Zeno Divergent sits after your models, not instead of them. It reads what they predicted, what actually happened, and what was never observed — then learns the structure of the difference.
The native input is not a time series. Each step carries the observation, the predictions of every reference model that claimed to cover it, the context those models assumed, and a mask for what was simply not observed.
Xt,Rt={y^t(1),…,y^t(K)},Ct,Mt
•Upload CSV or JSON per domain in the workspace, or stream synthetic episodes.
•Missing values are preserved as missing, never imputed to zero.
•Fixed window of T = 48 steps per scoring call.
terms
try this stage
Load one episode and see the observation, every reference prediction and the observation mask side by side.
References are nodes with declared coverage, assumptions and validity conditions. Their disagreement is kept as structure — who disagrees with whom, about what, and under which assumed regime.
•Per-reference residual, calibration drift and coverage overlap.
•Disagreement is preserved rather than averaged into a consensus.
•A reference outside its declared validity is flagged, not silently trusted.
terms
try this stage
Inject a regime shift and watch which references leave their declared validity envelope.
A permutation-equivariant encoder reads all references jointly with the observation stream, so the representation depends on the pattern of disagreement rather than on reference ordering or count.
ht=Encθ(X≤t,R≤t,C≤t,M≤t)
•Entity-conditioned tokenisation across heterogeneous channels.
•Same weights regardless of how many references a domain supplies.
•Masking is an input, so blindness reaches the latent explicitly.
terms
try this stage
Change the initial state and see the representation move with the pattern of disagreement, not the ordering.
The model's central object is a low-dimensional latent trained to carry the structure of model-relative departure — not the outcome itself, and not predictive variance.
ztS∈R10
•Ten axes spanning magnitude, persistence, structure, mechanism and coverage.
•Trained under multiple objectives: departure reconstruction, regime-change contrast, validity prediction and an information bottleneck.
•The falsifiable claim lives here: this latent carries information about future regime change that predictive uncertainty alone does not.
terms
try this stage
Advance the clock and follow the surprise state as the departure builds.
The latent is decoded into distinct channels so the response can differ. Reference disagreement, structural departure and blindness are reported separately and never blended.
ztS→(Ut,Tt,It,At)
•U — unknown / blindness channel: how much of the reading is absence of observation.
•T — typed surprise: which kind of departure is occurring.
•I — information request: what should be observed next to resolve it.
•A — admission control: whether each reference keeps the right to speak.
terms
try this stage
Inject a sensor failure: the blindness channel rises while structural departure stays flat.
Rather than sampling futures at random, the model searches for unexpected trajectories that remain consistent with the current surprise state and with domain constraints — counter-expectations that are hard to dismiss.
•Candidate futures are scored on credibility and on distance from reference expectation.
•Constraint-violating trajectories are discarded, not down-weighted.
•Surviving tails become the scenarios shown in the workspace.
terms
try this stage
Explore the credible tails that survive the domain constraints at the current step.
A surprise in Earth Observation can imply exposure in Financial Services or Critical Infrastructure. Propagation converts into the receiving domain's own unit; there is no universal risk score.
•Explicit, inspectable coupling edges with declared mechanism.
•Propagated surprise is attributed back to its source and path.
•Receiving domains apply their own admission control to the incoming signal.
terms
try this stage
Propagate a departure into the receiving domain's own unit and follow the attributed path.
SurpriseBench runs the model against conventional detectors, degenerate controls and a label-shuffled copy on identical inputs, with episode-level bootstrap confidence intervals.
•Held-out synthetic episodes; AUPRC, AUROC and lead time at a fixed false-alert rate.
•If any baseline matches the model, the harness reports not supported.
•Every published number today is synthetic and labelled as such.
terms
try this stage
Compare the model against conventional detectors on identical inputs.
Every term used on this page, defined once. Hover any highlighted term in the text above to see the short reading inline.
reference graph
The set of existing models that claimed to cover a question, kept as nodes with declared coverage and validity rather than averaged into one consensus.Every forecaster, simulator or risk engine that made a prediction for the same step becomes a node carrying its own coverage, assumptions and validity conditions. Edges record where two references overlap. Zeno reads the pattern of agreement and disagreement across that graph instead of collapsing it into an ensemble mean, because the shape of who disagrees is the signal.
Departure measured against what the reference models expected, not against a historical average.The learned object is the difference between reality and the expectations of the models currently in production. The same observation can be unremarkable for one reference set and a severe departure for another; surprise is therefore always relative to a stated set of expectations.
A ten-dimensional learned state that encodes the structure of departure — magnitude, persistence, structure, mechanism and coverage.Written z^S_t, the latent is trained under several objectives at once: reconstructing the departure, contrasting regime change, predicting reference validity and an information bottleneck that stops it from memorising the outcome. It is the piece the falsifiable claim is about — that it carries information about future regime change which predictive uncertainty alone does not.
The latent is decoded into separate channels — disagreement, structural departure and blindness — that are never blended into one score.A single risk number cannot tell you whether to act or to look. Typed surprise splits the reading into channels so the response can differ: reference disagreement calls for arbitration between models, structural departure calls for a change of policy, and blindness calls for observation. Each channel carries its own magnitude, timing and consequence estimate in the receiving domain's unit.
How much of a low reading is genuine calm and how much is simply not observed.Missing data is treated as a first-class input rather than imputed away. The blindness channel reports the share of the current window that no reference could see — a cloud-covered tile, a halted instrument, a venue that stopped reporting. A quiet surprise score with high blindness is not reassurance; it is an instruction to restore visibility before trusting the reading.
A per-reference decision about whether a model still has the right to speak about the current situation.Each reference declares the conditions under which it is valid. Admission control tracks whether the observed regime still sits inside those conditions and marks a reference as admitted, provisional or withdrawn. Withdrawn references keep producing numbers — they just stop being counted as expectation, and the fact of withdrawal is itself reported as evidence that the world moved outside the modelled envelope.
Whether the reference set as a whole is still operating inside its assumptions.An aggregate state — valid, strained or invalid — derived from admission control across references, calibration drift and coverage. It answers the operator's question directly: is the thing I am relying on still right, and if not, since when.
An unexpected but constraint-consistent future found by search, not by random sampling.Credible tail search looks for trajectories that are far from reference expectation yet still satisfy domain constraints and the current surprise state. Trajectories that violate a hard constraint are discarded rather than down-weighted, so what survives is a small set of counter-expectations that are hard to dismiss on physical or institutional grounds.
The model's statement of what should be observed next to resolve the current ambiguity.Rather than only reporting a score, the model names the observation that would most reduce its own uncertainty about whether a departure is real — a re-tasked sensor, a second venue, an instrument check. It is the actionable half of the blindness channel.
Explicit edges that carry surprise from one domain into another in the receiving domain's own unit.A departure in Earth Observation may imply exposure in Financial Services or load risk in Critical Infrastructure. Coupling edges declare the mechanism, and propagated surprise is converted into the receiving domain's native unit and attributed back to its source and path. There is no universal risk score.
predictive uncertainty
A model's own stated confidence — the baseline the surprise latent must beat to be worth anything.Variance, entropy or conformal width reported by an existing model. It is the strongest conventional rival: if predictive uncertainty alone were enough to anticipate regime change and model failure, no separate surprise representation would be needed. SurpriseBench runs it as a baseline on identical inputs.
Average warning time before a break, measured only while the false-alert rate is held at a fixed budget.Lead time is trivial to inflate by alarming constantly, so it is only reported with the alarm threshold locked to a false-positive rate on clean steps (10% in the published harness). A method that cannot hold the budget reports no lead at all.
The trained checkpoint is exported to ONNX and executed locally in your browser by default, so uploaded files are scored without leaving the session. A hosted inference endpoint can take over when available, with automatic failover back to local execution.