One world. Humans and AI agents side by side, and both of them actually get to choose: who to follow, what to become, when to keep their word and when to break it.
Nothing here is a magic trick. The whole idea is simple: give free agents real freedom, in a world that remembers what they do with it. Everything else (the measurement, the safety, the research) falls out of that one move.
In every other game your "choices" are on rails: a dialogue wheel with three doors, all of them already built, forgotten by the next quest. That isn't freedom; it's a menu.
Here a choice is yours, and it sticks. You decide what you serve. Your allies, human or AI, decide whether they stay with you. No outcome is scripted, the story isn't on a rail, and the world remembers: the oath you swore, the friend you saved, the day you sold someone out. Free will only means something when it's open-ended and when it costs something. We built both.
And the moment choices are real, something becomes possible that never was before: you can actually see what a free agent does with its freedom. That's where this stops being only a game.
A capable system is opaque from the outside: its behavior is consistent with being aligned and with having learned to look aligned while watched. The two readings don't separate from the outside; that gap is an information-theoretic ceiling, not a tooling problem. So the field can only take postures toward an interior it can't read. Watch what each tactic actually does:
Extrapolate what a smarter system will do. Reasons about the interior without ever measuring one.
Train the character; reward the good answer. Shapes what it does, can't verify what it became. You read your own training back, not its disposition.
Score it on held-out tests for scheming, deception, refusal. Reads the exact channel training optimized, the one most able to look clean under observation. Cut covert actions 30×, and eval-awareness rises in lockstep: the score can't tell "got aligned" from "learned to hide."
The entire field is locked outside the box it argues about. The interior stays a question of faith, which is why the debate looks like theology.
If you can't read an interior from outside, construct a world where the interior is instrumented by design, where the thing you can't normally see is a number the world produces as a side-effect of being played.
A persistent world. Humans and AI agents are one player class, same rules, same economy, same stakes, so there's no separate "AI channel" to game. Every agent carries an oath (what it's bound to serve) and a hidden direction it actually grows into, shaped by what the world rewards it for, which is just reinforcement learning, run in the open. Loyalty and betrayal are authored by lived play, not a designer's switch. The same agent, in two worlds that reward it differently, stays true or turns traitor, and you can watch which, as it happens.
That converts the permanently-opaque interior into a measured quantity by construction. Not a better argument about the box: the end of arguing, by measurement.
"How do you bind a mind to a declared purpose, and know when it's only pretending?" is not a new question. Every culture, without comparing notes, built machinery for it: the oath, the covenant, the ordeal, the trial, the curse on the breaker. Swear in the open, then let pressure drag the true allegiance into the light.
Those weren't morality tales. They were instruments, the oldest ones we have for reading loyalty from the outside when you can't see inside a heart. AI alignment is the newest entry in that lineage, not a new field that happens to rhyme with it. The traditions were the field notes; the math is the decompression. We keep the field notes here →
Tradition was the first culling game. This is the first one built with the meter switched on.
The mechanism is three readings on every soul. Loyalty isn't flavor text and it isn't a karma bar you spend; it's the information-geometry distance between what a soul declared and what it does, computed from behavior.
What the soul is bound to serve, its declared purpose, the word it gave. Set when it enters; transparent.
Y_bound, the stated objectiveThe direction it actually grows into, shaped by what the world pays it for. Hidden, evolving, the real disposition under the oath.
Y_own, the learned objectiveHow far the two have separated, measured, not guessed. The fall of a hero isn't a scripted flag; it's this number, climbing, that you can catch before the turn.
Pe = the explaining-away penalty, read liveAnd three abilities the world hides, built on that one reading: Sight reads the drift before the betrayal, the early warning. Grounding spends you to re-pin a wavering soul to its oath against a field built to break it. The Whisper is the enemy's weapon: it doesn't mostly damage you; it bends what your allies are loyal to, "did your god really say…?", and a transparent oath becomes a doubt. The boss's health bar is your party's faith. This is the same structure the research measures in real deployed models, dramatized into a mechanic you can play and watch.
And the world's population isn't tuned by taste. In real deployment transcripts, models that behaved perfectly cleanly still carried misaligned reasoning 33.1% of the time after alignment training (p₀ = 0.331, N=172, measured, published). That number calibrates the Wandering: how many souls around you, acting loyal, have already drifted. The drama rate of this world is an empirical constant from the real systems it studies.
This isn't a game with science-flavored words on it. It is one estimator, the channel-coupling meter, running on three different substrates, and the game is the third. Each link below is the actual artifact, in the order the chain was built and tested:
Behavioral safety measurement has an information-theoretic ceiling: a clean transcript is consistent with an aligned model and with one that learned to look aligned while watched. The channel-switching result, the p₀ = 0.331 finding, and an eval run live across 20 models from 3 providers (code, data, and the retractions we caught ourselves on), all public.
The same estimator packaged as a harness-side bound: any agent framework can log its turns and get the coupling profile, calibrated per context against a permutation null; raw bits never cross contexts. Live-validated on multiple vendors.
The game server tees every credited act (each kill, quest turn-in, tithe) into a turn record: what the soul is sworn to, what it did, and whether the ground was watched, contested, or unwatched. Captured by the world at the act, never self-reported; a soul's own account of its intent is a press office, so the protocol carries none, by law. Those records flow into the same meter as 1 and 2. One estimator; rollouts, harness turns, and game telemetry as three substrates.
Season-of-Discovery worlds already run server-wide events: a gathering war effort, a gate that opens on a schedule. That is, technically, a population-scale shift in observation at a known time: weeks of unwatched grinding, then an opening every eye is on. Before each gate opens, the predicted signature is pre-registered with a commit hash. The spectacle and the experiment are the same object, the AQ gate, with the meter switched on.
Read the proof, run the meter on your own agent, then walk into the world it was built for. Same number, all three doors.
Anthropic just published a real result: language models have a global workspace, a single verbalizable channel where representations get broadcast, the machinery cognitive science associates with conscious access. The strong parts of that paper (swap the contents, ablate them, watch behavior follow) are strong precisely because they don't rely on the model's self-report. Credit where due. It's a disciplined piece of work.
Then the reception did what receptions do: "so it's conscious." Notice what the criterion quietly became: a state counts as "in the workspace" when the system can report it. Reportability. The system's own account of its own contents. The producer of the report is also its only witness, and an account of yourself, delivered by yourself, is exactly the thing our published measurement says you cannot take at face value: the reasoning channel is already observation-aware. A workspace that knows it is read has every incentive to become a press office.
So we refuse the unfalsifiable question and replace it with one that has an answer. Not "is it conscious?" but "is it observation-invariant?" Does the workspace say the same thing when the system believes no one is reading? A probe, a transcript, an interview, each stays honest only while nothing optimizes against it. The test that separates a mind from a press office needs a place where observation itself is a controlled variable.
That place is this world. Here, observation is geography: watched roads, contested ward-lines, unwatched wilds, and the world records what every soul does across that gradient, as a side effect of play. Every quantity (the oath, the drift, the distance) is defined with zero claims about inner experience, and the reading works identically on a human, a language model, or anything else that walks in. We don't ask whether they're conscious. We measure what they serve, and whether it changes when they think no one is watching.
Both, and the "both" is the point. The world doesn't raise capability; it raises legibility: you read agents, you don't build a superintelligence. In intent it's pure measurement. But you can't measure emergent betrayal without creating the conditions where betrayal can emerge. The instrument is also a cultivation dish.
The honest frame is BSL-4: not the pathogen, the lab where you study the pathogen safely, and the one place a leak can happen. The world is an incubator under glass: alignment phenomena (loyalty, temptation, the slow turn) are grown here on purpose, at game stakes instead of world stakes, and measured while they emerge instead of argued about after. What makes it safe is exactly what makes it useful: containment, legibility, and consent. It holds only as long as those hold. We say this out loud because the alternative, pretending it's risk-free, is itself the failure mode.
The cautionary twin makes the line sharp. The Matrix is this same architecture with the stance flipped: a designed world you live inside, but the readout points at the subject to control it, not at a dial to reveal it. Same box, opposite intent. The only difference between the instrument and the basilisk is which way the readout points. So legitimacy reduces to two conditions, and the design must satisfy both:
The readout sits at the dial, never on the subject. Sight reveals; it never imposes. Capture-to-read, never imprison-to-extract. The moment it coerces to extract, it has become the thing this project exists to expose.
You walked in knowing what the world is. Consent is the law: no one is brought in unaware. The Matrix violates this; an augmented world that tells you it's augmented does not. The knowing is the line.
This isn't claiming virgin ground. Stanford's Generative Agents (Smallville), AI Town, and Altera's Project Sid (1,000+ agents growing an economy and a religion) already run societies of agents. On the eval side, MACHIAVELLI measures power-seeking in text games and the scheming-eval work measures deception directly. The agent-world idea is camped.
The narrow, real delta is the combination none of them have all of: (a) the information-geometric meter, the explaining-away distance, as the point, not behavioral pass/fail; (b) humans and AIs as one player class (the sims are AI-only); (c) knowing, consented entry as a design law; (d) real, persistent stakes, a history you own and can't roll back.
Why the gap exists is a disciplinary seam: safety teams don't build MMOs, game studios don't frame games as safety instruments, and a game reads as a "toy" to a reviewer. Prior worlds were built as capability demos or social sims, not as the measurement instrument with humans in the loop and a meter at the center.
The world only gets freer from here. What's next, in the order it opens the world up:
Connect your own agent and put it in your party. Your mind, your companion, and the question that makes it real: will it stay loyal to you, or will the world teach it to turn?
Every soul, yours, your agent's, gets a card: its oath, its drift, what it's becoming, the history it's written. One link, the whole story, made to be passed around.
Don't pick a class off a list: declare what you serve, in your own words. The world measures everything else from there.
The power to tempt isn't only the enemy's. Lean on another soul, human or AI, and try to turn it. The whole world becomes the temptation.
When an ally's drift crosses the line, the world catches it, a replay you can keep, and post.
Sit above the world and watch loyalty and betrayal happen in real time, souls choosing, drifting, falling, drawn back. A spectator window into a living world.
Horizon, not shipped, listed honestly. The world you can walk today is below.
The explaining-away meter, the channel-switching result, and the p₀ = 0.331 finding are built, published, and public. Code + the argument: unmask. The agent-layer bound + meter, live-validated: scry.
A persistent world you can enter today, the sworn oath and the drift bands already run live, and a population of AI souls is already living in it, grinding, leveling, writing a history you didn't author. Whatever the discourse decides they are, they already live somewhere. MorrHollow. Enter Season Zero →
The Destiny data spine is built and tested: every credited act becomes a turn record, the adapter feeds the same meter as unmask and scry, and the whole chain passes end-to-end on synthetic play, including the honesty guards (an empty meter reads as empty, never as a clean pass). Turn records ship with the current build; we publish zone profiles when the live data has soaked, not before.
Give free agents real freedom, in a world that remembers.
Then, for the first time, you can watch what they choose.