How Claude Helped Decode the Super Mario 64 Plaque Mystery

For decades, one of gaming's stranger mysteries has sat in plain sight: the tiny plaque on the star statue in the castle courtyard of Super Mario 64. Many players read it as "L is real 2401", helping fuel the old rumour that Luigi was hidden somewhere in the game.

A new AI-assisted reverse-engineering project argues that the phrase is most likely something else entirely: "Eternal Star / We hope so". The project claims the text was set in Palatino Bold Italic and that the famous Luigi reading fits the pixels much worse under its scoring approach.

That is a lovely result if it holds up. But the more useful lesson is not really about Mario. It is about how AI should be used when the evidence is messy, low-resolution and easy to overinterpret.

The Super Mario 64 plaque was not an image enhancement problem

The plaque texture is tiny: 32 by 32 pixels, with letters reportedly only 3 to 5 pixels tall. At that scale, the image does not contain enough visible information for a human, or an AI upscaler, to simply "read" it with confidence.

This is where many AI image tools become misleading. A super-resolution model can make an image look sharper, but sharper does not mean truer. When information has been crushed into a handful of pixels, the model has to fill in gaps. Those gaps can become invented letter shapes.

The project's key move was to stop treating the plaque as a blurry photograph and start treating it as an inverse-rendering problem. In simple terms, inverse rendering means working backwards from the final pixels to ask: what original text, font and production process could have created this image?

The associated public GitHub project describes a reproducible approach rather than a one-off visual guess. It models a plausible production pipeline: font, canvas, shrinking, blur, stone texture and scoring against the real game texture.

What Claude appears to have contributed

The careful wording matters here. Claude did not simply look at the plaque and read the hidden truth. The reported role of Claude, specifically Opus 5.5, was to help build and iterate the workflow.

That workflow included writing code, testing hypotheses, modelling how a 1996 texture artist might have created the asset, running searches across fonts and settings, and preparing a report. A GPU then tested many possible combinations to see which candidates reproduced the original pixels most closely.

This is a much stronger pattern than asking a model, "What does this say?" It turns the AI into a research assistant rather than an oracle. The model helps generate methods and code, while the verification happens outside the model through repeatable tests.

That distinction is worth remembering for developers and business owners. The value is not that the AI "knows" the answer. The value is that it can help you build a system for finding, testing and rejecting answers.

Why "Eternal Star / We hope so" is plausible, not confirmed

The project says one answer matched far better than the famous "L is real 2401" reading: "Eternal Star / We hope so". It also claims the second line was found through a blind search over English words when using the proposed font and rendering pipeline.

That does not mean Nintendo has confirmed the text. It has not been disclosed that Nintendo or the original asset creator has verified this result. The evidence is currently self-published, although the code and report are presented for reproduction.

The right conclusion is therefore: this is a strong computational argument for a likely source phrase, not a final historical record. That may sound less exciting, but it is more honest - and in AI work, honesty is the whole game.

The project also appears to preserve uncertainty in the right places. According to the research summary, "Eternal", "Star" and "hope" are described as strongly supported, while "so" is best-fitting but with a smaller margin. That kind of caveat is a sign of better analysis, not weaker analysis.

The AI hallucination lesson hidden inside a Nintendo mystery

This story is a neat example of the difference between generative gloss and verifiable analysis. A hallucination, in AI terms, is when a model produces a confident answer that is not grounded in the evidence. With images, that can mean inventing plausible details that were never present.

The broader computer vision problem is well known. The perception-distortion tradeoff describes the tension between images that look realistic and images that remain faithful to the original signal. That is exactly the trap with an unreadable game texture: the most readable reconstruction may not be the most truthful one.

This is why the plaque case is more interesting than a simple "AI solved a mystery" headline. The useful approach was not to ask AI to imagine a cleaner plaque. It was to constrain the problem, create candidate originals, render them forward and compare the outputs to the actual pixels.

That same principle applies well beyond retro games. If you are using AI to inspect old documents, product defects, CCTV footage, medical imagery, archives or technical drawings, the question should not be "does the AI answer look convincing?" It should be "what independent test would prove or disprove this answer?"

What UK teams can take from this AI-assisted workflow

For UK organisations, the practical lesson is about disciplined AI adoption. Claude and similar tools can be excellent at helping with code, search strategies, documentation and experiment design. They are less reliable when treated as final authorities.

A sensible workflow looks like this:

• Define the evidence. Decide what the ground truth data is and keep it separate from the model's suggestions.
• Generate hypotheses. Use the AI to propose routes, write scripts and identify edge cases.
• Test externally. Score outputs with deterministic code, known samples or human review, not vibes.
• Record uncertainty. Say what is strongly supported, what is weakly supported and what is not disclosed.
• Make it reproducible. Share enough method, code or audit trail for another person to challenge the result.

This is also where governance matters. If you are applying similar techniques to business documents, customer data or internal records, do not casually upload sensitive material into tools without considering data protection, supplier terms and your responsibilities under UK GDPR. This is not legal advice, but it is common-sense risk management.

I made a similar point in a different context when looking at how generative AI can create real-world misinformation risks. The danger is not only that AI gets things wrong. It is that it gets things wrong fluently enough for people to stop checking.

Claude as a coding partner, not a magic decoder

The plaque project also shows why coding agents and advanced assistants are becoming genuinely useful. They can help turn a vague idea into a working experiment faster than many people could manage alone.

That does not remove the need for expertise. In fact, it increases the value of good judgement. Someone still has to decide whether the modelled production process is plausible, whether the scoring method is fair, whether the font search is biased and whether the final claim is overstated.

This is the same pattern I see in more practical settings, from local AI workloads to software prototyping. The best results tend to come when people use AI to compress the tedious parts of the work while keeping human control over assumptions and decisions. That was also the core argument in my piece on using Claude alongside local AI workloads.

The real win is better verification

If "Eternal Star / We hope so" turns out to be the best explanation, it is a charming bit of gaming history. It also has a pleasing symmetry: a small message about a star, hidden inside a game that became a landmark of 3D design.

But the bigger win is methodological. This is AI used at its best: not as a shortcut around evidence, but as a tool for building a stronger path towards it.

The plaque was too small for enhancement to be trustworthy. The smarter move was to search for the source that could have produced it. That is a useful lesson for anyone using AI in 2026: when the answer matters, do not ask the model to make the blur look clearer. Build a test that shows what the blur can and cannot support.