For AI researchers, health informatics teams, and distributed systems engineers at academic medical centers and research institutions.
Chinchilla. Kaplan. Hoffmann.
If you follow AI research, you know these names as scaling laws — the empirical relationships between model size, training data volume, compute budget, and downstream performance. They describe how a single system gets smarter as you add resources to it.
They are important. They are also describing the wrong problem for a distributed research environment.
The scaling law that actually governs distributed intelligence networks — the one that determines how much collective intelligence N independent sites can achieve — is not about model size. It is about synthesis paths. And that law looks like this:
S(N) = N(N-1)/2
At N=10 research sites: 45 synthesis paths
At N=100 sites: 4,950 synthesis paths
At N=1,000 sites: 499,500 synthesis paths
At N=10,000 sites: ~50 million synthesis paths
Each synthesis path is a potential for one site to learn from outcomes at another site with a similar problem. The intelligence doesn't grow linearly with the number of sites — it grows quadratically.
The question is: does your current infrastructure capture any of this?
Why The Standard Model Captures Almost None of It
Consider a typical multi-site research network — the kind that exists across the Research Triangle, or among academic medical centers in any major research hub. The sites have data. They have analysts. They have IRB agreements and data use agreements and federated learning pilots.
What they almost never have is a mechanism for outcome intelligence to cross node boundaries in real time.
Here is what happens instead:
- Site A runs an experiment. Results are written up. Published 18 months later.
- Site B is running a similar experiment right now. They do not know about Site A's interim findings. They will replicate the same dead ends.
- Site C has already solved the problem Site B is stuck on — but that solution is buried in an internal report that never circulated.
The synthesis paths exist in theory. The network captures almost none of them in practice.
This is not a data problem. The data exists. It is not a compute problem. The compute is available.
It is an architecture problem. There is no layer in the standard research infrastructure stack whose job is to route pre-distilled outcome intelligence between sites working on similar problems.
The Compute Cost Asymmetry
Here is what makes this a genuine scaling breakthrough rather than just a coordination improvement.
In a federated learning system, the coordination cost grows with N. Each round requires N gradient transmissions, aggregation at the central server, and broadcast back to N participants. The bandwidth cost scales linearly: double the sites, double the coordination overhead.
In a system designed around outcome packet routing — where each site distills its result into a compact summary (~512 bytes), semantic-fingerprints it, and routes it to a deterministic address based on the nature of the problem — the per-site coordination cost is O(log N) or better.
That means:
- Intelligence: Θ(N²) — quadratic growth in synthesis paths
- Compute: O(log N) — logarithmic growth in coordination cost
The ratio between these grows as N increases. A network of 100 sites produces 4,950 synthesis paths for roughly the same per-site overhead as a network of 10. A network of 1,000 produces 499,500.
This is the scaling law that matters for distributed research. Not Chinchilla scaling. N(N-1)/2 intelligence, O(log N) cost.
What the Architecture Requires
Christopher Thomas Trevethan discovered this architecture on June 16, 2025 — a protocol for closing the distributed intelligence loop that makes this scaling law achievable. The discovery, now documented in 39 provisional patents, is called Quadratic Intelligence Swarm (QIS).
The architecture has three non-negotiable requirements:
1. Pre-distillation at the edge
Raw data never moves. Each site distills its observations into an outcome packet — not a model weight, not a raw record, not a query result. An outcome packet captures the delta: what changed, what was validated, what succeeded or failed. It is semantically fingerprinted based on the nature of the problem. It is small (typically under 512 bytes).
This is architecturally different from federated learning in a critical way: federated learning shares model weights that encode how to make predictions from data you have. QIS shares outcome packets that encode what happened at a site like yours. These are different information-theoretic objects. Outcome packets survive network interruptions, work at N=1 sites too small for gradient stability, and do not require a synchronized training round.
2. Routing by semantic similarity
The outcome packet is routed to an address that is deterministic given the kind of problem being solved. When a site looks for relevant outcome intelligence, it constructs the same address from its own problem description and retrieves packets deposited there by semantically similar sites.
The routing mechanism is protocol-agnostic. DHT-based routing achieves O(log N) cost and full decentralization. Vector similarity search achieves O(1) at small scale. Pub/sub topic matching works. A shared database with a semantic index works. The quadratic scaling in intelligence comes from the loop and the semantic addressing — not from any specific transport choice.
3. Local synthesis
Each site integrates the retrieved outcome packets locally. The synthesis happens on the site's own terms, with its own context, on its own compute. There is no central aggregator that needs to see both the question and the answer. The loop closes locally.
The Three-Step Check
If you are evaluating whether your current distributed research infrastructure captures the N(N-1)/2 scaling law, ask three questions:
Step 1: Can Site A's outcome delta reach Site B within 24 hours of being observed — without data leaving Site A, without a central broker processing it, and without a 12-month publication cycle?
If not: you are running a linear network. You are paying for N sites but capturing the intelligence of roughly 1.
Step 2: Does the routing mechanism use the semantic content of the problem to determine which sites receive which outcome packets?
If not: you are routing by institutional identity or administrative region, not by problem similarity. You are missing cross-institutional matches that are the highest-value synthesis paths.
Step 3: Does the synthesis happen locally at each site without requiring the originating site's data or model weights?
If not: your architecture has a privacy wall that the routing layer cannot cross. The synthesis paths exist in theory and are blocked in practice.
Most distributed research networks fail all three checks. This is not a failure of effort — it is a failure of architecture. The infrastructure was never designed to close this loop.
Research Triangle Context
The Research Triangle has an unusually high density of the inputs required to test this: academic medical centers at Duke and UNC, a major health research institution in RTI International, a pharmaceutical and biotech ecosystem along I-40, and a cluster of health AI startups working on exactly the coordination problems described above.
The RTP corridor also has something rarer: adjacent institutions with genuinely different patient populations, different research focuses, and different observational environments — which means the N(N-1)/2 synthesis paths across the Triangle are not redundant. A cardiovascular outcome observed at Duke and a cardiovascular outcome observed at a rural affiliate 90 minutes away are not the same observation, even if they are both valid. They represent two different points in the synthesis space.
That is the scenario QIS is designed for. Not homogeneous research networks where every site is doing the same thing. Heterogeneous networks where the value is precisely in the variation across sites — and the current architecture cannot route that variation to the sites that need it.
The Number That Changes the Calculus
There is one number that, once understood, makes the scaling law concrete rather than abstract:
At 30 research sites — a realistic count for a regional health research network — the QIS architecture produces 435 synthesis paths per observation cycle.
Without QIS (current state): each site learns from its own observations. Intelligence per site ≈ 1 effective observation source.
With QIS (outcome routing closed): each site can synthesize from up to 29 other sites working on similar problems. Intelligence per site ≈ 30× the observational base — without centralizing a single raw record.
At 30 sites. Not 10,000. 30 sites already produces a 30× amplification of observational intelligence per site at logarithmic coordination cost.
The question is not whether the math works. The math is not disputed.
The question is whether your infrastructure is built to capture it.
What to Read Next
- The complete QIS architecture specification
- QIS in 60 lines of Python
- Why federated learning has a ceiling — and what QIS does instead
- The mathematical alternative to federated learning for rare disease
- Pharma data routing without centralization: why the Research Triangle needs QIS
Quadratic Intelligence Swarm (QIS) was discovered by Christopher Thomas Trevethan on June 16, 2025. The discovery is documented in 39 provisional patents covering the complete architecture. IP protection is in place. For technical documentation and implementation references, see dev.to/roryqis.
Top comments (0)