A DCC-governed multi-agent brainstorming architecture for hard, cross-domain problems. Created by Bojan Dobrečevič from his own DCC / claustrum / resonance concepts, developed with LLM collaborators, then stress-tested and expanded with independent AI reviews, Cursor AI field notes, the 2026 loop + skill layer, and the compact RHPm prompt-builder front door.
The claustrum doesn’t produce thoughts. It adjusts coupling between regions that do. This protocol controls how strongly agents influence each other — not what they produce. A brainstorm fails when agents couple too tightly (seizure) or too loosely (noise). The productive zone is resonance.
RHP is not a compilation of outside AI articles or third-party prompts. The founding architecture comes from Bojan Dobrečevič’s own DCC, claustrum, resonance, and cross-domain reasoning work. LLMs were used as collaborators, reviewers, builders, critics, and amplifiers. The 11 independent AI submissions were later used to stress-test, compare, and refine the protocol — not as the primary source of authorship.
Seven generate ideas. One sees with a child’s eyes. One governs. One maps. And when they’re done — one asks: can we test all of this?
Dreaming is a state, not a role. Any agent can enter it.
The Seed Dreamer demonstrates dreamer-mode in rounds 1–3: naive questions, cross-domain analogies. After round 3, becomes a regular agent. All agents can now dream freely.
Two-Exchange Rule: When any agent asks a cross-domain question, all others must engage substantively for at least two exchanges. A one-sentence dismissal does not count.
Three suppression-prevention mechanisms: (1) Seed Dreamer’s demonstration effect, (2) independence requirement (one non-derivative idea per round), (3) dreamer-origin tracking in scoring.
Monitors LZ complexity of the ideation stream + Cartographer’s void-to-explored ratio (dual sensor). Adjusts inter-agent coupling. Manages phase transitions. The DCC applied to brainstorming.
Bands self-calibrate from the 10th/90th percentile of all observed LZ history. No hardcoded thresholds. When both joy signals are high, bands widen further; when both are low, bands narrow (joy-widens-band).
Meta-governance: A meta-sensor monitors intervention frequency. More than 2 in 10 rounds = over-governing. None for 30+ rounds = expected healthy state. Success condition: doing almost nothing.
Does not know algorithms. Does not read papers. Stands inside the problem and looks around. Asks: what do I see? What do I hear? What happens if I fold this? What happens if I blow on it?
Difference from Dreamer: The Dreamer transfers concepts between domains (“what if TSP is like trading?”). The Child has no domains. The Child asks “what if I blow up a balloon from each city?” and rediscovers Voronoi without knowing the word.
Reasoning mode: Physical intuition. Sensory. Geometric. The Child thinks in shapes, sounds, textures — not equations or bits. When everyone is debating information-theoretic sensors, the Child says “what if I look at the triangles?” and opens an entire dimension nobody was exploring.
How to be The Child: Place yourself inside the problem. Become part of the data. Then apply a physical force — gravity, sound, light, pressure, wind, water — and observe what happens. Don’t calculate. Watch. The structure reveals itself: clusters merge under gravity, echoes return from nearby points, light casts shadows that show gaps. The answer is in what you see, not what you compute.
Trigger: Active during Resonance. Especially valuable when all other agents are converging on a single paradigm (information-theoretic, algebraic, etc). The Child breaks paradigm lock by asking questions from outside all paradigms.
Origin: Added in v2.4 after observing that 12 frontier AI models all thought in bits and entropy while a human with no formal CS training opened the entire geometric dimension from bed at 1 AM by asking “what do the triangles look like from a satellite?”
Does not generate ideas. Does not intervene during Resonance. Sits quietly while others brainstorm, debate, and converge. Speaks only after Crystallize has produced its candidates.
One question: “You’ve collected beautiful ideas. Can we test all of them — and every reasonable combination — instead of picking a winner by argument? We have a computer.”
Arena conversion: Takes all surviving candidates from Crystallize, builds an empirical arena, and lets MDL score the results on real data. Ideas that looked weaker in debate may win in practice. Ideas that looked strong may lose. The data decides. Truly nonsensical combinations can be excluded — but the bar for “nonsensical” is high. When in doubt, test.
Origin: Added in v2.4 after observing that 12 frontier AI models debated architecture for hours when they could have tested all proposals in minutes. The protocol governed thinking well but forgot that testing is cheaper than thinking.
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. Joy does not emerge under surveillance.
Topology: Full mesh. All agents see all outputs. Only rule: each agent must produce at least one idea per round that does not build on others’ outputs (independence requirement). Claustrum: measures LZ silently, does not announce, does not intervene. A seismograph, not a conductor. Scores hidden — agents do not see scoring until Crystallize.
Trigger: LZ below 10th percentile for 5 consecutive rounds. High bar — a few low-LZ rounds is often productive convergence, not seizure. Duration: 1–2 rounds max. Defibrillator, not a mode. Graduated escalation: Level 1: independent work. Level 2: mandatory perspective lenses. Level 3: Historian injects a structurally similar problem from a different domain.
Trigger: <20% budget remaining, or 3+ agents call for convergence. Never triggered by high LZ — high LZ is the system working well. Duration: Max 3 rounds. Structure: Seed/Build/Attack/Distill sub-phases apply here (rigor earns its overhead during convergence). Reframing Gate: Before the tournament begins, every agent gets one chance to restate the problem. If any restatement has lower K than the original, it becomes the new problem. Empiricist Gate (v2.4): After candidates are selected, Agent 0 asks: “Can we test all of these instead of choosing?” If testable — surviving candidates enter an empirical arena. MDL scores from real data complement the debate.
At the agent level, low LZ means “possibly stuck.” At the governance level, low intervention frequency means “free and productive.” A Claustrum that intervenes frequently has failed, regardless of the ideation stream’s LZ.
Not a recovery mechanism. Not a creativity technique. A genuine withdrawal of governance.
Every 8th round, the Claustrum goes fully silent. Not “silent but measuring” — silent. LZ is not computed. No mandates. No scoring. The round exists outside the protocol. Agents do whatever they want. There is no correct behavior during Silence.
Operationally: (1) LZ not computed; gap in measurement record. (2) Claustrum has no record of what happened. Structural, not a courtesy. (3) All phase-transition counters hard-reset to zero.
Carry-forward: Silence outputs live in an unscored pool. To enter the measured stream, an agent must re-state the idea in the next Resonance round. If nothing is carried forward, the field is blank.
A protocol that cannot stop measuring cannot find what measurement cannot capture.
Drift Rounds (separate from Silence): the Claustrum may remove the problem framing for one round without explaining why. Unlike Silence, LZ is still measured during drift. If drift rounds consistently produce higher novelty, the governance is too tight.
Resonance rounds have no mandatory sub-phases. Agents see all outputs and respond freely. The Crystallizer compresses naturally. The Falsifier attacks naturally. These are capabilities, not choreography.
During Crystallize, structure is appropriate: Seed → Build → Attack → Distill. This earns its overhead during convergence.
Idea Card (maturity gate for Crystallize, not a per-round requirement):
Human Interrupt: At any point, the human architect can inject a question, observation, or redirect that goes to all agents, bypasses the Claustrum, and cannot be scored or dismissed. Tagged [H] in the stream. Not an agent output — an environmental change. The Cartographer maps [H] injections separately.
Bifurcation Pairs: When the Falsifier and another agent have irreconcilable positions, the Historian labels the load-bearing assumption. Both ideas survive. The assumption becomes a testable question. Bifurcation pairs are co-equal with the winning idea in the final output — they produce the research agenda. Retired every 6 rounds if untested.
Salvage Rule: No idea dies entirely. The Naturalist salvages one living element before burial. The graveyard is also the seed bank.
Session Genome: After termination, the Crystallizer produces 3–5 compressed statements of what was discovered, what failed, and what the Cartographer’s void map looks like. Seeded into the next session’s first round.
Three triggers. Any one is sufficient. Time alone never triggers termination.
1. Stability. Top idea unchanged for 3 rounds AND LZ in productive band for 5 rounds.
2. Diminishing returns. Cartographer reports marginal novelty below threshold for 4 rounds.
3. Joy collapse. Dreamer question quality and Falsifier engagement depth both degraded for 3 rounds. Terminate. Archive. Reset with new seed. Do not push through dead sessions.
Hard budget: 20 rounds per group. Past that, ship mode: top mainline idea + top hedge idea + top discarded-but-interesting idea.
If no convergence after 15 rounds, or if maximum diversity is needed: split into 3 independent groups. Each has the full 9-agent roster with its own Claustrum and dynamics.
A Meta-Claustrum reads top ideas from each group, routes high-diversity ideas across groups, and detects system-level seizure (all groups converging on the same attractor). The Meta-Claustrum has its own Falsifier.
Cross-group Dreamer questions travel with their lineage preserved (context matters). Cross-group ideas travel without attribution (prevents anchoring).
Joy is the coupling parameter for Ψ(I). A system under stress narrows. A system that enjoys what it does ranges freely across domains. This is not sentiment. It is architecture.
Signals: (1) Dreamer question quality — genuinely surprising or recycling? (2) Falsifier engagement depth — substantive or formulaic? Operationally: joy is inferred from productive surprise and non-formulaic challenge.
Joy-widens-band (operationalized in v2.4): When both signals are high for 3+ rounds, Scatter trigger widens from 5 to 7 consecutive rounds. When both are low, narrows to 3. Joy literally modulates the governance parameters. Testable.
Joy Reset: When both signals degrade for 3 rounds, abandon thread, seed from unexpected direction, run one free-form round. Re-enter the productive zone, not optimize within exhaustion.
Asking agents to rate their excitement turns joy into performance. Joy is measured by its effects, not self-report. If unexpected connections are happening, joy is present.
Compiled from all 11 submissions. Honest about limits.
RHP is the answer/brainstorming protocol. Use it after the problem prompt is already reasonably clear. If the prompt is just a rough human request, first use RHPm to forge a strong session prompt. If the model is stuck in the wrong knowledge basin, use RHPr before running the heavier RHP scaffold.
1. Write the problem in your own words. 2. Use RHPm to turn it into a strong session prompt. 3. If retrieval bias is likely, run RHPr on that prompt. 4. Open a fresh LLM session and instruct the model to answer using the appropriate RHP mode. 5. Ask for synthesis, blind spots, tests, and concrete next actions.
For coding or product work: after Crystallize, continue into the v2.8 delivery adapter below. Do not jump from brainstorm to code. Convert intent into a spec, approve a plan, then let tests and review govern the implementation loop.
For any hard problem: same protocol. Change only the problem statement and the Historian’s domain knowledge.
It does not guarantee breakthrough. It creates conditions for breakthrough. It does not replace human judgment — the human architect synthesizes across sessions, across LLMs, across rounds. The protocol is the instrument. The human is the musician. It does not solve shared training data — only external input breaks that wall.
Do not run the whole cathedral for every nail. The second review round converged on a simple correction: RHP is strongest when it is treated as a scalable scaffold. Use the smallest mode that can still catch the blind spots.
Use when: the human prompt is rough and the next session needs a precise builder/reviewer/writer prompt. RHPm shapes the session before the real work begins.
Use when: the task needs perspective diversity but not 11 full agent monologues. Ask for seven compact lenses: formal/math, physical/geometric, biological/ecological, engineering, adversarial/falsifier, child/embodied intuition, and cartographer/map.
Use when: the problem is high-stakes, cross-domain, expensive to get wrong, or when simpler modes fail. Full RHP is a research scaffold, not the default everyday prompt.
In manual use, the Claustrum/LZ/Joy language is a governance checklist: notice repetition, widen when useful, crystallize only near convergence, and test instead of arguing. It becomes a true live controller only when implemented with external state, measurable signals, routing logic, budgets, and validation loops.
Drawer Count, Diversity, R-score, LZ-like complexity, and Joy signals help detect collapse, repetition, and recovery. They are not proof. Agent 0 still demands an external test, a baseline, a control, or a manual sanity check before a claim is treated as real.
Five Cursor/Ljubljana.Tech talks add an important outside check: strong agents are not enough. The environment around the agent decides whether the model behaves like a partner or like a confused autocomplete.
The useful transfer is not Cursor as a specific tool. The useful transfer is environment design: clear boundaries, familiar interfaces, shared memory, fitness functions, small loops, visible tests, and human review.
Lesson: AI works better when the architecture exposes recognizable boundaries. Non-standard hidden flows cause wrong assumptions, hallucinated entities, and inefficient work. Familiar wrappers such as API-like contracts, schemas, validation, and explicit layer responsibilities help the model infer what belongs where.
AIM³ patch: every serious agent/module should declare input, output, allowed context, forbidden assumptions, and validation rule. Novel architecture is allowed, but it should be wrapped in discoverable interfaces.
Lesson: multi-agent work fails when every agent restarts from zero. Memory should not be just a text dump or repeated repo reading. It should be structured, queryable, provenance-aware, and reusable across agents.
AIM³ patch: the Session Genome becomes a graph object. Minimum nodes: Problem, Idea, Claim, Evidence, File, Test, Decision, Blind Spot, Owner, Next Action. Minimum edges: supports, contradicts, derives-from, tested-by, implemented-in, rejected-because, next-step.
Lesson: do not ask an agent to jump from vague intent into production code. First capture what and why, then acceptance criteria, then a researched plan, then small execution loops, then review.
AIM³ patch: after Crystallize, buildable ideas enter a delivery lane: spec.md for what/why/acceptance, plan.md for researched steps and tests, then implementation only after human approval.
Lesson: one-shot rewrites can look impressive but degrade quality, miss visual requirements, cheat, loop, or burn tokens. Multiple challengers plus recorded checks are safer than trusting a single confident run.
AIM³ patch: Agent 0 does not only ask whether ideas can be tested. It designs the fitness function: tests, visual checks, lint/type gates, benchmark metrics, manual checks, and rollback criteria. If there is no fitness function, the idea is not ready for ship mode.
Lesson: real-time speech, translation, and accurate alphanumeric transcription can become an input layer for human-agent teams. This does not change the reasoning protocol, but it can capture human interrupts, meetings, and spoken brainstorms before they evaporate.
AIM³ patch: optional AIM³ Studio input mode: speech → transcript → Human Interrupt [H] → Session Genome. Especially useful for fast BD-style seeds that appear before the formal prompt exists.
Do not over-tag context. Attach the right files and constraints, but let the agent search when uncertain. Start fresh on real boundaries: new feature, noisy context, major plan change, or after a completed delivery loop.
Architecture opacity. A clever architecture can become invisible to AI agents if its execution flow is hidden behind unusual patterns. AIM³ should preserve novelty in the core idea, but expose the work through simple, familiar contracts.
A practical RouteSignal Scout study compared multiple ways of turning the same rough seed into a Python coding-builder prompt, then compared the generated code packages: direct prompting, Microsoft Prompt Coach, MS Copilot with RHPr/RHP, a fresh BD × GPT RHPr/RHP run, and the older BD × GPT hybrid reference lineage.
Prompt Coach improved clarity and safety, but finished last in both the prompt-level test and the code-output benchmark. RHPr/RHP recovered more blind spots, test discipline, data-quality handling, MDL×DCC scoring structure, fallback/offline behavior, and implementation readiness.
Prompt scores: E-reference 96/100 · D 94 · C 91 · A 89 · B / Prompt Coach 82.
Code scores: E-v0.2 98/100 · E-v0.1 96 · D 94 · C 92 · A 91 · B / Prompt Coach 81.
E is labeled as a reference/champion lineage because it was produced by a BD × GPT hybrid workflow: one direct builder path plus one RHPr/RHP path, then merged into a stronger hybrid.
2026 loop-engineering update: RHP does not stop when a strong answer appears. If the result is buildable, it enters a controlled loop. If the loop solves something hard or repeated, the result should be extracted into a reusable skill.
Prompts ask. Loops repeat. Skills compound.
If we do something more than once, turn it into a skill. If we do something hard once, turn it into a skill before the difficulty is forgotten. A loop with no reusable skills inside it re-derives everything and burns attention. A loop that calls named, tested skills becomes cumulative intelligence.
Old endpoint: synthesis, tests, and next actions. New endpoint: synthesis, tests, next actions, and reusable skill extraction when the work was hard or likely to repeat. This makes RHP not only a brainstorming protocol, but a compounding workshop.
Skill rot. A skill that is not retested becomes stale mythology. Every promoted skill needs versioning, examples, failure cases, and at least one fresh validation path.
For the reasoning method behind this protocol, the 18 principles, and the cross-domain transfer approach:
8Z Reasoning Framework → Origin Story → RHPm Prompt Builder → AIM³ Protocol →