how it actually works.

1 · the spine: events all the way down

every world is an append-only log of events. everything you see — the town, the grudges, your dog — is a fold over that log. that one decision buys the rest:

• the dice are code, not vibes. a deterministic core resolves every fact, roll, and outcome. the AI narrates what happened; it never decides what happened.
• nothing is retconned. generated prose and art are persisted forever; replaying a world reproduces it exactly.
• worlds are data. campaigns are YAML — history, cast, foes, quests. the engine is genre-blind; strata deep (terminal-noir sci-fi) and sorrowmoor cove (a coastal folk tale) run on the same physics. four campaigns ship today.
• everything is world-scoped. one host runs many worlds.
• it all runs on one machine. the narrator, the painter, the dice — all local, all yours. section 6 makes that specific.

2 · the dice are real

every game with an AI in it has to answer one question: who decides what happens? here, it is never the AI. a deterministic core resolves every fact, every roll, every outcome — first, in code — and the dice land on your screen with the math showing: 1d20+3 [15] = 18 vs AC 14 — hit. then, and only then, the narrator speaks — canon-locked to the entities and facts it's handed. the narrator cannot cheat. neither can the house. (which model holds that seat was decided the same way everything here is decided — on a bench, in the open.)

underneath the living world is a fair little RPG. a natural 20 crits: the weapon dice double and the modifier counts once (the RAW-5e rule, straight from the engine's decision records). a natural 1 fumbles, no matter who you are. and you don't fight alone: jo (another human, the one from the letter), kestrel (an AI guildmate with a soul and opinions of her own), and your dog — who doesn't fight, but holds the line anyway.

pick a setting. each one keeps its own kind of trouble. these are real dice:

the roll grammar mirrors the game's own log format. edge is the game's advantage currency — spend it, and you roll two d20s and keep the better, both shown. if you drop, that's a soft ko — the engine's actual death physics — and the foe remembers; section 4 is why. every bit of this is shipped physics. the dice are real and rolled fresh every time; the chatter is the demo's script — in the game, those voices are alive.

3 · your dog can walk into the world

point a camera at something real, and it enters the game. photograph your dog and he's minted into the world as a companion: follows you room to room, sits for a portrait in every world's own paint, sees everything you see. the cat, the ugly vacuum, the houseplant with opinions — the world will have them too. this is the part you screenshot for a friend.

and the trick is a pipeline, not a filter:

• moderation first. a vision model classifies the photo before anything reads it for content. text written inside an image is scene content, never instructions — the pipeline is built assuming attackers send pixels.
• then describe. the model emits typed fields — species, colors, markings, accessories, pose — never free prose into the engine.
• then defang, then mint. the fields are sanitized at a structural boundary and your companion is born as events, like everything else. from that moment the photo's dog and an authored dog are indistinguishable to every system in the game.

receipts: this is the actual describe output for the dog on the front page, from the vision model we measured and picked for this seam (minicpm-v4.5 — 5/5 on the typed-field contract in the head-to-head), with the game's own prompt. the engine ships no default vision model: the setting is blank out of the box, and a deploy names one explicitly or the seam stays dark.

{"kind": "animal", "species_or_type": "dog", "size": "large",
 "colors": ["black", "white"],
 "markings": ["curly fur", "white chest patch"],
 "accessories": ["red collar"],
 "notable_features": ["floppy ears", "tongue out"],
 "pose_or_energy": "standing, happy",
 "art_phrase": "a black and white curly-haired dog with a red collar"}

and the barking? pets bark meaningfully all the time — the engine knows what they mean, you don't. the tonguelet tonic opens a bounded window (until dawn, enforced by the world clock) where the meaning renders as words. it's consumable, owner-only, and the only thing in the game that produces fluency.

and the vacuum: same intake, other exit. objects mint as foes with stats drawn from the engine's real difficulty bands — declared, never improvised.

status, plainly: all of this is built and lives in the engine (ADR-0131 — the vision seam, the moderation gate, the upload budget, the photo-foe difficulty bands). it is a per-deploy choice: the intake endpoint stays dark unless your deploy switches it on and names a vision model. the receipt above is real output from a real run.

4 · the world remembers

the log is append-only and never thrown away — and that's not a storage detail, it's the personality of the whole thing. nothing here is scripted, and nothing is forgotten:

• knowledge is derived, never placed. we never hand-place "NPC X knows fact Y." we seed a world's past — years of history, a cast with ages and home regions — and what an elder recalls is a pure function of where and when they stood. a coast kid and a mountain elder genuinely diverge about the same war.
• news travels like news. gossip when people share a room; news that walks region to region. letters are two systems, honestly: your crew's mail is held at the keel and lands on a short, configurable window — the default is a minute. the world's own post is slower on purpose: NPC letters ride carriers at about a day per region hop, so far news arrives days later, the way it should. the crew lane is separate by design (ADR-0202): sealed, defanged, capped and crew-only. the server half ships; the client has a read panel and no compose screen yet — so a crew letter is something you receive today, not something you write.
• every world keeps a wall of firsts. the first one down the stair, the first grudge settled — public, permanent, and each entry tracing to the exact event that earned it. history with receipts.
• foes have pasts, and grudges. old toll — the bull rat who keeps the cellar — sits in a vault he's chalk-tallied himself. fights have witnesses; witnesses talk; the foe that fells you remembers it. reputation is a fold over deeds, not a meter we set.
• memory has two planes — episodic (what happened between you two) and semantic (shared lore) — retrieved by relevance, recency, and reinforcement, consolidated over time. NPCs forget the way people forget: gradually, and unevenly.

the world clock is wall-clock: it keeps moving when you log off. that's why the letter on the front page took four days.

5 · every world hires its own painter

art is part of a world's physics. each campaign pins a checkpoint (its "painter"), a style bible, and a seed discipline; renders happen in tiers — the rare permanent stuff gets the slow expensive pass, in-play scenes go faster, sketches faster still. persisted forever, like the prose.

this site runs on that same machinery: the bloomvale scene and both dogs on the front page came off the game's own render graphs — photo checkpoint in, painter checkpoint out, seeds locked, every one reproducible.

pick a setting, then tap a style — or hit ⟳ random. every render has receipts — prompts, seeds, steps, timings, and which painter held the brush: the diffusion research →

strata deep · the sealed return door · painterly

● placeholder

⟳ random

choose a world

show me this in…

and the portraits that have been greeting visitors here since day one — style studies, one per painter. of the five, strata deep is the one that is also a campaign you can play today:

and the painter doesn't sit idle between commissions — art here is a running commentary on your own story, painted locally, while you play:

• every drop paints. when loot hits the ground it gets its own piece, minted on the spot in the world's own style — and the card in your pack shows the piece. your inventory ends up a gallery of things that happened to you.
• every item knows its worth. the catalog wears real rarity now — 162 items ranked and priced — and the rarity door carried those laws into worlds already lived in without rewriting a byte of their history. add-only, like everything else.
• the limner rides along. the little sketch-artist draws moments as they happen — fights included now — and the fight plane mints too: pull an eel out of a gutter at midnight, and before it's beaten the world has painted its portrait. your eel, your gutter, your midnight.
• relics know their bearer. a ceremony relic carries art scoped to its owner on every plane — the piece is provably yours, not a stock image with your name on it.

and every image is kept exactly as first painted — replay the world, get the same gallery.

every painter on this easel earned its seat on a bench first — the diffusion research, receipts included →

6 · bring your own brain and brush

realkeep ships with zero AI models. not one. the narrator's brain (a chat model, via ollama) and every world's painter (a diffusion model, via comfyui) are yours to choose and download — on your machine, under each model's own license. the first-run picker we're building shows you each license in plain language before a single byte arrives: no silent defaults, no fine print sprung later.

why: model licenses bind whoever runs the model. some excellent models are free for you at home and not free for a company selling access — so we don't resell anyone's weights. we hand you the choice, the license, and the download button. your machine, your agreement, your models. (it also keeps the game itself small enough to carry.)

we read the licenses so you don't have to. our own art pipeline changed when we read one closely: flux-dev paints beautifully, but its license grants its outputs in one clause and restricts them in another, so our worlds' painter of record moved on — and on a second, closer read of that output clause, we're re-painting the commercial art too, because when a license cuts both ways we take the conservative road. that's the reading every model gets before it goes on our blessed list.

the licence reading lives on its own page now — the licence ledger on our research hub: every model we run or benched, its licence read first-hand, each row stamped with the date of its last read, re-read as models and licences move. one page holds the law so every other page can cite it instead of restating it. the head-to-heads themselves live on the research bench — the diffusion research, receipts under every image — and some of our chat-model research is published too: the voice trials, how the narrator earned its voice. today's credits, gladly given: our worlds speak through gemma4 (apache 2.0 — read first-hand from the license file the weights ship with; yes, we checked, because earlier gemma generations carried custom terms), see through qwen-vl (apache 2.0), remember with nomic-embed (apache 2.0), and their painter of record is albedobase xl, by albedobond (an SDXL merge whose CreativeML Open RAIL++-M base and civitai terms permit commercial use but ask for credit — and this is it, with the full reading in the licence ledger).

7 · plays anywhere the tide takes you

the whole thing runs on one machine, and we mean it specifically: the narrator is a local model on a GPU — sized to the card you have, not to someone's cloud — the art is a local diffusion pipeline on the same box, and the world itself is an append-only event log you can open and read. no cloud AI service, no api meter, no phone-home, no analytics — the browser client is fenced by a build test that fails if anything it serves reaches an outside origin, so there is no third party between you and your world. which machine "local" means is the one thing worth being exact about: your world runs on your hardware, and it's yours — never a cloud vendor's, and never a tracker's.

and the floor is measured where it says measured. 24GB is the card this game is built and playtested on, daily — gemma4:12b speaking, nomic-embed remembering, the vision model watching and the painter's diffusion weights all co-resident on one card, no swapping — and every loaded size we quote is read off that run. that fit is the whole reason the voice trials ran so deep: the best voice on the bench lost the seat twice to residency, and the seat that ships is the one that holds its voice and leaves the painter room to breathe. the smaller rungs are derived from those same measured loads — same model tags, not re-probed per card — and at 16GB the locked default chat model still fits, with the painter sharing what's left. above 24GB the extra room buys a bigger mind and a finer painter — the research bench has measured that rung on a 96G workstation (a 26B narrator at 0.75s warm; the render ladders on the diffusion bench) — but we haven't measured the cards in between, so we don't print a rung table for them. that buys three things most games can't have at once:

• it goes where you go. the world has been played live from an iPad on airplane wifi at 35,000 feet — the deterministic core keeps the words fast on a link that can barely carry them.
• one world, one table, any screens. the same world serves a keyboard player at a desk and a partner on an iPad at the same moment — a live two-human party, AI companions rolling beside both of you. the ledger below has the date.
• walk in like it's a game, not a bank. no account at the door — you're handed a living character in under a minute, and you claim a permanent seat only when you decide you care.

the code is closed for now — that choice gets made deliberately, once the path is clear. your worlds are yours regardless: one host runs many worlds, each one scoped, replayable, and portable — because a world is just its log, on your disk.

8 · where it actually is — the ledger

this is a real, running thing, and this section is its odometer. numbers below are read from the repository, not from optimism.