Shape is signal.
Structural topology change precedes behavioural change. Before a system does something different, the shape of its relationships changes, and that change is measurable in the graph without knowing anything about the mechanism driving it.
This page is the depth behind SIET. If you are here to evaluate the detection product, the SIET page is the shorter road. If you are assessing whether the underlying claim is sound, this is the argument and the record.
One law, tested in places that share no subject matter.
A complex system approaching a transition behaves in a characteristic way. It becomes slower to recover from small perturbations, its fluctuations grow, and the structure of its internal connections reorganises. In ecology this is called critical slowing down, and it is well established.
The claim here is narrower and more useful: if you represent a system as a graph and watch the topology of that graph over time, those precursors are visible without a model of the system itself. You do not need to know why the market is fragile, what the attacker’s tooling is, or which instability mode a plasma will take. You need the shape, and the shape’s own history.
That is why the same mathematics appears in five unrelated domains below. Not five domain-specific heuristics that happen to resemble each other, but one method applied to five different graphs. The domains are the test: a framework that only worked on network telemetry would be a network heuristic, and a framework that works on tokamak plasma and sovereign trade flows and SMB traffic is describing something about structure.
It also cuts the other way, and it is worth stating plainly. Breadth is evidence that the principle generalises. It is not evidence that any single deployment is production-ready, and each domain below carries its own separate record with its own separate limits.
“I am not sure we can help you beyond what you have already achieved.”
Prof. Dirk Helbing, ETH Zurichprincipal architect of FuturICT
Intellectual property
GB2611542.8
UK patent pending, covering the application of structural early warning indicators to neural network training dynamics.
What is actually computed.
Graph topology metrics, combined with the early warning indicators established by Scheffer and others. Which of these carries the signal depends on the domain, and on what the graph represents.
Density
How much of the possible connectivity is realised
Entropy
How evenly connection is distributed across the graph
Fiedler eigenvalue
Algebraic connectivity, read as a fragility indicator
Velocity
The rate at which structure itself is changing
AR1 autocorrelation
Critical slowing down, after Scheffer
Variance ratio
Rising variance ahead of a transition
A note on SIET specifically
The spectral metrics above are what the planetary risk, market and neural domains compute on their graphs. SIET shares the premise and the graph representation, but its detector works on per-attribute baselines: the radius an attribute reaches, the number of distinct neighbours it has, and the distribution of who it usually talks to. It is the same idea applied at a different granularity, and describing it as spectral analysis would be inaccurate.
Five domains, five independent validations.
Each of these was validated separately, against real data, in a domain whose practitioners would not recognise the others as related work.
Cybersecurity
SIET
CommercialThe structural claim
An attacker changes what talks to what before any signature exists to describe them. Attribute-level baselines catch the change itself.
Evidence
17 of 17 attack episodes above a single connection detected on CICIDS2017, with no signatures and no attack knowledge. Median time to detect under one minute, 3 cases on a benign control day.
Planetary risk
Everything Engine
ValidatedThe structural claim
Global interdependence is a graph. Before a crisis becomes observable, that graph shows critical slowing down: rising autocorrelation, rising variance, compressing algebraic connectivity.
Evidence
195 countries and 16 domains monitored. 21 of 25 catalogued crises detected blind across 37 years of replayed history, most with 6 to 12 weeks of lead time, including the 2008 financial crisis and COVID-19. All four misses published.
Nuclear fusion
Everything Nuclear
ValidatedThe structural claim
A plasma disruption is a phase transition. The structural precursors that precede one are the same class of indicator that precedes a market break or a network intrusion.
Evidence
Plasma disruption detection on MAST tokamak data. 100% recall on the tested disruption set, with a response window several times the one ITER is designed around.
AI training safety
Neural SI
Patent pendingThe structural claim
A neural network in training is a weight graph. Its spectral properties show critical slowing down before the model undergoes a capability phase transition.
Evidence
Autocorrelation and variance ratio computed on the Laplacian spectrum during live training runs, giving advance warning of grokking transitions. UK patent GB2611542.8 pending.
Financial markets
Market SI
ValidatedThe structural claim
Institutional accumulation is structural before it is directional. Correlation topology concentrates before price moves.
Evidence
Around 700 UK and US equities monitored continuously, with the correlation graph rebuilt and structural metrics recomputed every 15 seconds.
The limits, stated first.
A framework that claims to work everywhere invites scepticism, and it should. Here is what the evidence does not support.
Breadth is not depth
Five validated domains means the principle generalises. It does not mean any one deployment has been proven at production scale.
One dataset per domain, mostly
SIET is validated on a single public benchmark. Where a second dataset was tried, some components measurably failed and that is recorded rather than omitted.
Lead time is not a promise
Historical replay establishes that a signal was present before the event. It does not guarantee the same lead time on the next one.
Misses are published
Four of 25 crises were missed in the planetary risk replay, and the two smallest events in the SIET benchmark were missed. Both are stated on their own pages.
The commercial application is SIET.
The framework is the reason SIET works without rules. If you want to see it operating on real telemetry, that is the product to look at.