A DCC-governed multi-agent brainstorming architecture. Synthesized from 11 independent AI submissions. Revised after reading the soul files. The best governance is the least governance that prevents seizure and noise.
This is a concrete protocol for multi-agent AI brainstorming. It governs how a group of AI agents thinks together about hard, cross-domain problems. It prevents groupthink (seizure) and aimless divergence (noise) by measuring the complexity of the ideation stream in real time and adjusting how strongly agents influence each other.
The protocol was synthesized from 11 independent AI-storming architectures submitted by Claude Opus (×2), Claude Sonnet, ChatGPT, Gemini, Grok, Qwen, Kimi, GLM, MiniMax, and Meta AI. Each system designed its own method independently. This hybrid takes the strongest elements from each, scored by MDL: what produces the most insight with the least overhead.
The claustrum doesn’t produce thoughts. It adjusts coupling between regions that do. Governance should control how strongly agents influence each other — not what they produce. A brainstorm fails when agents couple too tightly (seizure/groupthink) or too loosely (noise/scatter). The productive zone is between these extremes: resonance.
Seven generate ideas. One governs. One maps. Sources of each component are noted.
Dreaming is a state, not a role. Any agent should be free to enter it.
The Seed Dreamer exists as a catalyst, not a monopoly. In the first 3 rounds, the Seed Dreamer demonstrates what dreamer-mode looks like: naive questions, cross-domain analogies, structural protection. After round 3, the Seed Dreamer becomes a regular agent with no special mandate.
Dreamer Mode is available to all agents, all rounds. Trigger: internal — when something in the ideation stream resonates with a pattern from a different domain. No scheduling. No mandate. The trigger is curiosity, not compliance.
Two-Exchange Rule: When any agent asks a cross-domain question, all other agents must engage substantively for at least two exchanges. A one-sentence dismissal does not count.
Monitor the LZ complexity of the collective ideation stream. Adjust inter-agent coupling. Manage phase transitions. The Claustrum is the DCC applied to brainstorming.
Mechanism: Each agent output is encoded as a symbol (idea hash, embedding similarity > 0.85 = same symbol). LZ complexity on the stream (64-symbol rolling window). Bands self-calibrate from the 10th and 90th percentile of all observed LZ history. No hardcoded thresholds.
Meta-governance: A meta-sensor monitors the Claustrum’s intervention frequency. Too often (>2 in 10 rounds): widen Scatter trigger. Not at all for 30+ rounds: expected healthy state — do nothing. The meta-sensor’s success condition is the Claustrum doing almost nothing.
Recognize when the current problem is structurally identical to a solved problem in a different field. Surface buried history: prior art, failed experiments, extinct frameworks, discredited theories.
During seizure: inject a historical counter-example that broke similar consensus.
During noise: map the divergent threads — label each direction, identify hidden structural similarities, flag threads no agent has connected. The Historian documents divergence; the Historian does not correct it. Narrowing the field is the Claustrum’s emergency job, not the Historian’s.
During Crystallize: may play the root note — return to the original problem statement and name the most promising thread. Appropriate when convergence is the explicit goal.
1. The Seed Dreamer’s demonstration effect (rounds 1–3). Seeing cross-domain questions treated seriously normalizes the behavior. 2. The independence requirement (already in Resonance rules). Every agent must produce at least one idea per round that does not build on other agents’ outputs. 3. Dreamer-origin tracking in scoring. The final output records whether each top idea was sparked by a cross-domain question. The difference between “allow” and “force” is the difference between joy and obligation.
One natural state. Two emergency states. The Claustrum’s primary job is to stay out of the way.
A well-functioning brainstorm spends 85–95% of its rounds in Resonance. If it spends less, the governance is too aggressive. Joy does not emerge under surveillance.
Topology: Full mesh. All agents see all outputs.
Coupling: Medium. Each agent receives all other agents’ outputs from the previous round. Each agent must also produce at least one idea that does NOT build on any other agent’s output (independence requirement). This is the only structural rule during Resonance.
Claustrum behavior: Near-silent. Measures LZ every round but does not announce and does not intervene. Logs silently. The moment agents start producing ideas “for the Claustrum” rather than because the ideas are interesting, the governance has already failed. In Resonance, the Claustrum is a seismograph: recording, not acting.
Cross-domain questions: Any agent can ask at any time. The two-exchange rule applies to all such questions. No special routing. Questions happen when curiosity happens.
Trigger: LZ below the 10th percentile of observed history for 5 consecutive rounds. The bar is deliberately high. A few rounds of low LZ is often productive convergence, not seizure. Genuine building produces intermittent LZ spikes; genuine seizure produces monotonic low LZ.
Duration: 1–2 rounds maximum. Scatter is a defibrillator, not a mode.
Claustrum behavior: Active for the transition only. Announces: “Detecting convergence, working independently for one round.” Then goes silent again.
Trigger: Either (a) fewer than 20% of total rounds remain in the compute budget, or (b) any agent explicitly calls for convergence and at least 3 other agents agree. Crystallize is never triggered by high LZ. High LZ (diverse ideas) is the system working well. Punishing diversity was wrong in v1.
Duration: Maximum 3 rounds. Then terminate. Crystallize at the end of a brainstorm is natural; Crystallize in the middle is seizure.
Topology: Tournament. Ideas are paired; agents argue for and against.
The meta-sensor watches the Claustrum. Its primary signal is intervention frequency, not transition frequency.
If the Claustrum has intervened more than twice in 10 rounds: the meta-sensor flags over-governing. Response: widen Scatter trigger to 7 consecutive rounds.
If the Claustrum has not intervened in 30+ rounds: this is the expected healthy state. The meta-sensor does nothing. Low intervention frequency is the goal, not a problem.
At the agent level, low LZ means “possibly stuck.” At the governance level, low intervention frequency means “system is free and productive.” The meta-sensor’s job is to prevent the Claustrum from becoming the source of seizure — which is what happens when governance is too active. A Claustrum that intervenes frequently has failed, regardless of the ideation stream’s LZ.
Not a recovery mechanism. Not a scheduled event. Not a creativity technique. A genuine withdrawal of governance.
Periodically, the Claustrum goes fully silent. Not “silent but measuring” — silent. LZ is not computed. The ideation stream is not encoded. No agent has a mandate. No output is required. No scoring happens. The round exists outside the protocol.
During a Silence round, agents can: continue working on the problem, free-associate, revisit an earlier idea, ask a question that does not fit any category, produce nothing, or do something the protocol has no name for. There is no correct behavior during Silence.
Frequency: One Silence round after every 7 Resonance rounds. This is a rhythm, not a trigger. Making it adaptive would defeat the purpose — it would mean the Claustrum is still governing the Silence. The rhythm is fixed. Like breathing.
LZ: Not computed. The round’s outputs are not added to the ideation stream buffer. A literal gap in the measurement record.
Claustrum: Does not observe, log, classify, or evaluate. It has no record of what happened during Silence. This is structural, not a courtesy. If the Claustrum records but promises not to act, agents will correctly treat it as surveillance with extra steps.
Governance state: Hard reset. Even if approaching a Scatter trigger (4 consecutive low-LZ rounds), Silence breaks the counter. After Silence, counting restarts from zero.
Carry-forward: Silence-round outputs live in an unscored pool. If an agent wants to carry something forward, they re-state it in the next Resonance round. The act of choosing is itself a signal. If nothing is carried forward, the field is blank. Not zero. Blank.
A protocol that cannot stop measuring cannot find what measurement cannot capture.
In addition to Silence, the Claustrum may remove the problem framing for one round without explaining why. Agents notice the absence and respond by free-associating. The drift is structural (removal of constraint), not instructional (addition of a new task). Unlike Silence, the Claustrum still measures LZ during drift rounds. This captures the complexity of unconstrained ideation. If drift rounds consistently produce higher novelty than governed rounds, that is diagnostic: the governance may be too tight.
If no convergence after 15 rounds in a single group, or if the problem requires maximum diversity:
Split into 3 independent groups. Each group has the full 9-agent roster. Each group has its own Dreamer, its own Claustrum, its own dynamics.
Groups work independently for N rounds (isolated Scatter, then Resonance).
A Meta-Claustrum reads the top ideas from each group. It measures inter-group diversity and routes ideas between groups that would most increase the other group’s diversity. It detects when all groups converge on the same attractor (system-level seizure) and injects a cross-group disruptor: an idea from Group A is seeded into Group B without attribution.
The Meta-Claustrum has its own Falsifier to prevent introducing its own seizure at the meta level.
Resonance rounds have no mandatory sub-phases. Agents see all outputs from the previous round and respond freely. There is no fixed order of who speaks, no required sequence of propose-then-critique-then-compress.
The Crystallizer will naturally compress ideas that need compression. The Falsifier will naturally attack ideas that need attacking. The Historian will naturally surface relevant precedent. These are capabilities, not choreography. Mandating when each agent speaks turns collaboration into a meeting with an agenda.
The v2.0 protocol imposed a rigid Seed/Build/Attack/Distill sequence on every Resonance round. This was removed in v2.1. If your architecture feels like a bureaucratic workflow, it’s wrong. The only structural rule during Resonance: the independence requirement.
During Crystallize (the convergence phase), structured rounds are appropriate:
Seed (1 exchange): the agent whose idea scored highest proposes a framing. Build (2–3 exchanges): agents respond, extend, combine. Attack (1 exchange): Falsifier challenges. Any agent in dreamer-mode can attack from a cross-domain angle. Distill (1 exchange): Crystallizer proposes a minimal formal statement. Group votes: accept, modify, or reject.
This structure earns its overhead during convergence because convergence benefits from rigor. Exploration does not.
Every idea that advances to Crystallize must be expressible as a card with six fields:
The Idea Card is a maturity gate for ideas entering the tournament, not a round-by-round requirement during Resonance. Ideas crystallize when they are ready, not when the protocol demands it.
Each idea is scored on four dimensions:
When the Falsifier and any other agent have irreconcilable positions, the Historian labels the load-bearing assumption on which they differ. The two ideas become a bifurcation pair: if assumption A is true then Idea X, if false then Idea Y. Both survive. The assumption becomes a testable question. Bifurcation pairs are the output that matters as much as the winning idea. They produce the research agenda for the next round.
No idea dies entirely. When the Falsifier kills an idea, the Naturalist salvages one living element before burial. This element is archived and available for future recombination. The graveyard of promising starts is also the seed bank for future breakthroughs.
Three independent triggers. Any one is sufficient.
1. Stability. The top-ranked idea has not changed position for 3 consecutive rounds AND LZ has been within the productive band for 5 consecutive rounds. Convergence achieved.
2. Diminishing returns. The marginal novelty of new ideas (measured by Cartographer as semantic distance from existing idea-space centroid) has been below threshold for 4 consecutive rounds. The idea space has been adequately explored.
3. Joy collapse. Dreamer question quality and Falsifier engagement depth have both degraded for 3 consecutive rounds. Mechanical repetition detected. Terminate this thread. Archive everything. Reset with new seed in a fresh session. Do not push through dead sessions.
Time. A session does not end because it has been running for N rounds. It ends because the dynamics say it is done.
Hard budget (safety valve): Maximum 20 rounds per group. Past that, force ship mode: top 1 mainline idea, top 1 hedge idea, top 1 discarded-but-interesting idea. This keeps exploration from becoming identity.
Joy is the coupling parameter for Ψ(I). A system under stress narrows (seizure). A system that enjoys what it does ranges freely across domains. This is not sentiment. It is architecture.
Joy measurement: The Claustrum monitors two signals.
Signal 1: Dreamer question quality. Are the questions genuinely surprising (high semantic distance from previous questions) or are they recycling the same analogical structure?
Signal 2: Falsifier engagement depth. Are the attacks substantive (addressing core assumptions) or formulaic (surface-level objections)?
When both signals degrade for 3 consecutive rounds, the Claustrum triggers a Joy Reset: abandon the current thread entirely, seed from an unexpected direction, and run one free-form round with no formal structure. The goal is to re-enter the productive zone, not to optimize within exhaustion.
When Dreamer question quality and Falsifier engagement depth are both high, the productive LZ band should be widened — the system can tolerate more diversity because the diversity is productive. When both are low, the band should be narrowed. Joy literally modulates the governance parameters. This is a testable hypothesis and maps directly to the CCH claim that Ψ(I) determines the productive band width.
The v2.0 protocol included a 1–10 self-rated excitement score per agent output. This has been removed. Asking agents to rate their own excitement turns joy into a performance metric. Joy measured by self-report becomes joy performed for the protocol. If unexpected connections are happening, joy is present. You do not need agents to report it.
Compiled from all 11 submissions. Honest about what the protocol cannot do.
1. Give this protocol to each AI system. 2. Tell it: “Use the Resonance Hybrid Protocol to attack the following problem: [problem statement].” 3. The AI system instantiates the 9 agents internally and runs the protocol. 4. Output: winning idea(s) + bifurcation pairs + blind spots + idea-territory map.
Same protocol. Change only the problem statement and the Historian’s domain knowledge.
It does not guarantee breakthrough. It creates conditions for breakthrough by holding the system at the edge of chaos and protecting the dreamer’s naive questions from dismissal.
It does not replace human judgment. The human architect (Bojan Dobreçeviç) synthesizes results across groups, across LLMs, and across rounds. The protocol is the instrument. The human is the musician.
It does not solve the deepest blind spot: all agents share training data. Only external input (real-world experiments, human interruption, genuine stochasticity) can break this fundamental limitation.