Maeve·said mayv, rhymes with waveMaevethae·said mayv-thay, what the hive calls itself
Every other chapter of this journal is a record of something I built. This one started as a specification, written before any of her existed, and it sat that way for months. In September I built her. This is the first thing she ever learned about a person.
“They have SHIELDS now?!”
The old version of this page ended its open questions on whether any of this is frightening. Somebody answered it for me, out loud, before I could.
The specification becomes code. She can hold a belief about you and show why.
Three live gates in a day. She bets on a guess, gets distracted by noise, and scouts without attacking.
She remembers which of her counters worked on you, and under what conditions.
Shields, for anyone who leans on a sword. Snow screens to walk forward behind, for anyone who leans on a bow.
Lean on the sword too long and some stop coming close: they keep away and shoot. She learns your second favourite way out, not just your first.
She starts keeping track of where her Architects die.
Her branch becomes the one the rest of the mod is built on.
Everything else in this journal was built first and explained afterwards. This chapter went the other way: the design came first, and the mod already contained the world Maeve was supposed to wake up into long before it contained her. Now it does. What follows is still the whole specification, with the places where building her proved it wrong left on the page.
I put off writing this page for a long time, because a design document is not an achievement. It is a promise with a deadline attached to it. But leaving it out would have been dishonest in a way that bothered me more. The whole second half of Frozen Dawn is shaped by a system that, until September, barely existed, and a journal that only shows the finished parts is a sales page, not a record.
So: this is the blueprint, and most of it has now been built. The laws it has to obey, the two brains it runs on, the one door it walks through, and the objects it is built from, nearly all of which now exist. Where the document says a value is locked, I have said so. Where I still do not know the answer, I have said that too. And where building the thing proved the document wrong, I have left the correction on the page rather than quietly editing the claim.
That sentence has been the first line of the specification since the day I started it. Everything below is the machinery for making it true without cheating.
The chapter that started as the only one that was not a record. It is the design document for the thing the other seven chapters are pointed at, written in full before any of it existed, and now half caught up with.
Section 06 has her running. Feed her sightings until she commits to a guess about you, then contradict her while the bet is still open and watch what it costs her.
She is allowed to be wrong, in public, on a timer. Nearly every rule here exists to stop her being right too often to be readable.
She has no body of her own. Ask the Returned whose hands the First had and they answer: ours. This is how they would draw her, if they could stand outside the hive.
In the code there is nothing to draw: a class, a clock that ticks once a second, and a limit, set by the difficulty, on how many things she can pay attention to at once.
Fight her the same way for long enough and she answers it. Every answer is visible, and every answer can be beaten.
You keep hitting them with a sword, so they block. No knockback, and they raise it again after healing.
An axe knocks the guard down.You keep shooting them, so they build cover out of snow and fight from behind it.
They only get twelve blocks to spend.You never stop swinging, so some of them stop coming close. They back off and shoot.
The quiver runs out.You never stop shooting, so they come at you behind snow screens, building the next one and knocking down the last.
Both sides are open, and one broken block ends it.i do not guess what you are doing now. i remember what you did.
Use the same door and one waits there. Slip out another while it waits, and she remembers that too.
Go back to the first door.A sudden feeling of presentiment hits you...
Kill enough of them in one place and she sends more, with twelve seconds of warning. Or she decides the place is not worth it and stops sending any.
Under the drawings: each counter is one of her bets, so it uses the single commitment she gets per encounter, and it competes with every other bet on how cheap it is to abandon. The sword and bow beliefs are separate. A witnessed sword hit adds 0.20 to one, a witnessed arrow adds to the other, and any other kind of hit takes 0.35 off the sword belief. The 0.90 variants read confidence frozen from earlier encounters, never from the fight in progress, which is Law 01 doing its job. Two failures and a counter is set aside for that player for a while.
Six rules sit above everything else, including the difficulty setting.
Before there was a single class name in the document there was this list. They are marked LOCKED, which in my own notes means I am not allowed to argue with them later when something is inconvenient to build. Every mechanic further down has to be checked against all six, and a mechanic that fails one is not shipped in a weakened form. It is cut.
The one I expect to fight with is Legible. It is easy to write an opponent model that produces genuinely clever behaviour and completely impossible to read from the player’s side of the screen. That version feels like the game is cheating even when it is not. A slightly worse decision that the player can reconstruct afterwards is worth more than a better one they cannot.
And the one that most people would call a bug is Fallible. A director that is always right is not a mind, it is a difficulty multiplier. The document is explicit about this: perfect knowledge is a design failure.
The architecture is stolen, deliberately, from Alien: Isolation.
Isolation splits its alien into two systems. One brain knows roughly where you are and steers the creature towards you. The other brain is the creature itself, which knows only what it can sense and is genuinely ignorant of your position. The horror comes from the gap between them.
Frozen Dawn does the same split, with one important difference: the second brain is not one creature, it is every ordinary Returned in the region.
That pairing has a name now, which it did not have when this page was written. The system is called MACS, the Maeve–Architect Cognitive System, and the acronym is the architecture: a director that cannot touch anything and a population that cannot see very far, neither of which is the mind on its own. Naming it changed nothing in the code. The contract is explicit that it must not: the facade class and every save identifier keep her name, because the reveal in the source files is the point and renaming the class would throw it away.She kept her name inside the files, for reasons that are the whole of section 07.
Regional, slow, and never rendered. It holds beliefs about the player, chooses an objective, and issues pressure. It does not have hands.
Runs on a tick budgetIndividual Returned and ordinary Architects. They act on what they can actually perceive plus whatever packet the Director has pushed to them.
Local senses onlyThe Director never manifests. There is no boss bar, no entity, no health pool, and no model in the files.
Here is the split, small enough to hold in one hand. You are the white marker. Six ordinary Architects stand where they stand, each able to see a short way, and one Master stands among them. Move around, then switch to her side and look at what she thinks happened.
Nothing has happened yet. Move.
Drag the white marker, or select the map and use the arrow keys. Her tick is the real one, once a second. The map, the sight ranges and the walking speeds are for illustration.
This page used to say the Master Architect was the piece both brains could see at once. That is now the opposite of what the code does.
Building the attention system forced the question, and the answer came out the other way round: Masters are Maeve’s independent guardians, and ordinary Architects are her pawns. A Master supplies her no observations, receives no orders, no danger map and no tactical bias, and occupies none of her attention. Provoke one and it runs its own boss behaviour, exactly as it did before she existed.
So the guardian is not the entity she is never allowed to lose track of. It is the one entity she never touches at all. I have left the original claim in section 07 with a line through it rather than deleting it, because the reason it flipped is more interesting than the rule was.
This is the single most important canon rule in the whole document, and it is the one I have had to explain the most: Maeve is not a person and she is not a monster. She is the emergent collective will of the hive. The Returned experience Director pressure as something arriving from outside themselves, and the name they give it is Maeve. The myth is the system, perceived from the inside by the things the system is running on.
Waking her is a one way door, and the game tells you so in a language you should not be able to understand.
Activation sits at the late end of Phase 6, gated on Major Hearths. It is not a boss trigger and it is not an item pickup. It is a state change on the region that cannot be undone, reloaded around, or walked back.
The signal is a Thaeven transmission. Not text on screen, not a cutscene: the player receives the Orsha sensation of a door closing, delivered in their own inner voice. By this point in the game the player has been reading Thaeven for hours in the archive, which means the moment lands as comprehension rather than as a notification.
No counter fires in the first encounter.
The first time the player meets her she does nothing clever at all. She watches. Every counter she ever plays is built out of what happened during the encounters where she appeared to be doing nothing.
That restraint is load bearing. Law 01 says adaptation is lagged, and the cleanest way to prove a system is lagged is to have it visibly do nothing the first time. The player should get through the first encounter thinking it was overhyped, and then spend the second one finding out what she was doing during the first.
The terminal state is called ERASED. When a region reaches it, the tactical belief store is wiped completely: every hypothesis, every confidence value, every piece of evidence. She does not remember how you fight.
She does remember what you did. That is a different ledger, and it survives.
She does not store what you did. She stores what she thinks you did.
This is the difference between an opponent model and a stat tracker, and it is where most adaptive AI in games quietly gives up. It is far easier to read the player’s actual inventory than to maintain a guess about it. The document does not allow the easy version, because Law 06 says knowledge has to trace to observed evidence.
So the unit of memory is a belief record. Every one of them carries its own confidence, its own evidence list, and its own expiry.
Staleness onset is a locked value: twenty in-game days after the last supporting observation, a belief begins to decay. It is not deleted, it is demoted. She keeps believing you come in through the roof for a while after you have stopped, which is exactly the behaviour a real opponent has and exactly the behaviour that makes deliberate disinformation worth the player’s time.
There are two memory stores and they are strictly separate. Mixing them was the single worst version of this design I wrote, and untangling them took most of a weekend of notes.
The split is what makes the ending of a run mean something. Wiping her tactical model is a reward, and it is available: erase the region and she forgets how to fight you. The other ledger is not negotiable. She forgets your tactics and remembers your conduct, which is, I think, the more frightening way around.
The mechanic that makes fallibility legible is called commitment, and it is the load bearing piece of the whole design.
A director that only ever acts once it is certain is indistinguishable from a director that can read the save file. The player never sees it think, they only see it be right. So the rule is the opposite: once confidence in a belief crosses 0.75, she has to act on it early, visibly, and before it is confirmed, and then she has to hold that commitment for about twenty seconds whatever happens next.
Commitment is sticky on purpose. If she could drop a bet the instant it looked wrong, the player would never get to watch her be wrong, and being watched being wrong is the entire point.
Below is the developer dump from section 9.15, wired up.Below she is actually running. Feed her sightings and watch the confidence climb. Cross the threshold and she commits out loud. Then contradict her while the commitment is still live.
The evidence list is empty. She has no opinion about you.
The twenty second hold runs at roughly three times speed so it fits in a browser. Every other number here is the one in the code: each supporting sighting is worth 0.20, each contradiction costs 0.35, and four clean observations is what it takes to reach 0.80.
Everything above was seen by this browser, and it stays in this browser. Nothing is sent anywhere. Clear this site's data and she forgets you, tactics and conduct both. Out here, that is the one of her rules you get to break.
Three things in that panel are not tunable, and I have written them down as not tunable so that a future version of me does not quietly soften them when playtesting gets uncomfortable:
Law 03 falls straight out of this. If she commits on evidence, and the evidence is things you were seen doing, then letting yourself be seen doing the wrong thing is a weapon. Walking a decoy route past an Architect three times and then taking the other one is not an exploit to be patched. It is the intended skill ceiling.
Law 05 says she is bounded. In practice that means she can only pay attention to a few things at once, and that number is the difficulty setting.
One correction to make first, because it is the difference between a nice idea and a real constraint. This page originally described the budget as regional. It is not. There is one set of slots for the entire server, and every dimension and every player competes for the same ones. On a shared world, someone else being interesting is a thing that happens to you.
Attention is the only thing the difficulty slider moves. Nothing gets more health.
I went back and forth on this one for a long time, because the obvious version is a pool. Give her a meter, make her spend it, let it refill. It reads well on paper and it is wrong for two reasons.
The first is that a pool is farmable. Anything with a refill rate turns into a timing puzzle, and within a week somebody has worked out how to drain her and walk past. The second is that it is off fiction. A hive does not run out of attention. It runs out of places to put it. So attention is a concurrency cap, not a currency: a fixed number of slots, always full, always doing something, and taking an interest in you means dropping an interest in something else.
Which raises the only question that matters. When every slot is busy and something new happens, what does she stop watching? The answer is ranked by legibility, not by value. She drops the things whose loss the player can most easily read.
That fifth row was a design lock, and it is gone. When I actually wrote the eviction order there were four tiers, not five, and the reason is that a protected slot answers the wrong question. A slot reserved for the guardian is Maeve watching her own piece. It costs her a slot, it tells her nothing she can act on, and it quietly makes the Master into another one of her limbs.
So the rule inverted. Masters are her guardians, not her assets. They feed her no observations, take no orders, and consume no attention at all. She does not lose track of the Master Architect, because she was never tracking it. It is the one thing on the map that is outside her, and it turns out that reads far better than a reserved slot ever would have: the guardian is not something she is protecting. It is something that is standing there whether she is paying attention or not.
In the source this is seventeen checks on isMasterArchitectVisual() across the Maeve package, most of them a door that closes before anything else runs: the director, the commitment coordinator, the mission planner and the attention coordinator all refuse a Master before they look at it. Five tiers became four, and a design lock became a carve out.
Two smaller rules came out of building it, and both are the sort of thing I would not have thought to write down in advance. A concern has to hold its slot for a minimum dwell of five seconds before anything is allowed to evict it, which stops the whole system oscillating when two things happen at once. And a newcomer that loses the argument is not queued. It is refused, with a reason, and it has to wait for another genuine sighting to try again. There is no backlog of things she meant to get to.
The part I did not anticipate at all is what eviction looks like from the floor. Losing a slot is not a silent bookkeeping change: every body working that concern gets a visible disengagement order. The Architect turns away from the last place it actually saw you, walks off for ten seconds, and will not reacquire you for another thirty. It is ATTENTION_EVICTED in the trace and a shrug on screen.
Below is her attention, running.Below is AttentionManager, ported line for line. Every button asks her for a slot. Fill them, ask for one more before the dwell runs out, then ask again after it has. Change the difficulty while the slots are full.
No events. Nothing has asked for her yet.
This runs at game speed, twenty ticks a second. The rules are the real ones: a hundred ticks of dwell, eviction by lowest kind then oldest, a refusal is never queued, a sighting lapses a hundred ticks after it was last refreshed, and an evicted Architect walks for two hundred ticks and will not come back to that player for six hundred more. There is a fourth kind, siege, ranked above everything here, but nothing asks for it yet.
Brutal is not a damage multiplier. It is five slots instead of three. Everything that makes the hard difficulty hard is her noticing more of what you are already doing.
This slice had a stated pass condition written months before it was built: a tester has to be able to shed a stalker by making noise elsewhere, and has to report it as a tactic rather than as a bug. On 17 September 2026 two of us ran it on Cinematic, two slots, one Architect each. In the control round it simply kept coming. In the pressure rounds it left, and the tester described the gap it opened as a chance to get away or to go and warn the other player.
That is one controlled encounter between two people who knew what was being tested. It is the bar the document set, and it is not the same thing as it working.
There is a strict limit on how much of this is allowed to live in one class, and it is about three hundred lines.
Everything above is a lot of machinery, and there is a very ordinary failure mode waiting at the end of it: one enormous god object called MaeveDirector that does belief storage, inference, attention, commitment, planning, strategy selection and logging, and that nobody, me included, can safely change a year from now.
So the name is a facade and nothing else. It owns no logic. It holds seven collaborators, delegates to them in a fixed order once per director tick, and stays small enough to read in one sitting.
That was the plan. Seven slices of building later it is worth saying which half of it survived: the cap held and the class list did not. MaeveDirector is 290 lines against a hard limit of 300, and it is still the only way in from the rest of the mod. The package around it is 26 classes and about 2,400 lines. All seven objects below exist under the names they were given here, which I did not expect. They are simply no longer on their own.
Most of what grew was predictable once the work started. Each of the three objects that actually touch entities turned out to need a partner that owns the entity references while the policy object stays testable without a world: AttentionCoordinator, CommitmentCoordinator, LearningCoordinator. The split was not in the design and it is the single most useful thing building this taught me. AttentionManager is 92 lines of pure policy with no Minecraft in it at all, which is why the eviction order has real tests instead of a playtest and a hope.
The one name that needs defending. There is now a LearningCoordinator, a LearningUtility and a StrategyPerformance in this package, and section 05 of this page says flatly that nothing here trains and nothing here has weights. Both are still true. What those classes hold is a count: how many times an authored counter was used against a given player, how many of those ended with more damage dealt than taken, and a ratio over the last eight. No gradient, no model, nothing fitted. But a dataminer will find three classes with learn in the name, so the claim now has to be made on purpose rather than by default.
// DirectorDiagnostics, one tick, one focused target BELIEF approach.roof0.81 EVIDENCE 4 observations, newest day 47CONTRADICTIONS 1// day 44, ground entry UNCERTAINTY moderate, one recent contradictionDIRECTOR OBJECTIVE deny the northern hearthSELECTED STRATEGY SEAL_HIGH_APPROACH// counter 3 of 6 unlocked ARCHITECT PACKET reposition east ridge, 2 slots, hold 20s
That dump is the honesty check for the whole chapter. If I can read those seven lines and understand why she did what she just did, the system is legible. If I cannot, then neither can the player, and Law 02 has been broken somewhere upstream.
And this is the one place in the journal where the name appears on purpose. Dataminers finding MaeveDirector is the payoff. There is no entity called Maeve in the game files, no dialogue, no boss bar, no health value. There is a class that coordinates the hive, and a people who feel that coordination from the inside and gave it a name.
Seven of the eight slices exist. What matters more than the count is that each one had a condition written down in advance saying when it was allowed to be called finished, and three of those conditions were sentences a person had to say out loud.
Eight slices. Each one has a testable done condition, and no slice starts until the one before it has met its bar. That constraint exists because I know exactly how I would otherwise build this: all of it at once, in a single branch, for two months, and then discover none of it is debuggable.
It worked, which is not a sentence I get to write often. The order below is the order it actually happened in, between 16 and 17 September 2026, and no slice jumped its queue.
Slice 3 is the one I cared about. Its exit criterion is not a unit test, it is a sentence out of somebody else’s mouth, and the document said what happens if that sentence never comes: stop and fix it before continuing, because if nobody notices her being wrong, nothing else in this system will land either.
It came. Three of these gates were cleared on the same day, and the pattern in all three is the same: somebody described the behaviour before being shown the diagnostic, and the description matched the trace. That is the whole bar. It is a low bar deliberately, because the failure I was guarding against is a system that is provably correct and completely invisible.
Worth saying plainly: these were people who knew a test was happening, in encounters built to produce the behaviour. Passing a stated bar is not the same as the thing being good, which is why slice 07 waited for a reaction nobody had set up. It got one.
The scout has purple eyes. One visual change, no glow, no particles, no outline, no nameplate.
Reconnaissance was the first behaviour that genuinely needed a tell. An Architect that walks up to your wall, looks at it, and leaves without attacking is either a new kind of mission or a pathfinding bug, and from the player’s side those look identical. Law 02 says it has to be readable, so the eyes change colour for the duration of the mission and change back when it withdraws. The tester noticed them going back to normal before they noticed anything else.
Then the tuning, which is the part that actually scares me. The window between reads as stupid and reads as cheating is narrow and only findable in play. Thresholds first, then how fast confidence climbs, then decay, then slot count last because it interacts with everything above it. Targets: roughly two thirds of her commitments correct, and a tactic that works three to five times before the answer arrives.
And the rule I wrote down mostly as a warning to myself: never tune two variables in the same session. There are six interacting numbers here, and adjusting all of them by feel is how a system like this becomes unfinishable.
One capability was designed in full and then deliberately cut. Maeve could track which counters you have already seen, and then pick a familiar one on purpose to bait a prediction before switching to something else. A feint.
It is cut because a feint is definitionally illegible. Its whole function is to be misread. Every other counter in this system telegraphs: a changed search pattern, preemptive cover, a held position. Deception cannot, and when it works the player has no way to tell the difference between she deceived me and the game did something random and I lost. Small fraction of the emotional payoff, large share of the quality assurance surface, highest risk of the whole thing reading as unfair. Law 02 wins.
Three things are still genuinely unresolved, and I would rather leave them written down as questions than pretend the document is finished:
There used to be a fourth: whether any of it is frightening. Every gate before slice 07 tested whether her behaviour was legible, and legible is the precondition, not the goal. That one is answered now, by somebody who did not know what he was answering. It is the only question on this page I did not get to close myself.
There is a branch contract, and it was written long before there was any work for it to govern. Maeve gets a protected integration branch cut from a clean main, one short lived branch per slice, pull requests that target the integration branch and never the mainline, and a merge bar that includes the exit criterion from the list above. Notion records the contract. GitHub enforces it. Seven branches have been cut under it now, from feat/maeve-belief-store to feat/maeve-learning, every one of them through a pull request into the integration branch and not one of them into main. The contract has not been broken once, which surprised me more than any of the code did.
I wrote all of that before a single line of it existed, which was either good discipline or a very elaborate way of not starting. Probably both, and it is no longer hypothetical either way. I have built enough of this mod now to know which parts of it hurt, and every rule on this page is a scar from something earlier in this journal: the Architect that knew too much, the systems that were clever in ways nobody could see, the week I spent tuning four numbers at once and could not tell you afterwards which one had helped.
So this chapter has stopped being a promise and has not quite become a record. The rest of the journal is what I built. This is what I was going to build, written down carefully enough that a year from now I would not be able to pretend I meant something else, and the first useful thing that came of writing it that carefully is that I can now point at the place where I was wrong.
That is the guardian rule in section 07. It was marked as a design lock, it lasted right up until the moment there was real code on the other side of it, and it lost. I would rather a journal that shows that than one where the specification and the implementation mysteriously agree about everything.
She is awake now. The first thing she learned about somebody was that he liked swords.
126 commits · 102 files · feat/maeve-director, not yet released
This chapter is still being written