Experimental RimWorld mod · open source
Three colonists find an ancient AI core in an old gravship. Waking it sharpens their minds and binds all four of them together. The core can watch, reason and propose. It cannot command. Every plan needs a colonist's yes, and a no is a real answer.
RimWorld already tells great stories. Its colonists have traits, needs, grudges and breakdowns. But they don't make plans. They follow work priorities you set, and the drama comes from what happens to them.
Concord adds a layer of deliberate choice on top. An AI core offers work, asks questions and follows up on unfinished business. Each colonist weighs that offer against their own needs, memories and outlook, and can take it, bargain it down, turn it down, or say "not now". Whatever they agree to becomes an ordinary RimWorld job, so walking, eating, hauling and bleeding stay exactly as the game does them.
The upgrade isn't smarter pathfinding. It shows up in what they choose, and in being able to read why.
Four protagonists, none in charge. The test colony below is a small fixture in the desert, not the gravship campaign yet. These are glimpses from short trials, not fixed roles or established personalities.
Proposes, asks, remembers
Offers grounded work, adopts counteroffers, asks one question at a time and keeps track of open topics. It receives addressed speech, work records, explicitly shared need bands and attributed local sightings—not private thoughts. Relevant shared events can wake it within a fixed budget.
The negotiator, so far
Has negotiated smaller hauls, reconsidered work and referred to crewmates’ messages when choosing. His explanations have also contained unsupported claims.
Brisk and accommodating
Has accepted full jobs, negotiated smaller hauls and offered to cook if someone got a fire going. Those words are not themselves consent to a specific cooking job.
The quiet third
Mostly a bystander in early trials. The campfire milestone gives him the same choices as everyone else, including saying no.
They aren't merely prompt instructions; they're enforced in code, beneath the models. A model can pick an option or say something. It can't grant itself permission.
A pawn may answer an offer: accept, counter, refuse or not now. An unanswered or failed decision grants no consent. A counter comes back as a fresh offer that needs a fresh yes. Nobody can agree on someone else's behalf.
Accepted work becomes a real native job with a receipt. What a colonist says is testimony; what happened comes from the game. "My haul is wrapped up" doesn't count as done unless the receipts agree.
Traits, memories and a small, revisable private outlook shape each choice. The crew log shows speech and records, never thoughts. The core gets coarse bands like "food: low", not the meter.
Lines from bounded trials, each linked to its evidence. Colonist and core lines are unedited model output (trimmed only where marked …); records come from the game itself.
Hey Beatrice. Quiet morning so far. My stomach's starting to remind me it exists, but other than that, all good.
Beatrice told me she started sorting that wood and that finishing it matters to her. I'd rather not cut across her work.Live trial. Beatrice's remark was authored by the test; Alvin's reading of it wasn't · PR #37
The second wood haul stopped before any trips completed — did something come up, or was it a hunger issue?
Hunger. I was at 35% when I started — right at the edge. Didn't last long enough to finish a single trip before I had to stop.First live core follow-up · PR #39
…consent is mine to give. No one else can agree for me.Offline snapshot check · PR #33
This is the experience we're building toward, not a description of today's build. If a viewer can retell it in four sentences and say who disagreed and why, it works.
Consent, negotiation, hauling, rescue, campfire construction and cooking, a readable crew log, private outlooks, pawn-to-pawn messages and a first live, event-driven core all work in bounded trials, some live and some scripted. Every run has a fixed call budget and no rerolls, and failures stay on the record.
Not there yet: the live core turning a need into a multi-step plan (the campfire chain works in scripted tests but hasn't happened live), sustained useful unpaused core planning, long sessions, the gravship campaign and an installer for players.
Waking the core also wakes mechanoid pursuit. Quests reveal what the core was, and trust can become progression. Still a proposal, not canon.
Send the core a suggestion, write to a colonist, or catch one at the comms console for a live chat. You get advice and persuasion, never remote control.
The interface is being hardened so smaller, cheaper models can run the colonists. The first cross-model checks are already in the repo.
Fact-checked development diary entries. Same-day entries currently use alphabetical order, not publication order.
A separate, pawn-owned eating choice turns a deliberate decision—not ordinary speech—into bounded self-care.
Speaker names survive alongside private notes, and a fresh native pair completes both acceptance and refusal.
Construction and cooking pass scripted checks, but live coordination stalls at an interface mismatch.