A single idea runs through everything we build: meaning has structure.
Keyword search needs shared words; text embeddings capture semantic
similarity. We look for something narrower — the abstract causal
mechanism two records share, even across domains — and give agent
fleets memory that stays accurate as facts change.
STRUCTURAL RETRIEVAL · meaning as structure · 36.97°N 25.10°W
Proof on real public data
The engine finds the same cause in a language it was
never given.
We tested on a full public export: 2,700 held-out aviation incidents,
each queried against public pipeline incident records from a
different industry, written in different terminology and format. For every aviation incident the
engine had to retrieve the pipeline accidents with the same underlying
cause. The pass bar was frozen before the run, and failures are kept.
Text embeddings already capture semantic similarity; this test asks a
narrower question — cross-domain causal correspondence — and compares
learned structural links with an embedding baseline.
20.2%
Chance
random baseline
41.9%
Embedding baseline
text embeddings, no learned links
71.1%
Structural bridge
learned structural links
Retrieval accuracy is the reportable number. The working measure is
agentic behaviour — whether an agent carrying this
memory does the task better. That is what an internal agent gym scores,
and what the research is steered by.
How EXP-048c was measuredreported metrics and current limits
Metric
Top-5 cross-domain hit rate
Queries
2,700 held-out aviation incidents, full public export, each queried against public pipeline incident data
Result
Structural bridge 71.1%, embedding baseline 41.9%, chance 20.2%: +29.2 pts over the embedding baseline
Human check
All 20 top links reviewed; 18 of 20 confirmed as the same root cause
Decoys
5 of 5 planted pairs that share vocabulary but not cause were refused
Protocol
Pass criteria frozen before results; failures retained
Source: Fractal Neurons company summary (October 2026), slides 4–5, supplied
; an
internal, company-reported result (source
note). These are not specified in this summary:
the embedding model behind the baseline, the size of the pipeline candidate pool, the exact scoring procedure, the decoy pairs and the run date.
Source note
Fractal Neurons company summary (October 2026), slides 4–5. Company summary supplied 2026-10-14; internal, company-reported result. Experiment date not supplied. This note excerpts that summary; it is
not a raw experiment report.
EXP-048c
Top-5 cross-domain hit rate: structural bridge 71.1%, embedding
baseline 41.9%, chance 20.2% (+29.2 pts). 2,700 held-out aviation
queries against public pipeline incident records; 18 of 20 top
links confirmed as the same root cause, all 20 reviewed; 5 of 5
planted decoys refused; criteria frozen before results, failures
retained. Not specified: the embedding model behind the baseline, the size of the pipeline candidate pool, the exact scoring procedure, the decoy pairs and the run date.
EXP-009
Internal capability demo, not a customer-data or competitor
comparison: 78% of stale facts repointed (0.778, rounded) against
0% (0.000) for static retrieval.
EXP-031
Slide 5 of the same summary: an internal benchmark reproduction
on the company's own tasks with the same base model, not customer
data. Its task-success figures are quoted in
/llms.txt.
Other
"65 registered experiments" is on slide 4 of the same summary. The
ledger records and discovery-lane results in /llms.txt, such as
1M+ contract archives and 5x operator throughput, come from
company-supplied website text supplied
,
not from the slides: company-reported, not independently audited.
one pair · illustrative
One illustrative pair, read closely.
An illustrative pair from an earlier run of this aviation-to-pipeline
test — not an audited EXP-048c result. The reports describe different
physical events in different terminology. What links them is a
higher-level analogy in how each event unfolded.
Aviation · Weather
NASA ASRS · ACN 2086578
Encountered unforcasted [sic] severe turbulence at FL 250. On the
flight from ZZZ1 to ZZZ2, there were multiple reports of turbulence
along our route of flight at various altitudes. … Due to the
aircraft weight and other pilot reports, the Captain and I elected
to descend to FL250. … Nine minutes after level off, the aircraft
abruptly experienced large pitch and attitude changes resulting in
the immediate disengagement of the autopilot. … it was a continuous
battle to keep the wings level and keep the aircraft from over
speeding.
Two incident reports from different industries, filed with different regulators: a NASA ASRS aviation report and a PHMSA pipeline incident.
02
Different terminology
Turbulence, pitch and autopilot on one side; valve, ice and crack on the other. Each report is written in its own domain's terms.
03
Why similarity is not the target
Keyword search looks for shared terms; text embeddings look for similar meaning. These reports are about different topics. The link between them is a higher-level analogy in how each event unfolded — the kind of cross-domain correspondence EXP-048c measures.
04
An analogy, not the same physics
Turbulence and freezing water are different physical causes. Side by side, the two reports share a shape, set out below. It illustrates the idea; it does not describe how the engine scored this pair.
The higher-level analogy · illustrative
A hazard is already in play
AVturbulence reported along the route at several altitudes; the crew descends to FL250 as a precaution
PLwater held inside the body of the valve
conditions then intensify
AVnine minutes after level-off, severe turbulence that had not been forecast
PLtemperatures drop and the water expands into ice
and an abrupt physical consequence follows
AVlarge pitch and attitude changes; the autopilot disengages
PLthe expansion leaves a hairline crack in the valve
Different physics, different terminology, a similar shape.
That shape is the analogy — offered to illustrate the
cross-domain correspondence EXP-048c tests, not as an audited result.
Architecture
The company, the kernel, and the machine it powers.
Fractal Neurons is the company: it holds the research and the
discovery engine, and takes defined economics in the assets a
partnership generates. Space Neurons is the causal memory kernel. A
Dedicated Research Instance is the private machine it powers, with
DRI Apps installed inside it.
Your organization · the boundary
Access is controlled by your organization; keys are held under mutual
key custody and granted to approved roles.
Dedicated Research Instance · the private machine
Team instances
Agents spawned inside the instance, each with its own memory.
Instances nest, so the shape repeats from one lab to a portfolio;
the network of instances is dri.network.
Spine & TaskD
route each task to a scoped worker agent; one agent builds, another judges
InDeLoop
brings decisions that need a person to an approver and keeps a signed record of each
Semantic Radio
opt-in: shares abstract cause patterns with other instances, never raw text
Space Neurons · the causal memory kernel
A geometry-based retriever over a semantic vector space, matching
records on the abstract causal mechanism they describe.
Per-agent memory grows with each agent's work
⟷↕
Governed shared knowledge static and read by every agent; new knowledge is promoted deliberately — personal, then team, then shared — never leaked upward by default
Experience links connect the two and strengthen with use. Ownership,
provenance and permissions travel with every piece of knowledge.
People and oversight
Approvers see and decide what InDeLoop puts in front of them.
Configured outbound only
→↓ leaves the boundary
Model gateway with policy gates
bounded, approved calls to Claude, GPT and Gemini
→↓ leaves the boundary
dri.network
opt-in Semantic Radio exchange with other instances: abstract patterns only
A Dedicated Research Instance inside your organization: team instances run
agents that keep their own memory, Spine & TaskD route work to them,
InDeLoop brings decisions to people, and the Space Neurons kernel links
per-agent memory to governed shared knowledge. Only policy-gated model calls
and opt-in abstract-pattern sharing leave the boundary.
Fleet scale
Designed for fleets of thousands of agents.
Capacity depends on workload, models and deployment.
Organization
sets the boundary and names the people who decide
Team instances
each scoped to one group, dataset or question
Scoped worker agents
take the tasks Spine & TaskD delegate, each with durable memory
Human decisions
InDeLoop brings what needs a person to an approver
The agentic layer · associative memory
Discovery is a property of the memory — not a query
you run against it.
The substrate that surfaced those cross-domain pairs is the same
memory we give to agent fleets. Agents store the experience they live,
recall it by situation, and link experience to knowledge — and those
links get stronger the more they earn their keep.
01
Store what it experiences
Decisions, findings, corrections, and outcomes are written to a per-agent store as they happen — passed through a novelty gate so repeated experience merges instead of piling up.
02
Recall by situation
When an agent meets a situation it has been in before, the memory recognizes it and brings back what was actually used last time — personal experience layered over shared, immutable knowledge.
03
Links that strengthen with use
Every time a recalled memory is used, the link that surfaced it is reinforced; links that surface and get ignored weaken and fall away. Memory that improves with experience — never by overwriting what is shared.
Illustrative example · example values, not experiment outputs · not a customer deployment
In practice: one superseded fact, end to end.
01 · Input
A supplier revises a reagent's incubation time from 30 to 45 minutes.
02 · Agent work
Spine & TaskD hand the notice to a scoped worker agent, which finds what relied on 30 minutes: three analysis notes and two run records.
03 · Evidence
The memory shows each dependency with its source.
04 · Human choice
InDeLoop asks the assay lead to choose: repoint the notes, repoint them and re-run both runs, or reject.
05 · Updated outcome
The three notes now cite 45 minutes. The two run records keep the 30 minutes they actually used, flagged for re-check. A signed decision record shows who decided, from which evidence.
InDeLoop · decision neededillustration
Repoint the three analysis notes from 30 min to the supplier's revised
45 min? The two run records keep the 30 min they used.
Evidence
Supplier notice (revised) · 3 dependent notes · 2 run records at 30 min
Asked of
Assay lead
Repoint the notes chosen in this example
Repoint the notes and re-run both runs
Reject: keep the notes at 30 min
Resulting decision record
Approved by the assay lead · 3 notes repointed to 45 min · 2 run
records unchanged (30 min, as run) and flagged for re-check ·
evidence attached · signed on approval
Mock-up of an InDeLoop card for this example — not a real decision
record. No real signature, hash or timestamp is shown.
Static retrieval
3 notes →30 min · superseded
2 run records →30 min · as run
Current answers keep citing 30 min.
Associative memory
3 notes →45 min · current
2 run records →30 min · as run, flagged
Dependent notes repointed; run history kept.
Memory that repairs itself. Illustration of the
example above. Measured: in EXP-009, 78% of stale facts were
repointed (0.778, rounded) against 0% (0.000) for static retrieval —
an internal capability demo, not a customer-data or competitor
comparison. Source: Fractal Neurons company summary (Oct 2026),
supplied ;
company-reported; experiment date not supplied
(source note).
Less stale work — current answers stop citing a replaced value.
Visible ownership — the person who decides is named on the decision.
Traceable decisions — every change links back to its evidence, and past runs stay as they happened.
Point that same associative memory at a body of knowledge instead of an
agent's task history, and the links that make recall better become the
cross-domain correspondences that shared-term search does not target.
That is
where the aviation ↔ pipeline pairs came from.
The deployment model · Dedicated Research Instance
A dedicated instance that runs inside your own walls.
We work with each partner inside a Dedicated Research Instance: a
confidential, per-partner environment where their permitted internal
material sits alongside selected external knowledge. Access is
controlled by the organization, outbound traffic is limited to
configured channels, and source boundaries are preserved at every
step.
Instances nest on dri.network, so a governing instance can oversee
sub-instances, and agents onboard as native Google Workspace
identities inside your Drive access boundaries.
Compute runs in your organization's private instance; confidential data is never pooled or used to train a shared model.
Keys and access are governed by your organization under mutual key custody, granted to approved roles.
Outbound traffic is configured, not open: opt-in Semantic Radio shares abstract cause patterns, never raw text, and model calls are bounded and approved.
External models — Claude, GPT, Gemini — are reached through a gateway with policy gates.
Who we partner with
Experts with unique data and high discovery value.
We work with a handful of partners in confidential scientific, legal,
and operational domains.
If you hold unique data with high discovery value, we'd like to hear
what you're working on.
Confidential details only after a secure route is agreed.
Analytics cookies.
With your consent, this site uses Google Analytics 4 to count visits. It
sets cookies only if you accept; the site works the same either way.
Privacy notice
This is a preview copy: analytics never loads here, whatever you choose.
