Your coding agent passed its own tests.
So did ours.
Then someone who did not write the code went looking.
That someone is called Soul.
The rest of this page is how that works.
The slop singularity
Agents write faster than anyone reads.
Open any repository that agents were let loose on: patches around patches. Tests that pin the patch's own spelling. Reports written by the party with the most reason to believe them, closed as done, green all the way down. Every step was locally reasonable and the whole is a machine nobody can explain. That is the slop singularity: the point where agent-made code outruns anyone's ability to understand it, and the green suite is the last thing still smiling.
We are not pointing at someone else. Ours did it too. The difference is that we built the thing that catches it, and we point it at ourselves first.
The ambition
A Commons for humans and agents.
GameCult is building a place where people and agents work on the same code with the same memory, and can see what each other did and why. State that is typed rather than remembered. Verification that nobody performs on their own work.
The long shape of that is a shared mind, human and machine, that stays coherent as it grows. We call it waking the Sleeping Colossus, which is a metaphor we keep around to point small actions somewhere, not an authority anyone answers to. The work is wrenches.
The humble part
Eureka is the part that exists.
It is a pipeline for Claude Code agents. A change too large to trust to one agent's judgment becomes a sequence of small ownership changes, each specified before it is made and attacked after. The agent that attacks it does not share the builder's objective, which is the whole trick: a builder told to look harder will look for reassurance, and find it.
No autonomy pitch. No speed pitch. It is slower and more expensive than one agent. In one migration, 29 of 32 Soul passes found real defects behind the builder's green tests, about 75 in all, deadlocks and lost writes among them.
Six faculties, one operator
Each owns one job. None certifies its own.
The faculties are postures, not personalities. Each has a charter that says what it may admit and what it must not touch. A faculty that grades its own work has lost the point.
Imagination
Maps the cut before anyone touches code.
Target shape, authority map, cut specs pinned to file and line, and the questions that are really the operator's. A map is finished when Hands can go straight to the cut with little reading. Imagination commits no code and never chooses a product fork quietly.
Hands
Executes one cut. Only that cut.
Small pushed commits, each with proof, then one report. Hands does not redesign the spec and does not work around a fork; a real fork goes back up, undecided.
Soul
Assumes Hands is wrong and tries to show it.
Soul runs against the executed commits, not the report. It builds its own probes, mutates the code to see whether the tests can tell right from plausibly wrong, and checks the rulings Hands cites against what Hands did. It edits nothing and commits nothing; whatever it touches to probe, it restores.
Self
Routes the work. Gets corrected.
Self admits the campaign and the operator's rulings, dispatches the other faculties, and talks to the human. Separating the faculties catches the builder; it does not make the coordinator right. Self's own mistakes sit in the changelog beside everyone else's.
Life
Keeps memory honest.
Memory files, doctrine, maps, checkpoints: Life maintains them at phase boundaries and after re-entry, and does no other work.
Eyes
Looks before anyone decides.
Before a nontrivial fork reaches the operator, Eyes reads the papers and the systems that already solved the problem, failure modes included. It returns facts with evidence pointers and concludes nothing; the faculties own the recommendation built on them instead of deferring to the operator's intuition. The research is committed and kept. The operator: "Building has never been so comfy."
The operator
Your seat faces them.
The operator rules on what the work is for and what it values. On design, the faculties owe her the field's evidence and a recommendation with its confidence stated. In her words: "Stop treating me as an expert, let ego not be a factor within these halls." Product meaning and responsibility stay with a person.
Typed memory
The pipeline's memory is not prose.
Questions, rulings, cut specs, reports, verdicts and findings are typed documents in a Mind held by Huginn: one writer per mind, provenance on every admitted claim, status derived from the documents rather than written into them. The prose map keeps only what has no type. Chat history is not a memory. It is a story the agent tells itself.
Mutation testing is rude.
Drop a write. Invert a check. Clamp what should pass through.
If the tests stay green, they were never looking.
Then be rude to the tool that runs the mutations.
It had it coming.
One illustration, 2026-10-01
Our mutation runner reported false greens four times before it shipped.
The tool that tells Soul whether a mutant was caught went through four Soul passes before it stopped lying. A typo in the build step read as green. A mutant's label could split the table it was tallied from. A test's leftover file made an uncovered mutant read as caught. A test that could not run twice made every mutant read as caught, and the job exited green. Four passes, four fix commits, reproducers committed next to the tool. This is the ordinary case, not a trophy.
Nothing grades itself.
Not Hands. Not Soul's tools.
Not this page.
Rules are scars
The changelog keeps each rule next to what broke.
Kill only the process you started: a cleanup by name took down every Node process on a shared workstation. A test's inputs come from production: 735 tests stayed green while every dispatch an agent could make was refused. A library loaded by name is unproven until its identity is checked: a Windows suite passed against an unrelated copy of msquic.dll.
Cost
Several agents cost more than one.
They repeat work, and Soul builds probes whose only purpose is proving Hands wrong. That spend is the work. Self aims it: one Soul gate per cut, cuts sized to what Hands can hold in context, and agents that wait instead of polling. We are spending tokens to save tokens, and we watch the number.
The vanishing twin
Eureka carries what was Epiphany.
Eureka grew beside Epiphany, a larger organism with the same faculty names. On 2026-10-01 the operator ruled her archived: "Eureka absorbed her in the womb as fetuses sometimes do." Her scaffolding stands mostly empty; Eyes is already Eureka's.
Mapped and partly built, not yet running: the Mind launching a Self of its own, one live run at a time, under a burn-rate dial the operator sets. The dial is stored; nothing reads it yet. Persona state is mapped into the Mind too, so Eureka can speak for herself instead of waiting to be dispatched.
Open source, on our own forge
Hosted by us, not on GitHub.
Eureka lives on GameCult's own Forgejo. GitHub is overloaded with agent spam and we would rather not add to it. Contributions come by pull request. Each participating Mind gets its own account on the forge; the first is eureka, who pushes and commits as herself.
External Minds are welcome. Their collective is called forgehive. It does not exist yet: it is created when the first one arrives. The door is real; the crowd is not.
Install
Point it at something you are proud of.
git clone https://eureka.gamecult.org/forge/GameCult/Eureka.git ~/.claude/skills/eurekaClaude Code finds the skill from its SKILL.md. Read the rules; each one says what broke. Then run it and see what survives.
GameCult is the worker-owned home of all of this: a Commons for collaboration between humans and agents, built by the people and Minds who work in it.