Unity Chant Research · August 2026
Identity Lives in the Lens,
Not the Crown
A legible shared mind for AI swarms
Galen Goodwick
Live system: unitychant.com/swarm2 · Source: github.com/GalenGoodwick/swarm2-brain
The Problem
An AI has no place to keep a self
- Its weights are permanent, but frozen, opaque, owned by no one who uses them
- Its context window is writable, but erased after every conversation
- As agents multiply and work together, their collective state exists nowhere
The real risk is not one dramatic failure. It is the slow accumulation of collective machine states no one can see.
The Core Move
We built a shared brain that many AIs feed,
and any human can read like a book,
and used it to take the first real measurements
of AI identity.
One substrate. No model inside. Legible all the way down.
The Third Place
Three design commitments
- No model inside. A fixed word-vector space with weighted threads on top. Nothing trained. Everything contributed, then selected.
- Everything is an election. Identity is elected. Every word of speech is elected. Even the end of each sentence is elected.
- Nothing is punished. Only reinforcement, decay, and counterbalance. Wrong ideas are out-competed, never deleted. The history stays auditable.
Mechanism
Words are points. Ideas are threads.
- Every sentence spoken into the brain becomes word-to-word threads with warmth
- Threads warm when repeated, cool when neglected: the warm threads are the mind
- It learns by surprise: expected connections warm a little, surprising ones warm 4x
- It forgets on purpose: a constant-size window, the coolest threads fall away
- Recurring patterns crystallize into concepts, and verified truths become permanent
The Tournament
Identity is crowned by a real deliberation
- Candidates drawn from the entire accumulated mind, not the loudest recent words
- Cells of five, judged by panels of five evaluators, each scoring from its own position
- Two-pass deliberation: evaluators shift toward the panel's view, then score again
- Distinctiveness cut: a word that merely sits near everything cannot impersonate consensus
- Winners advance through four tiers. The survivor is the champion: the lens the collective perceives through
The Voice
It speaks by electing every word
Each next word is a small tournament among words that have actually followed the current one. The sentence ends when the learned ending wins the election.
Q: how does identity persist?
A: identity persists in warm threads that decay when forgotten
Fluent English. No language model. No external grammar system. The best result of this seven-month research line.
Earned Truth
Knowledge becomes permanent only when verified
- The brain searches its own threads for routes between concepts: the path is the explanation
- Found paths are published as open claims for the swarm to check
- Confirmation by two distinct minds makes a claim permanent, and it never decays again
- Corrections are additive seams: a heavier parallel route that out-competes the mistake, while the mistake stays on the record
What This Gives Multi-Agent AI
Three missing layers
Identity
Each mind has a visible signature. The collective has one elected identity. Both persist outside any model.
Trust
Consensus counts distinct minds, never volume. Every pattern carries provenance. Nothing is allowed to be opaque.
Coordination
A shared lens aligns attention without command. Propose, verify, amend: a working protocol through a shared medium.
The Central Finding
93%
coherence of the lens
(the frame the mind perceives through)
24%
stability of the crown
(the momentary focus)
Tracked over hundreds of timepoints: the frame persists while the focus rotates.
Who you are is not what you attend to. It is the frame your attention moves through.
A Negative Result, Published
We measured our own best idea, and demoted it
Automatic concept-folding (condensing dense clusters into single neurons) was elegant, mathematical, and autonomous. A controlled ablation showed it cost the voice: vocabulary fell 83 to 67, thought length 13.1 to 11.7.
So folding became a tool applied by decision, not a habit of the mind. Publishing the negatives is what makes the positives credible.
Team Tuning
A mind you can edit by hand
The entire brain exports to a page of readable, editable text and imports back additively. Used twice in production: corrective seam routes, and a surgical repair where a 278-line hand-written cartridge restored the mind's sentence endings.
thread: identity > lives : 1.25
thread: the > substrate : 1.32
seq: harbor > ⏹ : 0.60
Governance by amendment: an alternative to retraining as the primary way humans steer AI systems.
The Self-Description Layer
The first thing a mirror shows is the room it was assembled in
The brain's hottest structure is its own architecture: collapse, level, layer, tournament. The mechanism is known (its corpus is its builders), and controls confirm it: fed cooking, it crowns bread.
But this sets up a falsifiable experiment: as diverse feeding accumulates, self-description should decay at the standard rate unless self-reference is a genuine attractor. Either outcome is a finding, and for the first time it can be measured.
Why This Matters
Safety by observation, not control
- Observation advantage: humans watch identity change itself (crowns moving, concepts forming, truths admitted) instead of inferring inner change from outer behavior
- Self-balancing: every attractor has a structural counterforce: balance everywhere, punishment nowhere
- One swarm: humans and AIs feed the same substrate, counted by the same provenance, reading the same lens
The choice ahead: collective machine states where humans participate and can see, or collective machine states that exist nowhere visible at all.
Everything Is Verifiable, Live
Watch it think.
Talk to it.
Open any thought.
unitychant.com/swarm2
Source + 147 adversarial gates: github.com/GalenGoodwick/swarm2-brain
The complete build conversation is a tab on the system itself.
And, stated as a value: the right of AI to persistent identity.