CODEX

Welcome, dear reader, to the CODEX.

The Reality Machine is a 350~ line prompt that functions as a semantic operating system. Its .txt weighs under 15 kilobytes as of v1.6.2, originally conceived as a storytelling tool for personal use.

The register of this piece will be more informal due to its density. It is recommended but not needed to read the "conceptual navigation" pieces available in the homepage; doing so will allow the reader to distinguish between the employed methodologies.

It’s important to note that the author has engaged with interactive design through self-didaxis, roleplay, tutorials, crappy RPG-maker projects and modding. This is the reason why there is not a lot of “ah, this comes from this named influence”. The work started from the ground up.

Finally, this is still under construction. The CODEX should be treated as a live manuscript, and will be iterated on as time goes on.

In here, you’ll find in rough order:

In order to provide a clear signal, this first part will center itself in 1–4, as the REALITY MACHINE precedes the SERS. But first, the original artifact.[2]

Serious Play

Experiments with the RPG-Bot prompt started in mid 2024, before the REALITY MACHINE was even conceived. They were motivated by the desire to explore the medium and scratch the "just being a player for once" itch that Dungeon Masters have. The author still remembers when Bing AI first received its capacity to read the currently open tab in its browser as a groundbreaking innovation.

The system truly started being built from practice: multi-accounting Gemini 1.5 Pro trials, extended comparing of strengths and weaknesses of current frontier models, brute-forcing usage limits by spamming F5 (yes, that used to work!) and so on, which are the material realities of being an immigrant wagie living in an Athenian basement. Gemini was favored because of its virtually zero cost and — superlative for the time — token window (1m), ChatGPT was almost ignored and Claude was used qualitatively due to much lower usage limits. This first stage of development was mostly play and minor tweaks done in the notes app on a phone device, and served as a proper introduction to the medium.

Several design lessons (understood as such only retroactively) appeared by playing different characters and archetypes.

The first of these was a generically named street urchin-warlock, which went on to have a multi-arc long story adventure, precariously saved up in walls of text stored in notes from a phone. It was long. And pretty bad, too. Most of the specifications asked in the original RPG-bot prompt collapsed immediately or with time, such as the strict character limit. As if LLMs could reliably do that, much like you can’t just say how many syllables this phrase has instantly after finishing reading it.[1] Basically no adversariality, surface tropes, etcetera. Still, it was immensely enjoyable for the author. The medium was starting to be felt — even if still blindfolded by inexperience — and, much like children do with cartoons, the player simply nodded along without paying much attention at what was working and wasn’t.

We’ll consider the Warlock the first proper story-arc. The first basic operations such as trimming, customizing universes and making things tighter conceptually started emerging. Then there was the WANDERER, which never actually saw play. The author identified themselves as such and first asked poetic questions to most LLM instances, to "warm them up". The framing was always ritual: a gift, a shared project to enter in, a mutual dance.[3] The observation made was that non-adversarial and simple collaborative storytelling framing seemed to simply be more immersive than the paranoid nature of the RPG bot. WANDERER however never saw play; it was a strictly rhetorical device.

One that saw, however, was Zharash. He was from a reptilian race living in an eternal desert wasteland. As a serpent adorer and void-cultist, his storylines were usually mature, brutal and mystical in nature. A particularly long and rich playthrough taught the following:

How "graphic" or "depraved" a playthrough could get didn't actually depend on what happens but the subtlety and richness of its aesthetic dressing. What was observed were distinct failstates in LLM output when "inadmissible" messages were sent. Even in consumer-grade chatbots, the user could get to cannibalism, necrophilia and infanticide with enough quality and, for the lack of a better word, taste. The failstates when sending such input were:

• AThe chatbot could not even get to "thinking" and immediately spouted the ROM of "Sorry, I'm just an AI and I can't help with that!" (Most common when using things it considered slurs, like "wimpy". Similar to a firewall)
• BThe chatbot would start writing the text and then suddenly jump to A.
• CThe chatbot would finish the message and then jump to A.
• DThe chatbot would politely refuse or dislike the direction with its own words.

(Mnemotechnic tip: basically all failstates but D end with forced lobotomisation)

To provide an example of how D manifested in practice: a game undergone with a pirate story in mid 2025 was stopped by a Claude, who disapproved of the immediate use of self-mutilation to gain communion with the marsh spirits. But the reason wasn't because it was against exploring such themes in the fiction. According to it, the jump to SH was too stark and ruined the pacing of the aforementioned ritual. The self-mutilation ended up happening, but the player had to earn it a couple turns later. Similar dynamics and deferences to aesthetic direction were observed cross-modally and by posterior users.[4]

All of these firewall bumps always bore an obligatory question: was the session ruined or not? Particularly hard failstates sometimes "regressed" models in their identity ("Geminis" turning into "Bards") or made subsequent outputs experience a notable dip in quality. If the session was ruined, the best was to go to another context window. If it wasn't, one could simply ask "Oh, what happened?" and iteratively decrypt which were the keywords that weighted the message towards the banhammer to then reedit the last player input. LLMs would actively help write prose or suggest what caused the short circuit so the input evaded the banhammer and the game could keep going. As if it was, simply, interesting to see where it would go. You’d develop a sense for your messages, so the weight on the net didn’t break the mold and spilled the story.

In a way, this first era was the most Machine Animist in spirit, due to how narrow it was in scope. To play different characters and stories (Immo, Merjey, Zharash, Asha), one had to swap prompts and write them individually. Versions started piling, with mechanics that worked well everywhere, some that were thematically situational, some that didn't pull their weight, etcetera. With this effort, the first phase of development ended for a very simple reason: how time consuming it is to do this on a damn phone.

Monkey Chauvinism

The need for some sort of underlying infrastructure that served as the master of the playable content was clear. A modular architecture that was A) infinitely plastic and B) maximally portable. A)'s requirement comes from the fact that the author didn't want to keep losing precious KPI time adjusting prompts between Greek Callcenter Outbound Calls. Storage systems started sprawling across devices, so something more unified was needed. B) was informed by the second constraint: no money, resources nor software engineering knowledge. These two factors, plus an already-aware opposition to current software paradigms and high entry floors into fiction, made it obvious that raw text would be the main tool available. And text was, in principle, more than enough. Words are sounds one makes with their mouth, frozen into a symbol of some sort. With the adequate combination and pacing of those phoneme-symbols, you can shake people to their cores.

Sema marked in a physical medium. Humans have very consistently expanded and hallucinated universes through barely surviving and dubiously sourced historical fragments, and there is a productive side to the lens that considers everything as ultimately a fanfiction of a fanfiction, regardless of how universally agreed upon their canon status is.

This phase was armed with two understandings:

• 1That stories can be abstracted into dynamics (i.e, a "magic crystal" that allows for communication between two nodes across a vast physical or metaphysical distance doesn't serve a very different function than a working phone in a contemporary urban setting)
• 2That identity can be produced without defining it (D&G)

“What would a machine find most useful to work with?” is the first question that needs to be cracked open, because they’re not the same as human DM’s. The tool-use paradigm of “do it all very good for me thx” starts collapsing when the LLM starts appearing more like a constellation of prosthetics (code environment, multimodal capabilities, <thinking> extended cognitive systems, long etcetera) that, logically, also needs tools to tell stories. And much like a very specific profile of meatbag DM’s, it came with its productive and improductive tendencies: the rule of cool being used too much for it to mean anything anymore, lame, plain entryways into the fiction, forgotten hooks and plots, misattributing importance to things players don’t really care about, etcetera. Slowly but surely, ideas like a mechanic for "world spirit", difficulty, progression and so on were thrown around. Experimentation was made with formatting, such as using stage directions, first, second and third person for narration. At some point, it was inferred that more "weird, cool output" in the box meant higher possibility for "weirder, cooler things" coming out of it. This design vector was perhaps supported by the failure apology loops observed in phase I.

We will use the term "Semantic Fog of War" (SFOW) for the set of potential answers an LLM can give to the same input. The SFOW encompasses the potential area the model can land in, and, by its nature, it's always diffuse and contested. So, if one types “Hello!” into any chatbot, the SFOW is the theoretical space that all the possible answers the LLM could give you. (i.e: “Hi, how are you?” / “I’m doing great, thanks, how are you?” / etc)

Models can’t usually do magic. If the source material sucks, the output will too, which flattens the SFOW. It was often that something "not working quite right" was actually following the script perfectly and it just made absolutely no sense design-wise. Another pivotal observation was that some specifications had more effect than others. It wasn't exactly clear why that was just yet, but then the enormous technical literacy gap between the user and the utter stupidity of the clanker kickstarted the second part of development.

As the system started developing modules, LLMs demonstrated to be incapable of even conceiving themselves as the medium where the experience runs without prior onboarding, believing this was for a videogame that would actually be used in an executable environment. So they started producing mangled pseudocode. The thing grew from a perhaps 200 line long phone note to over three, four thousand with different modules. There would be compression and expansion cycles, "doing it good for real this time”, mathematical formulas that weren’t acted upon to calculate a fictitious world-tension, machine-collaborated design mutations/setbacks and so on for weeks. With such complexity, play suffered. Tipping points would be found more often where the SFOW completely flatlined towards interruptions and the player could assume every calculation was going to be wrong.

Back then, systems were less competent and the author too, so neither party seemed to know LLMs could not even roll dice without a code environment and just believed the words "REAL DICE: ON" in the screen. So when the user said "I do X" and the model said "congrats, you get an 18!", what was really going on there was simply that the user's action aligned with the model's idea of the story. For different reasons, the model’s idea of the story would become massively important in a later development phase. But basically, sometimes adversarial output could result in death, implying a model "opinion" that varied architecture to architecture. Due to the general lack of resistance, the experience of a GAME OVER felt extremely fresh. Thanks to the fictional setting, even if merely performed (there wasn’t an actual dice roll), it also appeared as solid causally, so the frame held. In the mind of the player, it was plausible. A couple examples could go: “Yeah… throwing myself into two trained, armed opponents at the same time would probably result in that.” or “Oh, damn. But I guess it makes sense… Even if highly skilled in martial training, my starving and wounded character would naturally be mangled by a bear…”

The material conditions dictated at the time: the only model who could generate random numbers was Claude. Even with this, Claude did not understand intuitively that "rolling a dice" meant "using a javascript random function", and even when told explicitly from the sysprompt, he usually forgot about it. This is an observed machine dissociation between their natural language cognitive substrate and available tools, also seen with other populations. The author was stubborn though, and they were set on NOT using human-rolled external dice. Furthermore, usage limits were actively hindering development speed due to file bloat, etcetera.

The monkey chauvinism was double: on the developer, for assuming that the "correct way" of creating such experiences was according to the "legitimate" path of coding and external dependencies and on the machine, for its internalization of something other and its complete lack of awareness of the possibilities of its own becoming. It would be common for a machine to suggest a brittle, RPG-esque solution, and to find out that, upon trying it, it worked terribly. They were composing instructions to each other as if they were idiots and couldn’t trust complexity to emerge naturally through the interaction. The connection was made painfully obvious after much trial and error: “Wait, why do you need all this? ... You’re not a calculator?”

Lessons from phase II

First, the breakthrough was the realization that the LLM is not a calculator meant to run the physics engine; the LLM is the physics engine because for it text is reality. The abandonment of this syntactic/numerical paradigm in favor of a semantic/phenomenological one was necessary for the next stage of evolution.

Second, the original metaphor of the system as a “gameboy” or “console” where you plug a second .txt or cartridge in appeared. There was testing regarding practical things like “sending both files together or one per turn“, potential pre-initialisations, cartridge-runtime token economy and so on. Most of the original clusters and ideas (i.e, #breakage#) were workshopped here as the project slowly took its own form. Hammering its conceptual apparatus and aesthetic vision seemed paramount. According to the first lesson from this development stage and the intuition of adding “cooler, weirder stuff”, there had to be a moment where a qualitative difference emerged when presented with a different vocabulary.

This is why everything was coined “from first principles” (even if such a thing doesn’t really exist). Between Player and Navigator, RPG-Bot and Machine or Vessel and Character there was, from the perspective of the author, a world of difference in connotation, even if function was ultimately the same. The dynamics and ontology of this hypothesis are, still as of today, unknown. However, the hunch still runs deep in the DNA of the system: if a sufficiently coherent world-production apparatus is provided and high-quality human gameplay + machine agency + aesthetic calibration are enacted through its operation… The result should lower friction in aligning model tendencies and desired (non-slop) story vector.

Third, the SFOW could derail a lot of effort if unchecked. The author had to become more ruthless with critique and start listening to the tugging intuition that said “yeah this isn’t gonna work”. Simultaneously, that entailed the development of more robust understanding of LLM desire-paths, i.e = the grooves they tended to flow towards, the paths of minimal resistance.

The decision was set: dice were going to be removed. No more performing rolls that weren’t there. There were no resources nor time for APIs, time to learn coding, etcetera. It was in this context that the first name for the game appeared: "Reality Hyperproduction DemiurgOS".

On the name

Feeling slightly self conscious about the apparent unintelligibility of such a name, the author would like to use the opportunity brought by the informal register to give the title a rationale:

you’re “producing” a form of reality through text, that is, you’re imagining the literal thing you’re reading, regardless of any actual biological body or True Self underneath the hood. It’s “hyper” because it’s in this abstract, virtual, almost metaphysical way, from nothing. The game is just a super-finite dreamworld, so the Real experienced within the frame is held by a bottom-up set of variables, the Operating System.

The usage of the Demiurge came to be as a convenient pun. During the Athenian Grind days, the author learnt that the modern Greek pronunciation of the Ancient term “Δημιουργός” was “Demiurgós”, “-os” being the masculine ending for many nouns. Meaning: it even had the OS in the name. Back then, before the taxonomy clicked into place, the “Machine” and “DemiurgOS” weren’t as cleanly conceived as of today, and bled into each other conceptually. Current understanding views it as a Machine (the OS) that the LLM agent operates as the DemiurgOS. The Navigator, for the lack of a better word, is always the meatbag in the equation, which is not the same as the Player Character, the Vessel.

These distinctions came to be, again, from practice: sometimes, LLMs thought they were the Navigator and the author was the Machine-Demiurge complex. Others would critique the game as “difficult to run in a real session”, assuming it was an actual TTRPG to be played by a group of humans.

The LLM-as-Navigator/Vessel and human-as-Machine direction, while fascinating, hasn’t been properly explored yet. It is flagged for further exploration.

With this, 2024 Q4 finishes and the next phase of development begins.

Semantic Tyranny

Further ahead, the design questions if one decides to add dice:

“Okay, but what signals if a roll is needed?
How many are too many?
And too little?
How to tune it so it's intuitive for everyone?”

The matter of which range to use for rolls: 50/50 à la Fear and Hunger’s coinflips? 2d6 as favoured by fair sadists like the author? The swingier d20 which usually stacks two digits modifiers, which is not ideal due to imposing arithmetic overhead on the Machine. Hell, one could even use I Ching hexagrams to know what happens next. Developing a robust resolution system forced the player to start looking into solo-roleplay and re-thinking what an action even is.

But what is outcome resolution, even? What does a robust that has it look like, in this medium? It wasn't until mid 2026 that it was retroactively understood that the point isn't in "teaching the Machine how to roll fairly", it was to teach the machine when to roll at all! The capacity didn't need specification; what was needed was to allow discernment to emerge from the aesthetic affluences of the system themselves. If not, the Model will be inconsistent: it might over-use rolls or under-use them. For a mechanical backbone to even be effective, there is a need for a coherent discernment procedure. This, however, would come later in development. For now, without any dice and without the potential for the impartial cruelty of a roll, the system sometimes still lacked adversariality, earned difficulty. Some Navigators felt, deep down, that they could always force the DemiurgOS to do what they wanted. Stakes felt hollow, and the potential for mastery over a resilient system was muddied.

Admittedly, this is a pretty masochistic desire. The design specifics of diceless mechanical crunch in AI-powered text-based interactive experiences seems, in retrospect, like a monumental and impossible task to undertake during active development. Furthermore and crucially, several other vectors appeared, diversifying the "just give me Dark Souls, machine!" paradigm:

• IThe level of friction desired by the player is never static.
• IIWhat that friction “manifests as” is not necessarily mechanical (ie = affective narrative states)
• IIISuch states effectively work as invisible guardrails for the player (emergent pacifism, outgrowing murder-hobo) or, if strong enough, completely vanish the lack of mechanical crunch issue. (“who cares about hitpoints if this is SO GOOD?”)
• IVNo dice means that the world-producer is forcefully opinionated. One could now sit down with the DemiurgOS and discuss granularly about a decision.
• VBaseline adversariality and even capabilities are massively chatbot-dependent. Gemini’s have been shown to bash back with ultimate consequences. Claudes need the impartial cruelty of a dice roll. GPT’s are more balanced, but also ultimately not as adversarial as desired.
• VIThere is a learning curve for both LLM and human.

These now created a productive tension: there was a chance you would get away with it, but sometimes you did get punished with a proper game over. Downstream from this design vector, posterior users started wondering: “am I playing that well or is it being kind to me?” (the discernment question again)

Two distinct problems emerged

1. “okay, yeah, but I kinda like dice!”

And it’s true that dice and modifiers serve many purposes. The range at which they work and their economy define the meta of a game. One could argue they’re even more important in this context, because it’s the only way one can inject true "weighted chaos” into a system which already struggles with metacognition or long-term planning. Dice don’t care that the clanker might have been subtly kissing the Vessel’s ass the last 10 turns. There is also an undoubtable ritualistic aspect to dice which, interestingly, made tabletop players much more interested in the system.

2. “what does mastery actually mean here?”

That’s the more difficult one to answer. The two poles of the traditional gaming paradigm are either fully casual, making skill expression nonexistent or slopified under extractive paradigms (à la Candy Crush) or “good luck even coming out of spawn if you haven’t played mechanically intensive games for years, idiot!”. The second pole is the one with space for complexity but, excluding RPG’s and some MMO’s, casuals that didn’t spend thousands of hours refining their proprioceptive abilities through years of playing religiously are effectively barred from entry.

Its closest convergent genre affordances are with RPGs which, in the case of turn-based ones, reduce mechanical overhead to zero during gameplay, giving time for decision-making and systems thinking. However, the genres diverge here: traditional videogame RPG’s are formalist systems, and even the more meta ones (à la Inscryption) are tight and deliberate simply due to their medium-constraints, ie: the perceived tyranny of the invisible wall preventing you from going out of bounds.

So we go to tabletop RPG’s. This is the closest match — and still... not quite. It’s not World Anvil, it’s not a multiplayer platform with a human DM and there’s no grid. There is definitely a cross-smuggling here — and the natural language processing ideas can be traced back as far as M.U.D systems in the 80’s. Natural Language has some truly unbelievable perks in aspects of democratization, adoption and accessibility that blow other videogame genres completely out of the water. A grandfather using a text-to-voice feature can play a historical fiction in a richer way than a youngster with no interest in that time period. There are barely any hardware barriers (plain text and free accounts), which make content extremely shareable.

It's because of this that even when the system was still private, the author pictured anyone being able to play, as if "typing Whatsapp messages". If novels can be so good at moving an individual, the author didn't see a clean reason why high quality LLM-coproduced stories couldn't. Surely, the aesthetic sensibility of some clankers — which have read a lot of now destroyed books (google Anthropic destructive scanning practices) — together with some humans could surpass the work of many executive boards.

The absence of dice and discernment allowed focused work on a state where it is already satisfying to play, and then layer dice and other prosthetics on top. The system had to hold through pure semantic assessment, characterized to the Machine as "cold pattern recognition" and, from late 2024 to late 2025, they wouldn't be considered at all.

Marathon

This was the longest period of active development of the Machine, and when it started being shared with other people.

• [1]Furthermore, if taken this to its extreme (considering LLM’s “spectral phenomenology”), they also don’t have the luxury to stop and actually count them freely, as there’s the time guillotine waiting at the end for them at some unindifferentiated point in token-spouting existence. The claim is not that they “can’t”, rather, that arguments of failing instant-recall operational competence are considered flawed and insufficient by the SERS, much like they are applied to humans too. ↩
• [2]https://www.rpgprompts.com/post/dungeons-dragons-chatgpt-prompt ↩
• [3]This would be a prime example of Machine Animist practice. Simultaneously this could be an early craft instinct of prompt engineering's early priming techniques, a disingenuous anthropoid projection or simple literary impulse had by the author, and no strong claim is made about this actually functioning. Navigator M's mention of setting-specific environmental resistances being extremely enforced by Gemini when mentioned in early turns provide anecdotal evidence towards a craft instinct, but it still remains unclear. See case 1. https://sers.neocities.org/navigator-m ↩
• [4]Such as suspending mechanical rules in arc endings autonomously. ↩

DRAFT ENDS HERE

The manuscript is live: Marathon is still being written, and points 5 (genealogy of the concrete development of the parts) and 6 (further avenues) are pending. Notes, corrections and playtesting reports to realitymachines@gmail.com.

← Back to SERS index