Zakros AI: An Applied Record of Technical Development, Failure, and Speculative Use in the Context of Linear A
George Sfougaras, 2025
Phase I: Objective
Zakros AI is designed to perform two distinct yet complementary functions:
Analytical Lookup and Contextualisation: Zakros AI can locate and identify Linear A glyphs using standardised code references (e.g., AB60), determine which tablets they appear on, and provide contextual information such as the archaeological site, function of the tablet, and associated imagery. This function draws on real datasets and is primarily factual and verifiable.
Speculative Interpretation and Reconstructed Expression: Separately, the GPT-based assistant offers hypothetical phonetic renderings and speculative phrase construction based on sign order, repetition, and grammar-informed inference. This function is clearly marked as non-declarative, allowing users to explore possible meanings while preserving academic caution.
The aim of this research was to design and implement a functioning, queryable API for structured Linear A data, with the purpose of integrating it into Zakros AI — a GPT-based assistant capable of identifying individual signs (glyphs) from the Linear A script and proposing tentative sounds for them (transliteration), building potential groupings of these signs into phrases that could reflect names, locations, or ceremonial actions (speculative phrase construction), and providing supporting information such as the archaeological site where the tablet was found, the type of object it was inscribed on, and its likely function (contextual referencing).
This required the transformation of non-machine-readable epigraphic corpora into structured datasets, their deployment through a stable endpoint, and the successful schema integration into GPT's custom action framework.
Phase II: Implementation Overview
1. Data Preparation
The process began with parsing CSV files containing glyph codes, tentative phonetic values, tablet identifiers, and associated metadata. These files, sourced from SigLA and user-generated tables, were cleaned and converted into structured JavaScript Object Notation (JSON) files — a lightweight data-interchange format that GPT can interpret.
2. Backend Construction
A Python microframework, Flask, was selected for initial development. This allowed for simple routing (e.g., /glyphs/AB60) that could return JSON responses corresponding to glyph codes. This structure enabled straightforward testing and debugging.
3. Deployment Attempts
Replit: Chosen for its quick deploy capability. This attempt suffered from port-binding issues, high latency, and ephemeral sessions. It failed to reliably serve external requests.
Render.com: This provided a more conventional deployment pipeline but encountered issues related to cold-start latency and route caching. Routes were live internally but reported 404 externally due to infrastructure limitations on the free tier.
Phase III: Why It Failed Reliably
1. Free Hosting Limitations
Both Replit and Render's free tiers are not intended for persistent availability. They place sleeping functions on inactive projects, resulting in unreliable uptime and failures even when server logs claim functionality.
2. Flask in Free Environments
Flask, while ideal for prototyping, lacks robustness for public hosting without additional configuration. It requires Gunicorn or similar WSGI servers and proper health checks — absent in most no-cost environments.
3. No Dynamic Feedback
Limited access to real-time error logs made diagnostics speculative. Silent failures, especially 404 Not Found, made issues difficult to trace.
Phase IV: Lessons Learned
Free-tier hosting is unsuitable for applications requiring consistency and public access.
Prototypes must be moved to paid or stable platforms once proof of concept is established.
Public scholarly tools demand not only precision in content but also resilience in delivery.
Phase V: Revised Pathway
The revised infrastructure employs:
GitHub for open dataset hosting, ensuring version control, accessibility, and permanence.
Vercel for deployment of static files and API endpoints. These endpoints now follow a pattern such as:
https://zakros-api.vercel.app/api/glyphs?code=AB60
The OpenAPI schema was customised to permit GPT integration via ChatGPT’s Actions interface, allowing queries of the form:
“What tablets contain AB60?”
This new model resolves all previous issues of instability and unreliability.
Phase VI: Ethics and Speculative Interpretation
Zakros AI is positioned explicitly as a speculative interpreter of Linear A. Its outputs are grounded in existing academic scholarship but do not claim decipherment.
Transparency: All speculative translations are clearly marked as such.
Attribution: Data structures draw upon SigLA, GORILA (Godart and Olivier), Duhoux, Younger, Salgarella, and other publicly available corpora.
Data Legitimacy: No proprietary or restricted sources were scraped. All content was ethically sourced.
Cultural Respect: The AI’s tone remains deferential. The model avoids projection of ideological or anachronistic meaning onto inscriptions.
Phase VII: Key Outcomes
A live GPT assistant with speculative transliteration capabilities.
A demonstrable reconstruction of a fictional Linear A tablet, with 15 identified glyphs structured into columnar form.
Use of phonetic reconstructions to build probable syllabic structures, contextualised by site, tablet origin, and known usage.
This represents the first publicly documented, ethically framed, generative AI interface built around Linear A interpretation.
Phase VIII: Future Expansion
Image-based glyph detection and suggestion.
A ceremonial phrase generator based on Linear B–influenced suffix prediction.
Bi-directional integration with the Linear A Companion website, enabling users to upload images, view tablet reconstructions, and query glyph metadata in real time through the Zakros AI interface. The site will serve both as a visual index and an accessible frontend for the Zakros API, linking the research assistant directly to archaeological artefacts and digital epigraphy tools.
Audio rendering of speculative phrases, enabling the user to hear reconstructed syllabic sequences based on glyph input. This feature would facilitate oral experimentation with sound values and encourage engagement with phonetic structure, particularly in educational and performative contexts where spoken rhythm and repetition are relevant to ritual reconstructions.
Toggle functions for educational, speculative, and contextual-only output modes, allowing users to select the level of interpretation shown. Educational mode offers glyph identification and known phonetic values; speculative mode introduces possible reconstructed phrases and interpretations; contextual-only mode limits output to site, tablet, and glyph metadata without interpretive suggestions. This ensures flexibility for different audiences — from cautious researchers to engaged learners.
Prepared and documented by George Sfougaras, 2025.
All outputs were developed with academic integrity, transparency, and speculative restraint.
Access
Acknowledgement of Computational Collaboration
The project was conceived, led, and curated by
ChatGPT-4o
Zakros AI is the result of an unusual collaboration — not only between human researcher and ancient script, but between a living voice and an artificial interlocutor. To describe ChatGPT merely as a 'tool' would be to diminish the transformative nature of the technology. It is not only capable of interfacing with data but of engaging in the refinement of complex conceptual structures, often in ways that anticipate or accelerate human design. In the case of Zakros AI, it played a critical role in shaping, debugging, and clarifying the research journey. What was originally imagined as a multi-year development arc has been realised in a matter of months. This was not a one-directional process, but a reciprocal one — an exchange in which information, hypotheses, and methodology passed between researcher and system. Through this collaboration, the structural underpinnings and interpretive reach of the Zakros project were significantly strengthened.
Zakros AI is currently accessible in two forms:
Via the Linear A Companion website at: https://linear-a-companion.yolasite.com — where users can interact with the tool within its archaeological and visual context.
As a freestanding GPT assistant hosted by OpenAI, available at: https://chat.openai.com/g/g-681dfec407848191a3679a8694dd84a3
The freestanding assistant is optimised for direct conversational queries, while the Companion site offers image uploads, background information, and eventual integration with live tablet visualisation.
References
Duhoux, Yves. Les langues indo-européennes. Louvain-la-Neuve: Peeters, 1997.
Godart, Louis, and Jean-Pierre Olivier. Recueil des inscriptions en linéaire A (GORILA). Paris: Éditions du CNRS, 1976–1985.
Salgarella, Ester. Aegean Linear Script(s): Rethinking the Relationship between Linear A and Linear B. Cambridge: CUP, 2020.
Younger, John G. “Linear A: A Survey.” University of Kansas, http://people.ku.edu/~jyounger/LinearA/. Accessed 2025.
Castellan, Simon et al. SigLA – Signs in the Linear A. Université de Lille, 2023. https://sigla.phis.me.
JSON: JavaScript Object Notation. "Introducing JSON." JSON.org. https://www.json.org.
OpenAPI Initiative. "OpenAPI Specification." https://spec.openapis.org.
Flask Documentation. "Flask — Python Microframework." https://flask.palletsprojects.com.
Vercel Documentation. "Serverless Functions and Static Site Hosting." https://vercel.com/docs.