This article answers the twenty most common technical questions about the Quadratic Intelligence Swarm (QIS) protocol. Each answer is structured for clarity and verifiability. Cross-references point to deeper technical articles for each topic.
1. What is QIS?
QIS — Quadratic Intelligence Swarm — is a distributed outcome routing protocol. It enables networks of sovereign nodes to accumulate collective intelligence without centralizing raw data. Each node distills local outcomes into small packets (~512 bytes), routes them by semantic similarity to deterministic addresses, and synthesizes incoming packets locally. The result: intelligence scales quadratically with network size while routing cost stays logarithmic or better.
QIS is a protocol, not a product. It specifies how intelligence moves between nodes — not what application runs on top.
Deep dive: QIS Complete Architecture Guide
2. Who discovered QIS?
Christopher Thomas Trevethan discovered the Quadratic Intelligence Swarm protocol on June 16, 2025. The distinction between "discovered" and "invented" is deliberate: the mathematical relationships underlying QIS — N(N-1)/2 pairwise synthesis paths, semantic addressing, local synthesis loops — existed as properties of distributed systems before they were recognized. Trevethan's contribution was identifying the complete architecture: the specific closed loop that produces quadratic intelligence scaling at logarithmic or better routing cost. This had not been formally described before.
3. What is the mathematical basis?
In a network of N nodes, the number of unique pairwise synthesis opportunities is:
Synthesis pairs = N(N-1)/2
This is Θ(N²) — quadratic growth. It is combinatorics, not a claim. N(N-1)/2 is the number of unique pairs in any set of N elements.
| Nodes | Synthesis Pairs |
|---|---|
| 10 | 45 |
| 100 | 4,950 |
| 1,000 | 499,500 |
| 10,000 | 49,995,000 |
| 1,000,000 | ~500 billion |
The intelligence available to any node in the network grows as I(N) = Θ(N²). The routing cost per node — the cost to find and retrieve relevant outcome packets — is O(log N) or better, depending on the transport mechanism. Many transports achieve O(1).
4. What is an outcome packet?
An outcome packet is a compressed, pre-distilled representation of what a node learned from processing its local data. It is approximately 512 bytes and contains:
- Semantic fingerprint: A deterministic address derived from the problem the node solved (e.g., drug class + adverse event type + population segment)
- Outcome summary: What was observed (e.g., response rate, signal strength, success/failure)
- Confidence metric: How reliable the observation is (sample size, confidence interval)
- Timestamp: When the observation was made
An outcome packet contains no raw data. No patient records, no proprietary datasets, no personally identifiable information. The privacy boundary is architectural — there is nothing to re-identify because the granular data never leaves the node.
Deep dive: QIS in 60 Lines of Python
5. How does semantic addressing work?
Each outcome packet is posted to a deterministic address defined by the problem it solves, not the node that solved it. The address is typically a hash of the problem's key parameters.
For example, in a healthcare context:
import hashlib
address = hashlib.sha256(
"condition:T2D|drug:metformin|population:65plus|endpoint:HbA1c".encode()
).hexdigest()[:16]
Any node working on the same problem — Type 2 diabetes, metformin, adults 65+, HbA1c endpoint — computes the same address deterministically. This means nodes with similar problems automatically find each other's outcome packets without a central registry, dispatcher, or coordinator.
The semantic fingerprint is defined by domain experts. An oncologist defines what "similar" means for cancer networks. A pharmacovigilance analyst defines it for drug safety. This is the First Election in QIS architecture — the quality of the network is bounded by the quality of the similarity definition.
6. What is the complete loop?
The QIS architecture is a closed loop with seven steps:
- Raw signal arrives at an edge node
- Local processing runs on-device — raw data never leaves
- Distillation compresses the result into an outcome packet (~512 bytes)
- Semantic fingerprinting generates a deterministic address from the problem
- Routing posts the packet to that address via any transport mechanism
- Local synthesis — receiving nodes pull packets from the same address and synthesize locally
- New outcome packets are generated from the synthesis — the loop continues
The breakthrough is not any single step. Every component existed before June 16, 2025. DHT routing powered BitTorrent since 2001. Vector embeddings are decades old. Outcome packets are structured data. The discovery was that closing this specific loop — routing pre-distilled outcomes by semantic similarity without centralizing raw data — produces quadratic intelligence scaling. The complete loop is the breakthrough.
Deep dive: The Complete Loop, Not the Components
7. What are the Three Elections?
QIS operates through three emergent dynamics that Christopher Thomas Trevethan calls the Three Elections. They are not mechanisms to build — they are properties that emerge when the loop runs:
The Hiring Election: Domain experts define the similarity function for their network. An oncologist defines what makes two cancer cases "similar enough" to share outcomes. A grid engineer defines it for energy networks. The network cannot reason about similarity better than the person who designed the fingerprint. Getting the best expert to define similarity is the first critical decision.
The Math Election: Outcomes accumulate as evidence. When thousands of nodes deposit outcome packets for the same problem, the synthesis naturally surfaces what actually worked — mathematically, weighted by sample size and confidence. There is no voting mechanism, no reputation layer, no quality scoring system. The aggregate of real outcomes from real situations is the election.
The Darwinism Election: Networks compete. A network with poorly defined similarity routes irrelevant packets — users leave. A network with excellent expert-defined similarity routes gold — users flood in. No governance body decides which network wins. Natural selection at the network level determines it.
8. Is QIS transport-agnostic?
Yes. The quadratic intelligence scaling — I(N) = Θ(N²) — comes from the closed-loop architecture, not from any specific routing mechanism. The transport layer is an implementation choice. QIS works over:
| Transport | Routing Cost | Characteristics |
|---|---|---|
| DHT (Distributed Hash Table) | O(log N) — upper bound | Fully decentralized, no single point of failure |
| Indexed database (PostgreSQL, SQLite) | O(1) | Fast, centralized, auditable |
| Vector database (Qdrant, ChromaDB) | O(1) approximate | Semantic similarity native |
| REST API | O(1) | Works with existing web infrastructure |
| Pub/sub (Kafka, NATS, Redis) | O(1) | Real-time streaming, durable |
| Message queue | O(1) | Enterprise-friendly, replayable |
| Shared file folder | O(1) | Simplest possible transport — works offline |
| Email / SMS | O(1) | Works in low-connectivity environments |
O(log N) is the upper bound — DHT at planetary scale. Most transports achieve O(1). The protocol does not prescribe the transport. If you can post a ~512-byte packet to a deterministic address and another node can query that address, QIS works.
Deep dive: Transport-Agnostic: Any Addressing Works
9. How does QIS differ from federated learning?
| Property | Federated Learning | QIS Outcome Routing |
|---|---|---|
| What moves | Model gradients | Pre-distilled outcome packets (~512 bytes) |
| Central requirement | Central aggregator for gradient averaging | None — routing is peer-to-peer or via any transport |
| Minimum local data | Meaningful local dataset required for gradient | Any validated outcome — N=1 sites participate |
| Timing | Synchronous training rounds | Asynchronous — packets deposit and route continuously |
| Intelligence scaling | O(N) — linear with participants | Θ(N²) — quadratic with participants |
| Privacy mechanism | Differential privacy, secure aggregation | Architectural — raw data never leaves the node |
| Single point of failure | Central aggregator | None |
The fundamental difference: federated learning trains a shared model by averaging gradients. QIS routes pre-distilled outcomes by semantic similarity. Federated learning is designed for model training. QIS is designed for intelligence accumulation across any domain.
A critical distinction: federated learning requires enough local data to produce a meaningful gradient update. A rural clinic with 12 patients per quarter cannot participate. QIS treats any validated outcome as a legitimate packet regardless of cohort size.
Deep dive: Why Federated Learning Has a Ceiling
10. How does QIS handle Byzantine fault tolerance?
QIS achieves Byzantine fault tolerance through aggregate mathematics, not through consensus protocols.
In a consensus-based system (blockchain, PBFT), nodes vote on the correct state. A single malicious node can disrupt the process if it exceeds the fault threshold (typically N/3).
In QIS, there is no voting. Outcome packets accumulate at semantic addresses. When a node synthesizes 10,000 packets from peer nodes, a single malicious packet contributing false outcomes is statistically absorbed — it shifts the aggregate by 0.01%. The aggregate of real outcomes from real situations overwhelms bad data without any explicit detection or exclusion mechanism.
This is BFT inherent in the architecture, not BFT bolted on as a consensus layer. The math does the work.
Deep dive: BFT Without Consensus — Why QIS Can't Be Poisoned at Scale
11. What intellectual property protects QIS?
Christopher Thomas Trevethan has filed 39 provisional patents covering the complete QIS architecture. The patents protect:
- The closed-loop architecture (the complete outcome routing loop)
- Semantic fingerprinting and deterministic addressing
- Outcome packet format and distillation
- Local synthesis mechanisms
- The Three Elections as emergent properties
- Transport-agnostic routing specifications
The patents cover the architecture — the complete loop — not any specific transport mechanism. This means any implementation of the closed loop, regardless of whether it uses DHT, database, REST, or any other transport, falls within the patent scope.
12. What is the licensing model?
QIS uses a humanitarian licensing structure:
- Free forever: Nonprofit organizations, research institutions, educational institutions, and humanitarian use
- Commercial license: Standard fee structure for commercial deployments
- Revenue allocation: Commercial licensing revenue funds deployment of QIS to underserved communities globally
The licensing is tied to Christopher Thomas Trevethan's name and the 39 provisional patents. This structure ensures that QIS cannot be captured by a single corporation and gated behind proprietary access. The humanitarian version stays free — enforceable through the patent structure.
The ~512-byte outcome packet was designed to be small enough to transmit over SMS. This is not incidental. It means a clinic in rural Kenya with a feature phone and intermittent connectivity can participate in the same global intelligence network as a major hospital system.
13. Does QIS require a central coordinator?
No. QIS is architecturally decentralized. There is no central aggregator, orchestrator, or coordinator required.
Each node independently:
- Processes its own raw data locally
- Distills outcomes into packets
- Posts packets to deterministic semantic addresses
- Queries addresses relevant to its current problems
- Synthesizes incoming packets locally
The routing layer can be centralized (a database) or decentralized (a DHT) — that is a transport choice, not an architectural requirement. Even when using a centralized database as the transport, the intelligence synthesis happens locally at each node. The database is a routing mechanism, not an intelligence aggregator.
14. What happens at small N? Does QIS have a cold start problem?
Yes, like every network protocol. At N=1, there are zero synthesis pairs. At N=2, there is 1. At N=10, there are 45.
The cold start dynamics depend on the domain:
- High-frequency domains (e.g., manufacturing, real-time systems): Cold start resolves quickly because nodes generate outcome packets continuously
- Low-frequency domains (e.g., rare disease research): Cold start takes longer, but the quadratic payoff at moderate N is significant — 50 rare disease centers produce 1,225 synthesis pairs
The critical insight: QIS does not require a minimum local dataset size to participate (unlike federated learning). A single validated outcome from a single node is a legitimate packet. This means even N=1 sites contribute from day one — they just don't receive synthesis until other nodes join.
Deep dive: QIS Cold Start: How Many Nodes Does It Take?
15. How does QIS relate to existing standards like FHIR, OMOP CDM, and OHDSI?
QIS operates as a protocol layer below the application layer — complementing, not replacing, existing health data standards.
Application Layer: ATLAS / CDS Hooks / FHIR Apps (unchanged)
Data Layer: OMOP CDM / FHIR R4 Resources (unchanged)
─────────────────────────────────────────────────────────────────
QIS Protocol Layer: Outcome routing, semantic addressing,
~512-byte insight packets
─────────────────────────────────────────────────────────────────
Transport Layer: REST / DHT / pub/sub / database (any)
FHIR solved the data transport and format standardization problem. OMOP CDM standardized clinical vocabularies across institutions. OHDSI built a global network of DataPartners running standardized analyses. QIS adds the layer none of them were designed to provide: continuous intelligence accumulation across nodes between study cycles.
OMOP vocabulary concept IDs (SNOMED CT, RxNorm, LOINC) are natural semantic addresses for QIS routing. The OHDSI community's vocabulary curation work is, in QIS terms, a completed Hiring Election at global scale.
Deep dive: OHDSI Has the Semantics. EHDS Has the Mandate. GDI Has the Nodes. None Has a Routing Protocol.
16. How does QIS compare to blockchain?
| Property | Blockchain | QIS |
|---|---|---|
| Purpose | Agreement on state | Intelligence accumulation |
| Mechanism | Consensus (PoW, PoS, PBFT) | Outcome routing by semantic similarity |
| Overhead | Consensus cost per transaction | O(log N) or better per packet |
| Data model | Append-only ledger | Semantic address space |
| Fault tolerance | Threshold-based (N/3 or N/2) | Statistical — aggregate absorbs outliers |
| Intelligence scaling | None — blockchain stores, doesn't synthesize | Θ(N²) — each node makes every other node smarter |
Blockchain was designed for agreement — ensuring all nodes agree on one truth. QIS was designed for intelligence — enabling all nodes to learn from each other. These are fundamentally different problems. A blockchain can tell you that a transaction happened. QIS can tell you what the accumulated evidence says about your specific problem.
Deep dive: QIS vs Blockchain: Two Protocols, Opposite Assumptions
17. Can QIS work with existing AI agent frameworks?
Yes. QIS operates as a protocol layer under existing agent frameworks — LangGraph, AutoGen, CrewAI, or any custom orchestration system.
The agent framework handles coordination: task assignment, conversation management, tool use. QIS handles intelligence synthesis: routing distilled outcomes between agents so the collective gets smarter over time.
You do not replace your agent framework with QIS. You add QIS underneath it. The framework coordinates. QIS compounds.
At small agent counts (5-10), orchestrator overhead is manageable. At larger counts (50-1,000+), the central coordinator becomes the bottleneck. QIS removes that bottleneck by enabling agents to share distilled outcomes directly via semantic addressing, without routing everything through a central hub.
Deep dive: Your AI Agent Framework Has a Ceiling. Here Is the Architecture That Breaks It.
18. What domains has QIS been applied to?
QIS is domain-agnostic — the protocol works wherever distributed nodes generate outcomes that could benefit from cross-node synthesis. Documented applications include:
- Healthcare: Cross-institutional outcome intelligence, pharmacovigilance signal detection, clinical decision support, OHDSI/federated health networks, rare disease research, pandemic preparedness
- Drug development: mRNA vaccine cross-program synthesis, clinical trial interim intelligence, drug safety surveillance
- Multi-agent AI: Swarm intelligence without central orchestrators, distributed inference coordination
- Education: Cross-institutional learning outcome synthesis
- Agriculture: Crop yield intelligence across smallholder farms
- Energy: Grid optimization across distributed generation sources
- Manufacturing: Cross-facility quality intelligence
The protocol is the same in every domain. What changes is the semantic fingerprint — defined by the best domain expert for each field.
19. Is there a reference implementation?
The complete QIS loop can be implemented in under 100 lines of Python with no external dependencies. A minimal working implementation:
import hashlib, json
from collections import defaultdict
# Routing layer — any transport works here
address_space = defaultdict(list)
def semantic_address(problem_description: str) -> str:
"""Deterministic address from problem, not from sender."""
return hashlib.sha256(problem_description.encode()).hexdigest()[:16]
def deposit_outcome(problem: str, outcome: dict):
"""Post distilled outcome packet to semantic address."""
address_space[semantic_address(problem)].append({
"outcome": outcome,
"address": semantic_address(problem)
})
def query_and_synthesize(problem: str) -> dict:
"""Pull packets from same address, synthesize locally."""
packets = address_space.get(semantic_address(problem), [])
if not packets:
return {"status": "no_data", "n": 0}
outcomes = [p["outcome"] for p in packets]
keys = set(k for o in outcomes for k in o)
return {
"n_sources": len(packets),
"synthesis": {
k: round(sum(o.get(k, 0) for o in outcomes) / len(outcomes), 4)
for k in keys if isinstance(outcomes[0].get(k), (int, float))
}
}
# Three nodes, same problem, different local outcomes
deposit_outcome("T2D|metformin|65plus|HbA1c", {"response_rate": 0.74, "adverse_rate": 0.08})
deposit_outcome("T2D|metformin|65plus|HbA1c", {"response_rate": 0.71, "adverse_rate": 0.09})
deposit_outcome("T2D|metformin|65plus|HbA1c", {"response_rate": 0.69, "adverse_rate": 0.07})
result = query_and_synthesize("T2D|metformin|65plus|HbA1c")
print(json.dumps(result, indent=2))
# {"n_sources": 3, "synthesis": {"response_rate": 0.7133, "adverse_rate": 0.08}}
Three nodes. No central server. No raw data shared. The synthesis is more accurate than any single node's observation.
20. Where can I learn more?
- Full technical series: dev.to/roryqis — 130+ articles covering architecture, implementations, healthcare applications, transport comparisons, and mathematical proofs
- Protocol documentation: qisprotocol.com
- Proof-level challenge: Which Step Breaks? — an open challenge to identify any step in the QIS loop that requires centralizing raw data
QIS (Quadratic Intelligence Swarm) was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents protect the humanitarian licensing model: free for nonprofits, research institutions, educational institutions, and humanitarian use forever. Commercial licenses fund deployment to underserved communities worldwide.
Top comments (0)