’ for app publishing and republishing.

Results and Discussion First Iteration

In the initial iteration we tested and vibe-debugged the initial app, based both on the spec doc from Figure 1 of [1] and the initial

Google AI Studio reply shown in Figure 2 of [1].

First, please note from Figure 2 that Google AI Studio completely ignored the first part of our prompt from Figure 1, in which we

thanked it and enumerated what we liked in the generated app. Consequently, we did not do it anymore in the next iterations.

Dually, we were then impressed not only by the fact that it understood and applied our first meta-axiom but added to table Excur-

sions the intelligent unique semantic key clasaId • obiectivId • dataExcursie, which does not contain the superfluous (for uniqueness) CostPerElev (CostPerPupil) column.

A3 was partially implemented: every form and report got a Copy (Copiază) button with which you can copy to clipboard their in-

stances, but you still cannot select and copy only subsets of them (e.g., cells, rows, columns).

A4 and ER3 were wrongly implemented: an Export .accdb button was added on the External Data page, through which you can obtain

such a db backup copy in your Downloads folder, but, unfortunately, MS 365 Access cannot open these files (see Figure 8). Consequent-

ly, we dared not test in this first iteration the dual Restore button. Moreover, the .json backups were removed, although .json restores were still advertised.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

42

Figure 8: MS 365 Access error message when trying to open a db backup copy. Reports’ columns may not be resized, only forms’ and queries’ ones are resizable (A5). Sometimes, attempts to resize columns crash

the app.

As you cannot select columns, you cannot either change positions between them or sort data on several columns (A6). However,

in this new version you may sort data in forms and queries on any column. Data filtering was generalized to all forms, queries, and

reports (A7 and ER1), but again only on the whole corresponding instances (e.g., if you add a “1” filter to the first financial query, its displayed instance will contain all rows containing the digit 1 in either pupil birth date, class, school, locality, or county name, trip cost, date, or payment date).

ER2 was also partially implemented: you can also modify SQL SELECT statements for both query and reports and, apparently, run

them but their updates are neither saved, nor at least temporarily updating the corresponding instances. ER5 to ER8 were partially implemented as well, by adding grand totals, but without group ones.

A8, A9, A10, A13, ER4, ER9, ER10, as well as bugs 1 and 3 were fully and correctly implemented.

On form Elevi (Pupils), requests 1, 2, 3, and partially 5 were correctly fixed. Unfortunately, number 4 was not solved: you can still

update birth date to any value, including in the future, cases in which Age (Vârsta) is set to 0.

All bugs (6 to 9) for form Scoli (Schools) and its subforms Clase (Classes) and Elevi Inscrisi (Enrolled Pupils), as well as those for forms

Excursii and its subform (10 to 14), and Judete (Counties) and its subforms (15 to 17) were corrected.

Bugs 18 and 19 (for both Debtors query and report), 20 and 21 (for the financial queries), as well as 22 (for the last report) were

also corrected.

Second Iteration In the second iteration we tested and vibe-debugged the second app version, based on the debugging specs from Figure 1 and on

Google AI Studio reply from Figure 2.

Meta-axiom A3 was fully and correctly implemented for forms and queries, but not for reports as well, despite Gemini’s claim (where

this pompous “meta-axiom” stands for “best practice rule”). The new app version provided several backup options: -

.accdb, which, unfortunately, was still useless, as MS 365 Access does not recognize its output; .xml, which can then successfully be imported into a .accdb file, but its tables lack constraints;

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

-

43

.sql, which exhibits the following issues, when trying to run it in a .accdb file:

 The comment line between the CREATE TABLE and INSERT sections must be deleted, as MS Access accepts comments only in the SQL script headers.

 The VARCHAR data type is not accepted and must be replaced by TEXT.

 The ON DELETE CASCADE subclauses of foreign key definitions are not accepted and must be removed.

 Each CREATE TABLE and INSERT INTO statement must be executed apart, as MS Access does not accept scripts containing more than one statement. Fortunately, you can create the db schema by running the CREATE TABLE statements one by one

-

(in the order in which they are generated in the .sql script) and then populate its instance by importing only the data from the .xml file.

.ps1, which, when run with Windows PowerShell, providing the .xml file is in the same folder, calls your installed MS Access ver-

sion to create a corresponding .accdb file (lacking table constraints, unfortunately); about it, we discovered an incredible junior

programmer type error: the generated .xml file is called “ExcursiiScoala_MSAccess365.xml” but the generated .ps1 file looks instead for a file named “ExcursiiScoala.xml”.

We also tried to run the .sql exported script in a db created under MS SQL Server 2017 [19], through the SQL Server Management Studio 21 [20], after making the following mandatory modifications:  Replace AUTOINCREMENT with INT IDENTITY.  Replace CURRENCY with [SMALL]MONEY.

 Delete the ON DELETE CASCADE from the foreign key definition of column ObiectivId from table Excursii.

 Enclose all inserts of each table (generically named T) between SET IDENTITY_INSERT T ON and SET IDENTITY_INSERT T OFF, as the script inserts values into the autonumbering primary keys as well.

We were very surprised by the ON DELETE CASCADE subclauses, because of the following 3 considerations:

1. Adding it by default to any foreign key is a very bad choice, as it can create cycles: this is why we had to delete it from the foreign key definition of column ObiectivId from table Excursii, as it would create the cycle shown in Figure 9, which is not accepted by the SQL Server (it displayed the error message “Introducing FOREIGN KEY constraint ‘FK__Excursii__Obiect__534D60F1’ on table

‘Excursii’ may cause cycles or multiple cascade paths. Specify ON DELETE NO ACTION or ON UPDATE NO ACTION, or modify other FOREIGN KEY constraints. Could not create constraint or index”.).

Figure 9: Cycle of the app’s db foreign keys that MS SQL Server refuses to close with ON DELETE CASCADEs. 2. Using it on fundamental data tables is very risky, as it can cause automatic, invisible data loss. For example, in this db scheme, deleting one county results in deleting all its localities, together with all their touristic objectives and schools, together with all their classes, together with all trips they organized, and together with the pupil enrollment in those classes, plus their partici-

pations in class organized trips; i.e., one row deleted by an end-user in the counties table can result in the automatic deletion of hundreds of rows in other 7 tables of this db.

3. Perhaps the worst issue about it is the fact that Google AI Studio added this subclause to all db foreign keys without being requested and without letting us know in the first two versions of this app.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

44

When pressing the Restaurare Bază de Date (Restore Database) button, the app opens the MS Windows file picker window and ac-

cepts as inputs files of type .accdb, .xml, .json, and .sql. As the .json export was removed, we tested only the other 3 possibilities: the

.xml and .accdb work ok (but only for .xml you get a “successful restore” message), while the .sql one did not (we got the message “No number after minus sign in JSON in position 1 (line 1 column 2)”).

Report columns may be resized in this version and attempts to resize those of forms and queries did not crash the app anymore, but

they did not work ok.

Form and query columns can switch positions between them. You can also sort on multiple columns, unfortunately even including

the same column several times (e.g., see Figure 10).

Data filtering finally worked ok, not only on the whole form/query data instances, but also only on selected cells/columns/rows.

Underlying query and report SQL SELECT statements (WHERE, GROUP BY, and ORDER BY) can be edited, saved, and run correctly

(see, e.g., Figure 11). We asked Gemini to also add the HAVING clause.

The Total Debt computed column has been added to the second financial report.

Figure 10: Sorting the Elevi (Pupils) form on multiple columns is possible even on several occurrences of the same one.

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

45

Figure 11: Editing and saving SQL SELECT statements of forms and reports.

Figure 12: Group and grand totals added to the orthogonal report. Almost all reports have the requested group totals as well (see, e.g., Figure 12). Unfortunately, the first hierarchical one lacks group-

ing and corresponding totals on schools, localities, and counties.

The instances of the 3 selection combo-boxes of the two action queries have been sorted ascendingly. Finally, the meta-axiom A14

was correctly implemented, and the associated bug was eliminated: not even manually can you save birth dates in the future or before 2013.

Third Iteration In the third iteration we tested and vibe-debugged the third app version, based on the debugging specs from Figure 5 and on Google

AI Studio reply from Figure 6. PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

46

Report cells may now be selected and copied into the clipboard. You still cannot do it for rows and columns, but we did not press

further with these features, as they are not provided by MS Access either. Consequently, we considered the A3 issue test passed. All A4, A5, A6, ER2, and ER6 related issues were perfectly corrected. Fourth Iteration

In the fourth iteration we tested and vibe-debugged the whole third app version, to make sure that no regression occurred or that

no new bugs were inflicted during debugging.

Fortunately, it was not the case; on the contrary, we discovered that the Creare (Create) submenu, which we did not ask for, was

populated with buttons opening the Debtors query and report, as well as with a new query that we did not ask for, computing some db

totals: total numbers of schools, classes, pupils, trips, pupils participating in trips, total trips revenue, and number of calendar years for which there is stored data (see Figure 13).

Figure 13: Db totals query. Of course, that other such totals might be interesting as well (e.g., total trip debt to be still collected, total number of touristic objec-

tives visited during trips, total number of localities and counties for which the db stores data). Moreover, this submenu (whose name is

rather misleading) could be suppressed and the db totals query button moved as the 7h hierarchical one in the 2nd queries submenu

(see Figure 20 from [1]), but, pressed by the deadline to provide this app to the colleagues of the second author of this paper, we finally considered it ok.

Please note that Google AI Studio spent, in total, some 4.5% of its time for publishing and republishing, 5% for building the initial

app version, and 90.5% for bug corrections and rebuilding the latter two app’s versions.

Please also note that, initially, we did not ask for anything related to the underlying db structure or SQL programming, so non-de-

velopers should not be at all afraid of the technicalities contained in this paper: we also addressed them only because Google AI Studio

provided them and the first author of this paper happens to also be a computer engineer curios to evaluate current AI tools state-ofthe-art.

It is also interesting to note, in this technical vein, that Google AI Studio is almost a true expert in db architectures and SQL devel-

oping, but still struggles with software (probably Python) developing, and, especially, with providing useful and user-friendly graphic interfaces, which are a standard, not only in MS Access, since over 3 decades already, as well as with automated testing. PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

47

Conclusion The main contributions of this paper are the following:

1. Providing a broader definition of vibe coding, by including non-developers and other natural languages than English. 2. Introducing the definition of a new concept called “vibe debugging”.

3. Providing a sound, complete, and optimal pseudocode algorithm defining our proposed methodology for testing and vibe debugging vibe-coded apps.

4. Applying this algorithm to a medium-complex app, step by step, thus proving that even non-developers may obtain with Google AI Studio a bug-free and functional satisfactory app, even if they do not use English as the interaction language with this AI tool.

5. Proving that Google AI Studio Pro based on the Gemini 3.8 Flash version has some excellent capabilities but is still making both architectural and developing errors (some of them even incredibly not intelligent at all), so that both testing and vibe debugging

vibe-coded apps is mandatory, despite its sophisticated Code Execution internal framework, and its claims, repeated at each iteration step, that the code it generates is verified.

6. Pinpointing Gemini 3.8 Flash engine powering Google AI Studio Pro’s failures, but also its excellent and good points.

We started with an initial vibe-coded real-world app of medium complexity generated by Google AI Studio Pro that we tested and

vibe-debugged recursively then, according to our methodology abstracted by the algorithm ATVBM from Figure 7. After 4 iterations, we managed to obtain a nearly perfect and bug-free app version, which we could then share with our interested colleagues. The work

has been done in Romanian, by a team of one non-developer (who did some 80% of the tasks) and one developer. The app’s GUI is in Romanian.

Conflict of interest The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be con-

strued as a potential conflict of interest. Acknowledgements

This research was not sponsored by anybody and nobody other than its authors contributed to it. The authors are grateful to Mihaela

Virginia Mancas and Diana Christina Mancas, who are always carefully checking all our manuscripts.

References

1. Mancas Ch and Mancas C. “Vibe Coding with Google AI Studio Case Study”. PriMera Scientific Engineering 9.4 (2026): 50-73. 2. Google AI for Developers. “Build with Google AI Forum” (2026). https://discuss.ai.google.dev

3. Google AI Studio. “ExcursiiScoala: Bază de Date” (2026). https://excursiiscoala-gestiune-excursii-colare.ai.studio/ 4. Codd EF. “A relational model for large shared data banks”. CACM 13.6 (1970): 377-387.

5. Abiteboul S, Hull R and Vianu V. “Foundations of Databases”. Addison-Wesley, Reading, MA (1995).

6. Mancas C. “Conceptual Data Modeling and Database Design: A Completely Algorithmic Approach. Volume 1: The Shortest Advisable Path”. Apple Academic Press, Waretown, NJ (2015).

7. Microsoft Corp. “Access help & learning” (2026). https://support.microsoft.com/en-us/access

8. ISO/IEC/IEEE. “Welcome to the Home of Software Testing Standards” (2026). https://softwaretestingstandard.org/

9. Pezze M and Young M. “Software Testing and Analysis: Process, Principles, and Techniques”. Wiley Inc., Hoboken, NJ (2008). 10. Myers GJ., et al. “The Art of Software Testing. Third Edition”. Wiley Inc., Hoboken, NJ (2012).

11. Jorgensen PC. “Software Testing: A Craftsman’s Approach. Fourth Edition”. CRC Press, Boca Raton, FL (2014). 12. Mili A and Tchier F. “Software Testing: Concepts and Operations”. Wiley Inc (2015). 13. Selenium. “Getting Started” (2026). https://www.selenium.dev/

14. iiTWORKFORCE. “HP UFT Real-Time Project-Based Training” (2026). https://www.iitworkforce.com/building-work-experi-

PriMera Scientific Engineering

https://primerascientific.com/psen

Recursively Testing and Vibe Debugging a Google AI Studio Generated Vibe-Coded Database Software Application

48

ence/quick-test-professional/

15. Smartbear. “Agentic QA for modern software teams” (2026). https://smartbear.com/

16. Lops A., et al. “A System for Automated Unit Test Generation Using Large Language Models and Assessment of Generated Test Suites” (2024). https://arxiv.org/html/2408.07846v2

17. Chudic A and Calikli G. “Automated Test Suite Enhancement Using Large Language Models with Few-shot Prompting” (2026). https://arxiv.org/pdf/2602.12256v1

18. Google Corp. “Google AI Studio performs automatic testing and verification of its generated code primarily through its native Code Execution mechanism” (2026). https://www.google.com/search?q=documentation+on+automated+test+technologies+used+by+code-generating+LLMs

19. Microsoft Corp. “SQL Server Technical Documentation” (2026). https://learn.microsoft.com/en-us/sql/sql-server/?view=sqlserver-ver17

20. Microsoft Corp. “What Is SQL Server Management Studio (SSMS)?” (2026). https://learn.microsoft.com/en-us/ssms/sql-server-management-studio-ssms

PriMera Scientific Engineering

https://primerascientific.com/psen

Related documents

• The Case for Vibe Modeling: A Missing Step in AI-Based Trustworthy Software Development#414174
• FlowCheck: Helping End-Users Specify and Verify Intent in Vibe-Coded Web Apps#793115
• FlowCheck: Helping End-Users Specify and Verify Intent in Vibe-Coded Web Apps#789111
• Vibe Coding with Google AI Studio Case Study#1324704
• Vibe Coding with Google AI Studio Case Study#1319136