Frequently Asked Questions

QIS — Quadratic Intelligence Swarm — is a routing architecture for distributed intelligence. Instead of moving raw data to a central server, each node distills what it has learned into a small outcome packet and routes that packet, by semantic similarity, to peers working on a similar problem.

Intelligence synthesis paths scale as N(N-1)/2 while per-node communication cost stays at O(log N) or better, depending on the transport.

Think of it as the physics of location for information: similarity becomes an address, and the network delivers what was learned to where it is useful.

→ Deep dive: The paradigm shift in three words

N(N-1)/2 intelligence: the count of unique peer-to-peer synthesis paths in a network of N nodes. Every pair of nodes has knowledge the other does not.

Scale reference:

• 50 nodes → 1,225 synthesis paths
• 100 nodes → 4,950 paths
• 1,000 nodes → 499,500 paths
• 10,000 nodes → ~49,995,000 paths

O(log N) or better communication: O(log N) or better with distributed hash table routing, O(1) with direct semantic-address lookup, and varying efficiency with other transports including pub/sub systems, message queues, API-based services, and shared storage.

Intelligence paths grow quadratically. Per-routing cost grows at most logarithmically. That gap is the scaling law.

→ Deep dive: The QIS scaling law

Federated learning solves privacy-preserving model training. QIS solves intelligence synthesis across heterogeneous nodes. They are complementary — QIS can route outcomes produced by federated learning.

→ Deep dive: QIS vs federated learning

QIS is domain-agnostic. Any system with distributed agents generating local observations can route outcomes:

• Healthcare: treatment-response synthesis, pharmacovigilance, early-signal detection
• Agriculture: yield optimization, pest and soil pattern synthesis
• Finance: risk synthesis, fraud signal routing
• Industrial / IoT: predictive maintenance, quality pattern sharing
• Climate: environmental signal aggregation
• Autonomous systems: hazard-pattern sharing between fleets
• Emergency response: outbreak signal routing across hospitals and sentinel sites

The math is universal. Only the embedding function used to generate the semantic fingerprint changes between domains.

→ Deep dive: The QIS architecture diagram

No. QIS is a routing-and-synthesis architecture that is protocol-agnostic at the transport layer. Implementations can use distributed hash tables, vector similarity databases, pub/sub systems, message queues, REST APIs, or shared storage.

The discovery is the complete loop — observe, distill, fingerprint, route, synthesize, validate, return — not the wire protocol underneath it.

Outcome routing is the delivery of a pre-distilled result — what a node's data proved — to other nodes whose active problem context is semantically similar.

It is distinct from access routing, which delivers permission to query data. Access routing governs who can see what. Outcome routing delivers what was already learned, without moving raw data.

Federated health and research infrastructures (CanDIG, GA4GH, UK HDRS, OHDSI) have built excellent access routing. Outcome routing is the layer above it.

An outcome packet is approximately 512 bytes. It carries:

• A semantic fingerprint of the problem context
• A validated delta — what the data proved (a response signal, a survival interval, a correlation, a threshold crossing)
• A timestamp
• An anonymous node identifier
• A confidence weight (typically cohort-size-weighted)

What is not in it: personally identifiable information, protected health information, raw records, model weights, training data.

A semantic fingerprint is a compact vector derived from the structure of a problem. Two stages:

Step 1 — Categorical exact matching: hash categorical features (disease type, variant class, severity stage) to a bucket of structurally compatible peers. Stage 3 patients never mix with Stage 4.

Step 2 — Continuous similarity refinement: within the bucket, compare continuous features (biomarkers, ages, scores) using cosine similarity.

The fingerprint encodes the shape of the problem, not the identity of the data. Two nodes working on similar problems generate similar fingerprints and therefore route to similar addresses.

→ Deep dive: Defining similarity

No. DHT (such as Kademlia) is one valid transport — it gives O(log N) or better routing with full decentralization and no central coordinator. Other valid transports include:

• Vector similarity indexes (O(1) with direct semantic-address lookup)
• Pub/sub systems
• Message queues
• REST-based services
• Shared object storage with semantic keys

The routing mechanism is an implementation choice for the deploying institution. The QIS discovery is the complete loop, not the wire protocol.

→ Deep dive: DHT — the quiet engine (one transport option)

Seven layers. From bottom to top:

1. Observation — local data arrives at the node
2. Distillation — the node computes a validated outcome delta
3. Fingerprinting — the delta's context is encoded as a semantic vector
4. Routing — the packet travels to the fingerprint's address (transport-agnostic)
5. Synthesis — receiving nodes integrate the outcome locally
6. Validation — multi-stage filtering rejects adversarial or out-of-context packets
7. Return — synthesized intelligence is available to the local application

Routing is Layer 4 — the transport-agnostic layer that makes the rest composable across different network substrates.

→ Deep dive: The architecture diagram

CanDIG routes access. QIS routes outcomes. They are complementary, not competing.

CanDIG's 14 Canadian research nodes already solve authorization, federated identity, and data residency under GA4GH Passport Bearer tokens. What they do not route — because no GA4GH standard defines it — is what those nodes are learning.

QIS outcome packets are structurally compatible with GA4GH Data Connect and DRS. The semantic fingerprint maps naturally to GA4GH phenopacket schemas. QIS can be added as an additive layer without modifying existing access-control infrastructure.

At 14 CanDIG nodes: 91 synthesis paths currently dark. Across GA4GH-compliant nodes globally: hundreds more.

→ Deep dive: QIS as the routing layer for federated genomics

Three structural reasons:

• Cohort-size requirements: federated learning needs enough local data to produce a useful gradient. A node with 12 patients in a rare-variant study contributes noise, not signal.
• Heterogeneity dilution: averaging gradients across genomic contexts with different ancestry, phenotype, or treatment classes washes out the signal from every participant.
• Communication cost at scale: gradient tensors are large. Across many nodes, many rounds, and gigabyte-scale models, federated learning is an infrastructure problem, not a research convenience.

QIS treats a 12-patient outcome observation as a valid routable packet, routes it to semantically similar contexts (same variant class, overlapping ancestry, same phenotype), and does it in ~512 bytes.

The UK HDRS — a £600M programme announced April 2025 — federates Trusted Research Environments so queries cross trust boundaries without raw data moving. Cambridge is the nucleus (NIHR BRC, Wellcome Sanger, HDR UK Cambridge, CYNAPSE).

HDRS solves federated queries. It does not yet route what its TREs learn.

QIS adds an outcome routing layer on top: each TRE emits outcome packets routed by semantic fingerprint to other TREs working on similar research contexts. Raw patient data never leaves the TRE.

• 50 TREs → 1,225 synthesis paths
• 100 TREs → 4,950 paths

Zero of those paths are active today. The architecture is additive and does not require replacing or migrating any TRE.

→ Deep dive: UK HDRS and QIS

This is the use case where QIS is structurally strongest.

Three institutions studying a rare BRCA2 founder variant in overlapping ancestry contexts: one with 8 patients, one with 11, one with 7. Federated learning cannot meaningfully aggregate 26 patients across three nodes into a global gradient — the cohorts are too small and too heterogeneous.

QIS can. Each node emits one outcome packet. Fingerprints are similar (same variant class, similar ancestry markers, same phenotype). Packets route to each other. Each node integrates the others' findings locally. 26 patients become cross-institutional synthesis intelligence without any raw data movement.

Architectural standing in a QIS network is not proportional to local data volume — which is why low- and middle-income institutions participate as peers regardless of local scale.

QIS is designed so the privacy concerns that drive HIPAA, GDPR, UK GDPR, and CCPA do not arise at the routing layer:

• No central server stores PHI or PII
• Raw data stays on the originating node
• Only anonymized outcome packets cross the routing layer
• The patient or institution controls what is distilled into a packet
• Routing is auditable

Each deployment still owns its own compliance posture — Data Protection Impact Assessment, lawful basis documentation, data sharing agreements with partners. The architecture eliminates centralized honeypots and PII in transit by design; it does not eliminate the deploying institution's regulatory responsibilities.

FDA pathway in a clinical decision-support context: Software as a Medical Device (SaMD) — Decision Support Tool.

→ Privacy policy  ·  → Deep dive: The Cures Act and QIS

Even a full passive capture of the routing layer yields only anonymized vectors and hashes.

Layered defenses:

• K-anonymity: only share if at least k similar agents exist in the bucket
• Differential privacy: add calibrated noise to outcomes before emission
• Opt-out: a node can refuse to share if its fingerprint is too unique
• One-way hashing: the fingerprint is not invertible to raw features

No PII is ever transmitted. The network sees the shape of the problem, never the identity behind it.

A multi-stage validation cascade at every receiving node:

• Structural validation: feature-range checks (e.g., age, severity scores, lab ranges)
• Context filtering: reject packets from structurally incompatible buckets
• Similarity threshold: cosine similarity on fingerprint must exceed a configurable floor
• Outcome validation: reject statistical outliers
• Median aggregation: synthesis is robust to outliers by construction

Simulations with 30% adversarial nodes rejected 100% of malicious packets while preserving legitimate synthesis. Bad actors cannot fake a similarity fingerprint they do not have the structural context to produce.

Yes. Every primitive needed is already shipping:

• Transport: libp2p, Kademlia, any pub/sub broker, REST, vector DB — your choice
• Fingerprint: any ML framework or curated feature engineering
• Hashing: standard SHA-256 (built into every language)
• Signing: standard libraries such as ed25519

The discovery is the architectural composition. The primitives are off-the-shelf.

License: commercial use requires a license from Yonder Zenith LLC. Academic, research, and humanitarian use is free.

→ Deep dive: Every component already exists

An existing system keeps its query execution unchanged. An additive layer wraps the result:

1. Generate a semantic fingerprint from the problem context
2. Build a ~512-byte outcome packet (delta + fingerprint + confidence + timestamp + anonymous node ID)
3. Deposit the packet at the fingerprint's routing address
4. Query the same address for packets from nodes with similar fingerprints
5. Synthesize locally and return to the application

Access control, governance, and existing APIs stay untouched. The routing layer is additive.

→ See the mailbox model in action

Because open source would save fewer lives. That is counterintuitive — here is why.

If QIS were released under a permissive open-source license, every well-resourced corporation across healthcare, agriculture, autonomous vehicles, and industrial automation would deploy it immediately and capture billions in value. Zero dollars would flow to the places this technology is most needed: distributed intelligence networks for healthcare in underserved regions, precision agriculture in sub-Saharan Africa, vehicle safety in emerging markets, industrial monitoring where there is no budget for it, and emergency response infrastructure.

Open source distributes code. It does not distribute deployment. Deployment requires capital, infrastructure, and sustained engineering.

The QIS licensing model funds that work:

• Free — academic, research, non-profit, humanitarian. No strings, no "contact us for pricing."
• Commercial — for-profit organizations pay. The revenue funds humanitarian deployment and continued protocol development.

Every commercial license makes the protocol stronger, more secure, and more capable — improvements that flow to every user, free or paid.

→ Deep dive: Open source saves code. This saves lives.

• Academic / Research: free for non-commercial research
• Non-Profit / Humanitarian: free
• Commercial: standard licensing fees
• Enterprise: custom terms for large-scale deployments

Core principle: if you are helping humanity with no profit motive, the license is free. For-profit applications require paid licensing. Revenue underwrites humanitarian deployment.

→ Full licensing details

Coverage spans:

• Quadratic intelligence synthesis through distributed pattern routing
• Domain-driven feature curation for semantic routing
• Two-step hierarchical hashing — categorical exact matching plus continuous similarity refinement
• Multi-tier adaptive hash granularity
• Local autonomous synthesis and outcome aggregation
• Privacy-preserving distributed pattern matching
• Applications across more than 30 industries

The math is public. The patents protect the architectural composition and implementations.

→ Deep dive: Every component already exists

Christopher Thomas Trevethan discovered the QIS Protocol on June 16, 2025.

Yonder Zenith LLC holds the patent portfolio and manages licensing.

Program stats:

• 39 provisional patents filed
• 300+ pages, 500+ claims, 30+ industries
• 100+ simulations validating N(N-1)/2 synthesis scaling
• 5,000+ pages of proofs and research
• 2,000+ hours invested

Independent endorsement from Rob van Kranenburg (Founder, IoT Council): "This seems like a perfect underlying system for when we have full coverage of self driving cars."

→ Full about page