This piece is a departure from The Rosetta String’s usual localization analysis - a personal retrospective on Blizzard and twenty-five years of gaming. Overwatch turning ten felt like the right occasion to finally write it - especially since my own 35th birthday was two days ago.

My mom came home from the store a bit apologetic.

She had gone to get me StarCraft - the game my friends had been telling me about before my first PC even arrived. Space, alien races: I was immediately sold. When she asked what game I wanted, there was no question. But when she came home, she said the disc had come with a few extra games she hadn’t expected. “I figured they wouldn’t hurt,” she said. I was like, “All right, why not? Extra games are nothing to complain about.”

Those extra games were Warcraft and Diablo 2.

If you grew up in the West, you might ask how those three could possibly come on the same disc. I grew up in Uzbekistan, and licensed CDs were nowhere to be found. Only pirated games were available, no matter how hard you tried. They came in these bundles, and those bundles came stripped - I hadn’t seen any Diablo cinematics or heard its audio dialogue until I found them on YouTube years later. Everything that could be cut while keeping the game playable was cut.

I was 10, and the PC had come pre-installed with Quake 3, Red Alert 2, MechWarrior, and Delta Force. StarCraft was the first game I tried from the bundle. The campaign was interesting, but RTS was hard for a kid with no one to ask for guidance. So I moved on to Diablo 2 - and that one got me right away. I was absolutely lost to it.

Just getting through the first act took me weeks. I remember finding my first green item - an Infernal Cranium cap - and being completely flabbergasted. I didn’t even know such items existed.

Eventually I moved on to other games, exploring the world of PC gaming that had just opened up in front of me. I played Warcraft, quite a lot of Quake and Red Alert, gradually got into other things. But I’d return to Diablo every once in a while.

The biggest boost came from an unexpected source. When my mom had a computer guy come to upgrade the PC, he noticed Diablo and mentioned a big expansion - Lord of Destruction - with a lot of new content. He also got me a Diablo 2 encyclopedia: an offline program with information on every item, monster, and location in the game, like modern wikis. That was an absolute game-changer. I couldn’t imagine there was so much left to discover. When I got a worm virus (remember those?) and it corrupted a bunch of files, this encyclopedia was what I lamented the most.

Eventually I got my first dial-up internet - around 2003, which was still quite late even for Uzbekistan. It was expensive, but I found pirate servers to play on, and just patching the game to the latest version was its own adventure. Downloading a few dozen megabytes on dial-up was a ginormous endeavor. But it opened the door to online play, and for the next two years I played Diablo on Russian pirate servers and loved it - while my offline life centered on Counter-Strike 1.6 with a team of childhood best friends.

Warcraft, meanwhile, became my other obsession. The second part was dated by then, but when I got my hands on Warcraft 3, I finally fell in love with RTS. I loved the campaign, played online using circulated license key lists, even entered a couple of tournaments - without any success. I tried hard to get my offline friends into it, looked for local communities, but had little luck until Dota came out and suddenly everyone was playing Warcraft 3.

That was around the time I was finishing high school and about to leave the country. My teenage years had been significantly shaped by Diablo 2 and Warcraft - and both of them were Blizzard.

Then I turned 18, left Uzbekistan for Russia, and the LAN era was over.

No more gaming with childhood friends, but a fast internet connection and access to licensed games for the first time. The first thing I bought was World of Warcraft. I really enjoyed it for a week or two, but reality kicked in quickly: new country, first year of university, too much going on. I left and never returned to Azeroth.

That began a few years without a Blizzard game that truly held me. World of Warcraft hadn’t scratched the itch. Diablo 2 felt played out. There was no sense returning to Warcraft 3 since Dota had become its own thing under Valve. I checked out StarCraft 2 and completed the campaign, but had no interest in multiplayer. Later on, Heroes of the Storm briefly caught my attention when it launched, but didn’t hold it long. I was playing plenty of other games - just nothing from Blizzard that stuck.

Things started to shift in 2012 with Diablo 3. My laptop at the time was too outdated to run it, and I was planning a work-and-travel program in the US that summer with the explicit goal of bringing back a proper gaming laptop. Spoiler: I did. A 5 kg Asus machine with a 17-inch screen that served me faithfully from 2012 to 2020. Quite the beast.

Diablo 3, though, never became a true successor to Diablo 2. I won’t rehash what everyone already knows - different setting, different atmosphere, different overall vibe. Not bad, but not what I had anticipated. In two or three waves of play over the years, I clocked maybe a hundred-something hours. Not nothing, but nowhere near my time in Diablo 2.

Subscribe now

In 2016, something came out of nowhere.

It was mid-May, a new Blizzard title was in open beta, and I downloaded it immediately. That was Overwatch. I was hooked before it even launched, bought it right away, and that became one of those moments you can point to. Overwatch would become my most played game for years - in the decade since, almost certainly my most played game ever. It became part of my personality; even the snapback hats I wear are Overwatch-branded.

Playing it with friends in hot-seat style at my place in 2016 and 2017 - one quick game each on my laptop, then switching seats, like it was 2003 and we were playing Heroes of Might and Magic 3 - is one of my favorite gaming memories.

But Overwatch was clearly the exception, not a sign that Blizzard had found its footing. By 2018 and 2019, their troubles were hard to ignore. The Taiwan controversy, harassment scandals, a management direction that felt badly off. Even the Diablo Immortal announcement - spectacularly botched as it was - I read as a problem of presentation rather than ideation. The idea of a Diablo on mobile was sound. I was genuinely excited for it. The uncomfortable reality, though, was that beyond Overwatch, Blizzard hadn’t produced anything truly great in years.

I stayed a believer. Always convinced they still had it in them.

Then in 2021, they proved it.

Blizzard announced the Diablo 2 remaster, and to say I was excited is to say nothing. Friends got me a pre-order for my birthday, and when it released I played the hell out of it. It was perfect - absolutely like the old game, with just enough quality-of-life polish and updated visuals. The best remaster anyone could ask for. I played through the release window and a couple of seasons, got active in the community, traded, had a great time.

Then Diablo Immortal finally launched. I had played the betas so I knew what I was getting: basically Diablo 3 on mobile, with its own story and some new features. And I played the hell out of that too. I was among the more active players on my server, joined a clan, took over leadership when the original leader left, helped form alliances, led the server for a stretch. We had a great international community, a few semi-famous streamers - a genuinely fun period, until I got married and my wife got pregnant and there was simply no time for it anymore.

Looking back, each Diablo game had pushed the social dimension a little further. Diablo 2 had trading and chat, but it was cumbersome - you relied on sites like D2JSP, made do with primitive tools. Diablo 3 tried to expand it, made a few bold decisions that backfired and were eventually retracted. By the time D2R came out, Discord communities and specialized trading sites had transformed that side of the game entirely. Immortal then introduced MMORPG-style faction activities: alliances, clan warfare, a rhythm of communal events. I didn’t stay long, but I was there for the best of it.

Overwatch 2 launched around that time and, as most people felt, underwhelmed. The PvE content I had been most excited about never materialized as promised. But Diablo 4 was due in June 2023, and I was ready.

The release of Diablo 4 was not the most significant event of the first week of June 2023 for me. My son was born that week. Still - I played, and it delivered. A strong launch with things to iron out, but not a single moment of disappointment. Diablo 4 became my second go-to game.

In the years since, both games have settled into a comfortable rotation. Overwatch found its footing once it dropped the PvE ambitions and committed to regular content. In early 2025 the game got a major overhaul: perks, Stadium mode, full third-person view - a package that finally felt like a genuine second chapter. In 2026, perhaps fittingly, they dropped the “2” from the name. The game had finally become distinct enough from the original to deserve that number, right as they decided it no longer needed it.

Meanwhile, Diablo 2 received its first major content update in a quarter of a century: the Warlock class, dropped in February 2026. I jumped straight in and played for a couple of months - a breath of fresh air in a game that’s been around longer than I’ve been gaming. And in April, Lord of Hatred wrapped up the Diablo 4 story arc that started in the base game. I enjoyed it, despite being fairly indifferent to Diablo lore in general. I’ve always been more in it for the gameplay.

All three games - Diablo 2, Diablo 4, Overwatch - are, in mid-2026, arguably in the best places they’ve ever been.

Diablo 4 has a persistent shared world. You’re always surrounded by other players, strangers from around the world wandering through the same landscapes. It’s strange to think back to that stripped pirated disc in Uzbekistan - a 10-year-old alone, in silence, discovering that green items existed. From there to running a clan on a mobile server, to just existing in a game alongside other people all the time.

One last thing: the first text I ever published online was a short fanfiction about the life of a Necromancer in Diablo 2. It’s gone now - the site has been taken down and relaunched too many times over the years. A small footnote, but a fitting one.

If you’ve read Part 1 and Part 2 of this series, you’ve seen the results. The suffix breakdowns, the color-coded localization philosophies, the moment where Brazilian Portuguese and Spanish turn out to be running entirely different strategies despite looking identical from the outside. Part 1 mentioned, in passing, that LLMs helped parse languages I don’t personally speak. What it didn’t mention was the scale of what that actually meant - or that the entire analysis was already sitting in a finished spreadsheet before that first article went live.

The series presented the findings in a logical order for the reader. That wasn’t the order in which the work happened.

Here’s how it actually went: the framework existed, the pipeline was running, and hundreds of entries across 18 languages were already classified before a single word of Part 1 was published. The three-part structure - puzzle, rubric, machine - was a deliberate editorial choice, not a chronicle of the research process. Part 3 is where that curtain comes down.

The decision to use AI as a classification engine was conceptual before it was technical. The game had already been translated by professional teams - that work was done, and done in ways no language model would have the cultural intuition to replicate. What I needed was something that could look at those finished translations and categorize them: map each localized string against a pre-defined rubric, identify which linguistic strategies were used, and do it across hundreds of entries without taking three to seven minutes per item. Not judging the translations as good or bad - just naming what kind of thing each one was. That reframing changes everything about how you build the tool.

The Stack (Nothing Fancy, On Purpose)

If you’re expecting a Python environment, a vector database, and a CI/CD pipeline, you’ll be disappointed.

The entire setup runs inside Google Sheets. That’s it. Google Sheets as the UI and database, Google Apps Script as the glue connecting the spreadsheet to the outside world, and the APIs from Anthropic and OpenAI as the logic engines doing the actual classification. If you have a billing account with either provider and access to Google Workspace, you have everything you need. Analyzing a full language tab costs literal pennies in tokens.

I want to be upfront about the trade-off: compared to dedicated, enterprise-grade LQA software, this setup is slower and more brittle. It was never designed to replace that. It was designed so that a solo researcher - or a single localization coordinator without an engineering team - could run a meaningful linguistic experiment without writing a line of Python.

Subscribe now

The Asset Drop

Rather than walk you through building this from scratch, I’m handing you the finished thing.

👉 Neologism Evaluation Matrix: Clair Obscur: Expedition 33 Analysis

(Open the link, click File > Make a copy, and you have a working version in your own Drive.)

When you open it, resist the urge to dive into the language tabs immediately. Start with the Settings tab.

One of the more common mistakes in AI-powered spreadsheet tools is baking the logic directly into the script. Change a parameter and you’re hunting through JavaScript for the right string to edit. Instead, the entire framework is decoupled: the prompt, the taxonomy definitions, and the provider selection all live in Settings, exposed and editable. The data validation dropdowns and the conditional color-coding across all 18 language tabs are physically linked to that sheet. If you want to rename “Pure Borrowing” to “Exact Transfer,” you change it once in Settings and the document updates everywhere.

The Settings tab has three things worth knowing:

The Provider Switch. A dropdown to route requests to either OpenAI or Anthropic. The script reads this dynamically, so switching models requires no code changes.

The Prompt. The full text of what gets sent to the model on every request. This is where the framework lives - not hidden in a function somewhere.

The Source of Truth. The exact definitions of every taxonomy category used across the matrix. This is what keeps the AI’s classifications consistent rather than creative.

The Prompt Engineering

Asking an LLM to grade a localization isn’t as simple as asking “is this a good translation?” You have to constrain it into behaving like a classification system, not a conversational partner.

The prompt is built in three functional blocks.

Context & Goal sets the ground rules before the model sees any data. Critically, it establishes that the French text is the semantic source of truth - even though English functioned as the pivot language during the actual localization effort. This is a non-trivial distinction. Without it, a model asked to evaluate the Russian Бульбарам might trace it back to the English “sack” and miss the sac à patates connection entirely. The prompt forces the model to do what a good localization researcher does: follow the meaning upstream.

The Typologies are the Molina & Hurtado Albir, Lincoln Fernandes, and Mangiron & O’Hagan category definitions, pasted in verbatim. Not summarized. Not paraphrased. If you give a model a loose description of “Structural Calque,” it applies the term loosely. If you give it the exact definition plus in-context examples, it applies it predictably enough to be useful across hundreds of rows.

JSON Enforcement is what makes the whole thing programmable. A model left to its own devices will produce a beautifully reasoned paragraph about why Cultam became Sectam in Spanish. That paragraph is useless to an Apps Script trying to populate a spreadsheet. Instead, the prompt ends with a hard instruction: return a JSON object with these exact keys, nothing else, no Markdown fences, no preamble.

For Monoco’s suffix (-aro), a translation that takes “Mouth,” renders it as the native “Rot,” and appends the suffix to create Rotaro, produces exactly this:

{"Analysis": "Translates the semantic root 'Mouth' to the native word 'Rot' and appends the character suffix '-aro'.", "Suffix": "Exact Copy", "Semantic": "Match", "Root": "Native", "Molina": "Structural Calque", "Lincoln": "Re-creation"}

The script grabs those values and drops them into the correct columns. No parsing ambiguity, no judgment calls.

The Automation Layer

You don’t need to touch the Apps Script to run this tool. But if you’re curious what’s happening when you click 🤖 AI Analysis > Process Active Tab, here’s what the code is actually doing.

Dynamic Column Mapping. The script doesn’t assume a fixed column structure. It reads the header row of whatever tab you’re on and builds a map of where each JSON key should land. Add a “Reviewer Notes” column, rename a header, rearrange the sheet entirely - the script adapts without breaking.

Safe Batching. Rate limits are real. The script processes rows with automatic retries and exponential backoff, and skips any row that already has an analysis populated. If you lose connection at row 340 of 600, you don’t restart from the top. You pick up exactly where you left off.

Fast Mode. Toggle this in Settings. With it on, the script runs in the background and populates the sheet when it finishes. With it off, you can watch it happen - cells filling in, color codes appearing, the matrix assembling itself one row at a time. This is completely unnecessary and I did it anyway, every single time.

What the Matrix Actually Revealed

Once the pipeline runs, the shift from Part 2 to Part 3 becomes almost visceral. In Part 2, patterns were things I could describe. In Part 3, they’re things you can see.

The color-coding immediately surfaces the macro-level philosophies. Languages that leaned on Foreignization - German, Polish, Indonesian - show up as solid, uniform blocks of “Exact Copy / Opaque / Pure Borrowing.” The matrix makes visible in a glance what took paragraphs to explain: these teams made a coherent, consistent choice and applied it across the entire item list without deviation.

The more interesting view is what happens at the exceptions. Within the Cognate Cousins group, Brazilian Portuguese and Spanish technically share the same localization strategy. The matrix tells a different story. Brazilian Portuguese shows a Pure Borrowing rate of 82.24%, with nearly a third of its terms categorized as opaque. Spanish sits at a 53.27% Semantic Match, having actively reconstructed native roots through Structural Calques. From the outside, both languages look like they’re playing the same game. Inside the matrix, they’re running entirely different philosophies - and neither team was wrong, which is the interesting part.

That’s what quantification does for you. It replaces “these feel similar” with “here’s exactly where they diverge.”

A Few Honest Disclaimers

Before you run the pipeline, a few things worth knowing:

Permissions. The first time you run the script, Google will ask for authorization. You have to grant it, or nothing executes. This is standard for any Apps Script making external API calls.

Security. If you found this code on the internet and you’re not sure you trust it - good instinct. Open Extensions > Apps Script, copy the code, and have your AI assistant of choice review it before you run anything. I’ve made it as readable as I could, but verify before you execute.

AI Variability. Depending on the model and reasoning settings, results will shift. The same item might be classified as a Structural Calque on one run and a Morphological Adaptation on the next. The matrix is a tool to accelerate human judgment, not replace it. Treat the output as a first-pass draft that a human reviewer confirms, not as ground truth.

API Costs. You need your own keys. A full language tab costs pennies. All 18 in one session costs a bit more, but still cheap.

(It is possible to get a daily allocation of free tokens from OpenAI, but you still need an active API account with them.)

Closing

I’m aware that I just spent three articles analyzing weapon name suffixes in a French Belle Époque RPG. There’s no dignified defense for this - only the honest acknowledgment that this is exactly the kind of rabbit hole I’d fall down again, and that the problems lurking underneath the surface-level absurdity are ones I run into on real commercial projects. How do you systematize creative judgment? How do you build a framework flexible enough to be useful without being so loose it becomes meaningless? Those questions don’t get smaller just because the dataset involves potato-themed mechs.

The framework in the spreadsheet is a starting point, not a locked box. The prompt in the Settings tab is yours to rewrite. Swap the taxonomies for your own theoretical frameworks, add columns, break the script and put it back together. If you build something interesting on top of it - or if you use it to analyze an entirely different game - I’d genuinely like to hear about it. The best place to reach me is here on Substack or on LinkedIn, and if you’d prefer to try the matrix without the API key setup, or need it adapted to a specific pipeline you’re already running, I’m happy to help with that too.

The Clair Obscur: Expedition 33 series ends here. The suffix puzzle, the academic rubric, and finally, the machine that grades them. If anything across these three pieces made you look at a localized item name differently, that’s more than enough.

Translating a made-up fantasy word usually devolves into a vibe check.

When a translator is handed a localized proper noun - a legendary sword, a fictional alien species, or a piece of magic technology - the evaluation process is notoriously subjective. Managers and editors often find themselves asking, “Does this sound right?” or “Does this fit the lore?”, while translators wonder, “How far can I go with creativity here?”

But pure subjectivity is the enemy of scale. When you are managing thousands of localized terms across 18 different languages, you cannot manage by vibes. You need a system to measure if a translator successfully adapted a made-up word without micromanaging their creative process.

As we explored in Part 1 of the series, Sandfall Interactive’s Clair Obscur: Expedition 33 presents the perfect stress-test for this exact problem. The developers enforced a rigid, UI-driven morphological rule: a weapon’s name must end in a specific suffix to indicate which character equips it (e.g., Gustave’s items end in -am, Lune’s in -im).

Translating these items meant solving a complex linguistic puzzle over a thousand times across the game’s supported languages. To analyze how the different teams solved it, I built the Neologism Evaluation Matrix. Based on both structural grammar analysis and the frameworks of renowned localization researchers like Carme Mangiron, Minako O’Hagan, Lucía Molina, Amparo Hurtado Albir, and Lincoln Fernandes, this 6-tier system is designed to categorize, evaluate, and scale the translation of fantasy terminology.

To be clear: no framework can completely eradicate the inherent subjectivity of human language. A borrowed English string might be perfectly understood by a German gamer, blurring the line of what is truly opaque. Similarly, the boundary between an established equivalent and a structural calque can sometimes depend on who you ask. However, by breaking these neologisms down into structured categories, we can replace gut feelings with a highly deterministic, defensible system.

(Note: If you want to see the raw data behind this system, I’ve made my complete 18-language dataset and interactive rubric open-source. I've dropped the link to the master Google Sheet at the very bottom of this article so you can copy the framework for your own projects).

Here is how to build a blueprint that quantifies the unadaptable.

Subscribe now

Section 1: The Architecture of a Fantasy Word

To evaluate a localized proper noun predictably, you must separate what the final word physically looks like from the linguistic strategy used to create it. To do this, the Neologism Evaluation Matrix is split into two halves.

Part A: The Physical Anatomy (Custom Metrics)

The first half of the matrix focuses purely on the physical anatomy of the localized string. In our raw data, this was tracked across three custom categories, which act as the ultimate product of the translation effort:

1. Constraint Adherence: Did the localization physically obey the developer’s rule?

• Exact Copy: The rule is perfectly preserved (e.g., Spanish turning Chevalam into Cabalam).

• Adapted: The rule is tweaked for target-language phonetics or scripts (e.g., Russian adapting the -am suffix into Cyrillic as -ам in Бульбарам / Bulbaram).

• Dropped: The rule is abandoned entirely (e.g., Simplified Chinese dropping the suffix to append standard characters for sword or blade).

2. Semantic Fidelity: Does the root word still carry the original in-game reference or lore hint?

• Match: The root perfectly translates the source meaning (e.g., translating the aquatic reference in Seashelum to the Spanish Caracola).

• Partial/Creative: The root relates to the item but deviates from the exact reference (e.g., Simplified Chinese translating Blodam poetically to 月光 / Yuèguāng, meaning moonlight, befitting a crescent-shaped axe dealing Light Damage).

  

• None (Opaque): The localized word is a phonetic copy with no decipherable meaning in the target language (e.g., Polish importing the English word Jarum directly).

3. Root Lexical Origin: What is the player’s perception of the word?

• Native: Built from recognizable vocabulary (e.g., Abismo in Latin American Spanish).

• Borrowed: An imported foreign string (e.g., Abysseram in German).

• Invented/Transliterated: A cluster of syllables built from phonetic conversion (e.g., Japanese mapping terms entirely to Katakana).

Part B: The Linguistic Strategy (Academic Taxonomies)

If those three custom metrics define the physical outcome, the second half of the Matrix defines the linguistic strategy. To categorize how and why translators made their choices, we map the data against three established academic taxonomies.

(Note: Depending on specific language pairs, there is natural overlap between these three academic frameworks. However, in the context of this specific analysis, this redundancy was intentionally maintained to cross-verify accuracy and capture different dimensions of the localization process. Teams can trim or combine these taxonomies based on their own project requirements).

1. The Micro-Linguistic Techniques (Molina & Hurtado Albir, Translation Techniques Revisited: A Dynamic and Functionalist Approach, 2002) Lucía Molina and Amparo Hurtado Albir developed a foundational taxonomy for translation techniques, focusing on the micro-level steps translators take to solve specific problems in a text. Their framework analyzes the mechanical construction of how the target word was built.

• Pure Borrowing: Taking the source word exactly as it is, without orthographic changes (e.g., English Abysseram imported directly into Indonesian as Abysseram).

• Naturalized Borrowing (Transliteration): Borrowing the word but adapting it to the target’s phonetic rules (e.g., Abysseram adapted to Korean Hangul as 아비세람 / A-bi-se-ram).

• Structural Calque: Translating the structural formula of the source word while swapping the lexical components for native ones (e.g., The Russian team mapping the formula Abysse + -eram to the native Хляб- / Khlyab- + -eram).

• Established Equivalent: Using a recognized translation for a known root rather than inventing a new one (e.g., translating the common RPG root Cadavre directly to Corpse or Trup).

• Discursive Creation: Establishing a temporary, highly contextual equivalence that is entirely unpredictable outside the specific context of the game. This covers massive creative leaps, such as the Russian team turning Sakaram (a reference to the French sac à patates or sack of potatoes) into Бульбарам / Bulbaram to preserve a lore-specific idiom about clumsiness.

2. Naming & Neologism Strategies (Lincoln Fernandes, Translation of Names in Children’s Fantasy Literature, 2006) Lincoln Fernandes focused specifically on the translation of names and proper nouns in fantasy literature. His taxonomy is perfect for gaming because it treats made-up words not just as vocabulary, but as entities with identity, lore, and structural weight.

• Copy: Reproducing the name exactly in the target text.

• Transcription: Adapting the name at the level of the alphabet.

• Morphological Adaptation: Altering the spelling of a name slightly so it looks right in the target grammar (e.g., The Ukrainian team adapting the weapon Choralim - referencing the Chorale enemy - by taking the native root Хорал / Khoral and attaching the Cyrillic suffix to make Хоралім / Khoralim).

• Rendition: A transparent translation where the semantic meaning is translated literally, often dropping the neologistic suffix entirely (e.g., The Vietnamese team taking a complex neologism and rendering it simply as a clear, descriptive native weapon class like Kiếm for sword).

• Re-creation: Inventing a completely new compound word in the target language to recreate the specific effect of the source text name (e.g., The Italian team taking Verso’s -so suffix and the root for confuse, but realizing Confuso is already an everyday Italian word. To maintain the artificial game feel, they invented Confusiso).

  

3. Game Localization & Transcreation (Mangiron & O’Hagan, Game Localization: Translating for the Global Digital Entertainment Industry, 2013) Carme Mangiron and Minako O’Hagan are pioneers in game localization studies. Their work shifted the academic focus away from strict equivalence and toward transcreation - the idea that game translators must prioritize the player’s emotional and gameplay experience over rigid literal accuracy. Their categories are best understood as guiding priorities.

• Transcreation (Playfulness & Feel): Departing from the source text completely to ensure the target audience has the same emotional experience (e.g., The Traditional Chinese team abandoning Western morphological formulas entirely to make items read like natural, poetic Wuxia equipment).

• Functional Priority (Mechanic over Semantics): A core concept where the UI or gameplay function dictates translation. The suffixes in Expedition 33 (-am, -aro) are functional; they tell the player who equips the item. A team prioritizing this will preserve the suffix at all costs, even if the root becomes opaque.

• Domestication: Adapting terminology so it feels inherently native to the target audience’s fantasy vocabulary (e.g., The Ukrainian team adapting the reference to the Bourgeon enemy. They took the translated root Бруньк’є / Brun’kie (bud/shoot), adapted it to Брунькй- / Brun’ky-, and appended the suffix to create Брунькйон / Brun’kyon, ensuring the fantasy world feels organically integrated for a local player).

• Foreignization (Exoticism): Intentionally keeping the terminology alien to maintain a specific atmosphere (e.g., The Polish and German teams leaving the French roots largely untouched, preserving an exotic, European Belle Époque flair at the expense of native semantic clarity).

Section 2: Abstracting the Formula

In the Expedition 33 data, the defining constraint was the character-specific suffix. Every weapon name had to obey that morphological rule to function correctly within the game’s UI. But not every project relies on suffixes to communicate mechanics.

To apply this matrix to other games, you simply swap out the first category. Constraint Adherence is not a fixed metric; it is a variable slot designed to hold whatever non-negotiable rule your specific project demands.

Consider a retro port tied to legacy UI architecture, like the classic 10-character limit for proper nouns in Generation 1 Pokémon games. In this case, your universal constraint is a strict character limit. Your evaluation tiers for adherence simply become: fits limit exactly, requires UI truncation, or breaks limit entirely.

Or perhaps you are managing a lore-heavy RPG where worldbuilding dictates weapon taxonomy. In The Elder Scrolls, a grouping derived from the race or species must be clearly indicated, such as Dwarven Weapons or Elven Weapons, just as Eevee evolutions in the Pokémon franchise are globally constrained by the -eon suffix rule. Constraint Adherence tracks whether that specific prefix or formula survives the translation process.

The constraint could just as easily be a procedural generation requirement, such as the dynamic loot systems in action RPGs like Diablo or World of Warcraft. In Romance languages, translated adjectives must often match the grammatical gender of the base weapon noun to plug cleanly into the engine’s text strings.

Defining this universal constraint before localization begins sets the baseline for the entire project. It gives translators a clear boundary to work within, and it gives QA an immediate, objective metric to check before they evaluate the cultural nuances of the translation. If the constraint fails, the string fails, regardless of how evocative or native the localized word sounds. Once the constraint is secured, we can look at how the translators actually handled the meaning behind the words.

Section 3: The Matrix in Practice (Data from the Trenches)

Once the universal constraint is established, the matrix allows us to map the actual linguistic strategies at play. In Expedition 33, the goal was to keep the character suffix intact while ideally preserving the semantic meaning of the root word.

By applying the matrix and mapping the data to the localization groups we established in Part 1, distinct operational philosophies emerge. Teams achieved the exact same goals using entirely different methods.

Group 1: The “As-Is” Adopters (German, Polish, Indonesian) This group opted for a strict non-interventionist strategy. The data shows they relied on Pure Borrowing for 100% of their terms, importing the original strings directly into the game. Consequently, they achieved a 100% Exact Copy rate for the mechanical suffixes. The constraint was technically perfectly preserved, but it came at a semantic cost. The matrix categorizes these items as highly opaque to the player: 72.90% for Indonesian, 70.09% for Polish, and 67.29% for German. However, it is ultimately up to the editor or linguist to evaluate just how understandable these borrowed strings might be for their specific audience. A German gamer might perfectly understand an English root, making the term functionally native despite its categorization. Still, on paper, the UI works flawlessly while the lore is muted in favor of a foreign atmosphere.

Group 2: The Phonetic Adapters (Japanese, Korean, Thai) When a target language uses a different writing system, an exact orthographic copy of a Latin string is impossible. The matrix accounts for this through the Adapted constraint tag. These teams utilized a 100% uniform approach of phonetic transliteration. Every source term was converted directly into Katakana, Hangul, or Thai script. The suffix constraint was entirely adapted - meaning the phonetic sound of the suffix survived the script change, successfully satisfying the functional UI requirement. However, the semantic meaning was sacrificed, resulting in inventories that are almost entirely opaque to local players unless they reverse-engineer the sound back to the English or French concept.

Group 3: The Structural Navigators (Simplified Chinese, Traditional Chinese, Vietnamese) Sometimes, the matrix reveals when a localization team decides a constraint actively harms the target player experience. Appending meaningless phonetic syllables to translated isolating languages creates visual and grammatical friction. Recognizing this, all three teams in this group exhibited a 100% Dropped rate for the suffix constraint. Instead of preserving the rigid morphological formula, they dismantled it. They appended natural, semantically descriptive native weapon classifiers to the translated roots. Vietnamese, for example, achieved a staggering 90.65% use of Native roots and completely eliminated opaque terms, hitting 100% semantic fidelity between direct matches and creative partials. They consciously sacrificed the exact game mechanic to maximize semantic understanding.

Group 4: The Cognate Cousins (Spanish, Latin American Spanish, Italian, Brazilian Portuguese) Thanks to shared Latin roots, these teams maintained the mechanical constraints flawlessly, all hovering at or near a 100% Exact Copy rate for the suffixes. However, the matrix exposes major philosophical differences within the group. Brazilian Portuguese leaned heavily into Pure Borrowing (82.24%), leading to nearly a third of their terms being opaque. Spanish, on the other hand, actively utilized Structural Calques (38.32%) to adapt native roots, successfully achieving a 53.27% Semantic Match while still perfectly honoring the UI constraint.

Group 5: The Deep Transcreators (Russian, Ukrainian, Czech, Turkish) This group successfully balanced high constraint adherence with high semantic fidelity by rebuilding vocabulary from the ground up. They heavily utilized Structural Calques (Russian at 64.49%, Ukrainian at 57.94%) to translate the root words into native equivalents. Because of this, they achieved strong semantic match rates (Russian 65.42%, Ukrainian 61.68%). To honor the developer’s UI rule, teams like Czech utilized a 100% Exact Copy strategy, while Ukrainian and Russian utilized the Adapted constraint to seamlessly graft the suffixes into Cyrillic phonetics. Turkish also required the Adapted constraint for 29.91% of its terms to respect native vowel harmony rules, proving that constraints can be flexible enough to respect target language grammar.

By tracking these decisions objectively across a matrix, localization managers no longer have to guess why one language feels different from another. The data exposes the exact pivot points where linguists chose to prioritize mechanics over meaning, or vice versa.

Section 4: Scaling the System (Governance & QA)

The real power of this matrix is not just grading translations after the fact; it is defining the rules of engagement before a single word is translated.

When managing a massive localization project, subjective instructions like “make it sound natural but keep the fantasy vibe” inevitably lead to inconsistent glossaries. Different vendors or freelance linguists will interpret that vibe differently, resulting in a fractured player experience.

Instead, localization managers can use the matrix to issue precise, verifiable directives. A glossary brief can explicitly state: “Prioritize Structural Calques to maintain Semantic Fidelity. If native roots break target language phonotactics, fall back to Pure Borrowing. The Universal Constraint must maintain an Exact Copy rate of 100%.” This removes the guesswork for linguists and establishes a clear benchmark for success.

Once the translations are delivered, checking thousands of morphological adaptations across 18 languages is historically a QA nightmare. This is where the matrix scales the operation.

Because the framework translates subjective creative choices into a rigid categorization matrix, it provides the exact logical architecture needed to feed an evaluation engine. It is entirely agnostic to who or what is doing the checking. You can hand this rubric to a dedicated team of human LQA testers or build an automated QA pipeline, whichever is available and applicable in your case. The system works because the tags are standardized and repeatable.

Conclusion

Ultimately, localization leadership is about balancing the art of the linguist with the operational reality of game development. There is a persistent fear that implementing rigid frameworks will stifle the creative soul of a translation, but the data proves the exact opposite.

A strong framework does not kill creativity; it gives that creativity a track to run on. By systematizing the structural constraints and defining the operational goals upfront, you free translators to focus their energy on the art of language, secure in the knowledge that their creative choices will align perfectly with the game’s underlying mechanics.

The Open-Source Rubric

If you are a localization manager or linguist looking to implement this framework, you don’t have to start from scratch. I have made the full 18-language dataset, along with the grading methodology, available to view and copy.

👉 Neologism Evaluation Matrix: Clair Obscur: Expedition 33 Analysis

(Open the link, click File > Make a copy, and you have an editable version in your own Drive.)

In Part 3 of the series we break down the AI automations that can help a solo researcher perform this analysis without a whole team of linguists.

When you pick up a weapon called Brulerum in Clair Obscur: Expedition 33, its purpose seems straightforward enough. Equip it, and your attacks deal fire damage. If you have a passing familiarity with French - or just a fondness for crème brûlée - you might even smile at the etymology: brûler means “to burn.”

But what about the “-um”?

It isn’t just a generic fantasy flourish. It is a strict, mechanical rule hidden in plain sight. As the game celebrates its one-year anniversary, players are still uncovering the philosophical and artistic secrets of Sandfall Interactive’s masterpiece. But for localization professionals, the game’s most fascinating puzzle isn’t hidden in the lore - it’s explicitly spelled out in the game’s source code.

In the item files for Expedition 33, the developers left a very clear directive for their translation teams:

The same rule applies to the rest of the cast: Maelle’s items end in -um, Verso’s in -so, Monoco’s in -aro, Gustave’s in -am, and Lune’s in -im. The challenge wasn’t discovering this rule; the challenge was executing it across 18 different localized languages.

As a localization professional and a dedicated fan of the game, the intersection of rigid system mechanics and fluid human language is my daily bread. I have no inside connection to Sandfall Interactive or Riotloc, the studio that handled the game’s localization. While localization diaries have revealed that Sandfall was incredibly supportive and collaborative with the translation teams, the exact resources, deadlines, or internal glossaries remain behind closed doors. Interestingly, despite the game's undeniable French soul, the localization team treated French as a translation target alongside the others. I am simply an enthusiast fluent in English, French, and Russian (and familiar with a few more) - relying on LLMs as a handy analytical tool to help parse the remaining languages - who wanted to see how the world solved the Brulerum problem.

Translating Expedition 33 was a masterclass in prioritizing constraints. Localization teams had to balance a rigid UI mechanic against the rich, layered meanings of the French source words. Here is how they did it.

The Anatomy of a Charactonym

To successfully localize an item in Clair Obscur, a translator had to navigate a hierarchy of escalating obligations. In literary theory, there is the concept of a charactonym - or what is known in Russian as говорящее имя (a “speaking name”) - a name that reveals the inner nature of the subject. Expedition 33 weaponizes this concept.

To grade the difficulty of the translation, we can break a name down into four levels of meaning:

Level 1: Functional Priority (The Suffix)
This is the developer’s golden rule. Gustave = -am, Maelle = -um, Sciel = -on. This isn’t just flavor text; it is User Interface design. It allows a player to instantly know who an item belongs to without opening a sub-menu. If the suffix is lost, immediate visual attribution is lost.

Level 2: The Gameplay Mechanic
The root word dictates utility. Brulerum must convey fire damage; Coldum must convey ice. Losing the root means losing combat readability.

Level 3: In-Game Reference (Lore Consistency)
Many names are directly tied to specific enemies or world objects. Take Volesterum, a weapon for Maelle tied to the flying Volester enemy. If a localizer translates the enemy name into their native language, but leaves the weapon as “Volesterum,” they introduce a logical inconsistency into the world.

Level 4: Flavor and Ambience
This involves contextual worldbuilding that doesn't affect mechanics but strengthens immersion. For example, Abysseram is a weapon found in an underwater level.

The Localization Spectrum: How 18 Languages Responded

When faced with a source word that demands a specific phonetic suffix but also contains a pun, a lore reference, and an elemental clue, something usually has to break. Looking across the 18 translated languages, we can group the localization strategies into four distinct approaches.

Group 1: The “As-Is” Adopters"

This group includes: German, Polish, and Indonesian.

The Strategy: Direct retention of the English names. These languages bypassed the puzzle entirely. By keeping the English terminology, they technically preserved the Level 1 suffix rule (since the English version preserved it), but at a severe cost to world cohesion.

While this might sometimes be a result of tight resources, it was often a deliberate stylistic choice. Some localization teams opted to leave original names untouched to preserve the game's Belle Époque flair. However, this stylistic preservation leads to tangible immersion breaks. For instance, consider Jarum, a weapon tied to the Jar-nevron enemy. The Polish translation team localized the enemy as Słój (Polish for jar). However, because they kept the English item names, the corresponding weapon remained Jarum. The Level 3 lore connection is completely severed. It is the most economical approach, but the least cohesive.

Group 2: The Phonetic Adapters

This group includes: Japanese, Korean, Thai.

The Strategy: Transliteration across writing systems.

Languages in this group mapped the items phonetically. Lune’s weapon Snowim simply became 스노임 (Seu-no-im) in Korean, スノイム (Su-no-i-mu) in Japanese, and สโนวิม (S̄-no-wim) in Thai.

Unlike Group 1, this isn’t necessarily a missed opportunity; it’s often a deliberate cultural choice. In Japanese and Korean gaming markets, transliterating Western fantasy terminology is a standard stylistic practice used to preserve a specific exotic or foreign “vibe.” The “charactonym” mechanic is lost, but the atmosphere is intentionally preserved.

Group 3: The Structural Navigators

This group includes: Simplified Chinese, Traditional Chinese, and Vietnamese.

The Strategy: Categorical substitution.

Isolating languages face a nearly impossible task with Expedition 33‘s suffixes. Because languages like Chinese and Vietnamese rely on characters and words representing distinct concepts, appending a meaningless phonetic syllable (like “-am”) to a translated noun is grammatically jarring and visually chaotic.

Instead of breaking their languages, these teams found a brilliant structural workaround: they replaced the phonetic suffixes with weapon classifications assigned to each character.

• Maelle: Attributed via 刺 (Thorn / Stinger / Stab) in Simplified Chinese, 劍 (Sword) in Traditional Chinese, and Kiếm (Sword) in Vietnamese.

• Sciel: Attributed via 镰 (Scythe/Sickle) in Simplified Chinese, 鐮 (Scythe / Sickle) in Traditional Chinese, and Liềm (Scythe) in Vietnamese.

They bypassed the literal developer rule to honor the functional developer intent: ensuring the player instantly knows whose weapon it is.

The Special Case: The French “Source”

Before looking at how other Latin-based languages adapted the text, we have to acknowledge the unique position of French itself. Even though English served as the primary master text for the global localization effort, the game’s cultural heart, puns, and aesthetic are deeply French. The French localization team had the paradoxical task of “translating” a French-feeling game back into French. They had to ensure the mechanical suffix rules aligned flawlessly with the native cultural puns that the developers originally conceptualized. Because of this, French acts as the true semantic source code for the rest of the world’s adaptations.

When considered in isolation from this unique trait, French displays the same properties as the following group with other Romance languages: from minuscule diacritic adjustments (Trebuchim - Trébuchim) to more substantial spelling changes (Coralim - Coraillim) to full root adaptations (Rangeson - Forestion - as in Sakapatate forestier) - everything is handled gracefully to preserve all levels of meaning.

Group 4: The Cognate Cousins

This group includes: Spanish, Latin American Spanish, Italian, Brazilian Portuguese.

The Strategy: Smooth linguistic adaptation.

This group successfully hit almost every mark on the hierarchy, but they had a distinct advantage. Because these are Romance languages, their vocabularies share identical or highly similar Latin roots with the original French.

They didn’t need to completely rebuild the words; they simply had to rely on linguistic cognates. In many cases, they could easily swap the French root for their own native equivalent and attach the required suffix without it feeling forced or structurally broken. In Spanish, Chevalam becomes Cabalam. In Brazilian Portuguese, Stalum becomes Estalum. In Latin American Spanish, Ballaro becomes Bailaro. Minimal changes lead to a fully native experience.

However, this linguistic proximity occasionally created its own unique hurdles. For example, when applying Verso’s “-so” suffix to the root for “confuse”, the resulting word - Confuso - is already an actual, everyday word in Italian. To maintain the artificial “game” feel and preserve the developer’s suffix rule, the Italian team had to add an extra syllable, cleverly turning it into Confusiso. Overall, it is a highly effective localization, but one born of historical and geographic convenience.

Group 5: The Deep Transcreators

This group includes: Russian, Ukrainian, Czech, and Turkish.

The Strategy: Rebuilding the vocabulary from the ground up.

This is where the heavy lifting happens. These languages utilized their flexible morphological systems to invent entirely new words that satisfy all levels of the hierarchy.

In Turkish, for example, Monoco’s weapon Grandaro (Grand + aro) was creatively transcreated into Heyhularo, deriving from the word Heyhula (meaning phantom or huge) and appending the -aro suffix.

In Ukrainian, Sciel’s weapon Moisson is a reference to the Moissonneuse enemy. The Ukrainian team localized the enemy as Жниця (Zhnytsia / female form of Reaper), and then flawlessly adapted the item root to match, creating Жницон (Zhnytsia + on).

In Czech, Lune’s Scaverim referencing the Scavenger enemy, was translated as Mrchorim which, in its turn, is based on the enemy’s translated name of Mrchožrout.

But transcreation is a double-edged sword, and the Russian translation offers a fascinating case study in semantic nuance.

Consider the weapon Jarum that we looked at in the Polish example earlier. To maintain the link to the enemy while keeping the suffix, the Russian team translated the item as Zhbanum (Жбанум). They took the localized enemy name Zhban (Жбан—an uncommon but perfectly understood word for a jug) and seamlessly attached Maelle’s -um suffix. It is a perfect structural calque.

However, one could argue that simply transliterating the French version’s Amphorum to Амфорум would have been the safer choice. A zhban can vary wildly in appearance and, arguably, doesn’t closely resemble the physical shape of the Jar-nevron. But more importantly, the word carries a secondary, vernacular meaning in Russian slang: the head. Specifically, it is used in the phrase дать в жбан (to punch someone in the head).

Because fighting the Jar-nevron is an optional encounter, possessing a weapon that inadvertently sounds like an invitation to “bash its head in” gives the item a slightly aggressive flair. It’s a subtle psychological shift that might subconsciously steer players toward a more violent playthrough—a choice that carries actual narrative consequences later in the game, potentially locking the player out of a trophy.

Still, the absolute crown jewel of transcreation in this game goes back to the Russian team with Gustave’s weapon, Sakaram. In French, this weapon is a reference to the mechanical Sakapatate enemies—literally a ‘sac à patates’ (sack of potatoes, meaning dressed badly or clumsy).

To quote the Russian editor, Maria Yunger: “how do you bottle that distinctly French whimsy, attach an ‘-am’ suffix to it, and uncork it for a foreign audience?” The Russian team’s solution was a masterstroke of multi-level translation. First, they localized the mechanical Sakapatate enemy as Bulbomech (Бульбомех), utilizing the culturally resonant word Bulba (Бульба - potato). Then, they took that exact root and forged the weapon Bulbaram (Бульбарам). By affixing Gustave’s -am suffix to the localized enemy name, they preserved the character attribution, the physical lore connection, and the hidden joke in one elegant swing.

The Invisible Art of Localization

It is important to acknowledge that item nomenclature is just a microscopic fraction of the colossal effort that went into translating this game. Clair Obscur: Expedition 33 presents localization challenges that would make seasoned veterans sweat - from an entire race communicating in a seemingly nonsensical, minion-like language, to incredibly subtle allusions to real historical figures and French philosophy that you have to possess the cultural literacy to even catch, let alone adapt for a global audience.

A puzzle like the item suffixes highlights exactly why localization remains a deeply human endeavor. While AI and LLMs have become fantastic tactical companions for modern translators - perfect for parsing vast data sets or brainstorming target-language puns - they are just tools. Judging the hierarchy of obligations requires strategic intuition. Knowing when to sacrifice a combat mechanic for a lore reference, or understanding when a target language’s grammar makes a structural calque impossible, is a choice about what matters most to the end-player’s experience.

When we equip an item, we rarely think about the linguistic acrobatics required to put it in our hands. But for those willing to look closely, the grammar of gameplay is a story all its own.