Every term we use, defined.
Most vendor glossaries are marketing with a heading. This one is written to be quoted. If a term on our site is not here, it is a term we should not be using.
50 of 50 terms
A
Action framework
The declared rules by which a conclusion becomes an action: which interventions exist, what they cost, what they may affect, which need a human, and the limits and budgets within which anything may be automatic. Owned by the business.
CapabilitiesAutonomy ladder
Five levels of granted authority (inform, recommend, approve, auto-apply within limits, auto-apply with review), granted per action class and never globally.
Capabilities
B
Baseline dynamics
The part of state movement we model from how the business works: seasonality, schedule, throughput, decay, the effect of known events.
The engineLearn moreShow less
Available on day one, and carries most of the state.
Related: Learned residual
Belief, plausibility, uncertainty
Belief: how much evidence supports a conclusion. Plausibility: how much could support it if unknowns resolved favorably. Uncertainty: how much is genuinely undecided.
Evidence
C
Corpus provenance
The record of what data trained a model, and specifically which teacher produced each row.
EvaluationLearn moreShow less
A corpus dominated by a previous model’s own output is self-referential: it teaches the model what it already believes. Such corpora disqualify a promotion.
Related: Teacher · Promotion gate
Cost to act
The intervention’s price. A property of the intervention, so it is consistent wherever it is recommended.
DecisionsCost of waiting
What the entity’s exposure becomes if its state evolves unimpeded over the horizon.
DecisionsLearn moreShow less
Makes “do nothing” a priced option rather than a default.
Related: Cost to act
Counterfactual replay
Running an entity’s actual history again with an intervention applied at a chosen point, to compare the two trajectories.
DecisionsConflict
How much the sources contradict each other.
EvidenceLearn moreShow less
A high-conflict case is flagged as contentious rather than averaged into false confidence: the number most often thrown away, and the most useful.
Related: Evidence fusion
Conflict gate
The control that halts automatic action when sources of evidence disagree, regardless of the autonomy level granted.
CapabilitiesLearn moreShow less
An automated system that acts confidently on contradictory evidence is worse than one that does nothing, because it acts at scale.
Related: Autonomy ladder · Conflict
Cross-domain connection
Inside one tenant, domains connect through acyclic bridges and evidence transfer observers. Nothing connects between tenants.
CapabilitiesLearn moreShow less
Learned calibration does not transfer between domains today.
Related: Tenant · Domain
Cloud native
Elastic, horizontally scalable, independently upgradable, observable and deployable where you need it.
Platform
D
Domain
A world with its own entities, state dimensions, dynamics, interventions and horizon.
Core conceptsLearn moreShow less
Domains are defined as data, so a new world is a configuration change rather than a code change.
Related: Ontology · Reference world
Divergence
A failure mode in which a model’s output is degenerate (non-finite or pathological) rather than merely inaccurate.
The engineEvaluationLearn moreShow less
Treated as a distinct verdict, because the remedy is different from “get more data”.
Related: Promotion gate
Decision horizon
How far ahead the domain must be right, declared by the business.
EvaluationLearn moreShow less
The model is trained and judged at exactly these horizons. It is an input, not a tuning parameter.
Related: Per-horizon error · Persistence baseline
Digital twin
A live, decision-grade model of an operation: every entity as state, continuously corrected by real events, and branchable so forward paths, counterfactuals and policy changes can be run before they are committed to.
CapabilitiesDomain pack
The shipped definition of a domain, including its baseline dynamics, interventions and horizon.
PlatformDrift
Change over time in the relationship between a model’s predictions and reality, monitored per serving model.
PlatformEvaluation
E
Entity
The thing in your operation whose state we model: a store, a payment lane, a member account, a machine, a shipment, a service case.
Core conceptsLearn moreShow less
Entities are the unit of attention; metrics are properties of them.
Related: State · Domain
Elapsed time (dt)
How much time a tick represents.
Core conceptsLearn moreShow less
In production it is the observed elapsed time between an entity’s events. In the demo edition, with the built-in simulator, it is a fixed simulated step. This is why the engine handles irregular feeds correctly.
Related: Tick
Evidence fusion
Combining multiple sources of evidence on an entity rather than stacking them into a weighted score, so disagreement is preserved.
EvidenceEarly warning
Detecting that an entity is diverging from its expected behavior before any static threshold is crossed.
CapabilitiesLearn moreShow less
Useful early warning states the entity, the forward position, the direction, the time at risk, the evidence and the value at stake.
Related: Value at stake · Time at risk
Explainable AI
Conclusions derived from explicit state and evidence, so every inference can be traced, reproduced and audited. Not a language model and not token-based.
CapabilitiesLearn moreShow less
It never means the absence of a learned component: there is one, and it is bounded, weighted and visible.
Related: Learned component
Edition
The capability set a deployment runs. The demo edition includes a built-in simulator and reference worlds; production editions run on real sources only.
PlatformLearn moreShow less
The simulator is a capability gate, not a toggle.
Related: Reference world
I
Intervention
An action available in a domain, carrying a currency cost and a modeled effect on the primary dimension.
DecisionsLearn moreShow less
Block a tender, review an account, schedule maintenance.
Related: Cost to act
Impact analysis
What an action changes (the expected effect on the primary dimension), what it costs, and what else it touches: the cascade across related entities.
CapabilitiesLearn moreShow less
In the product this surface is called Impact.
Related: Intervention
Inference framework
The declared rules by which state becomes a conclusion: ontology, thresholds, evidence weightings and the decision horizon. Owned by the business.
Capabilities
L
Learned residual
The correction a learned model applies to the baseline.
The engineLearn moreShow less
We ask the model to learn the difference from a known baseline, a smaller and more interpretable problem than learning state from scratch.
Related: Baseline dynamics · Residual weight
Learned component
The specialized, bounded model that produces the correction to the declared baseline.
The engineLearn moreShow less
It is not a language model, it is not token-based, and it generates no text. It outputs a numeric correction whose influence is weighted, visible and overridable.
Related: Learned residual · Explainable AI
Long-term memory
The governed, attributed record of entity history, decisions taken, the reasons, and the outcomes.
CapabilitiesLearn moreShow less
Distinct from the model: memory changes what the platform knows; learning changes what it predicts.
Related: Receipt
M
Model record
The per-model report: per-horizon error beside persistence, the furthest horizon that still wins, the verdict, corpus provenance and promotion history.
EvaluationLearn moreShow less
It always includes the horizons where the model does not win.
Related: Per-horizon error · Promotion gate
Multi-tenant
Many isolated customer organizations on shared infrastructure, with isolation enforced in the data layer rather than in application code.
PlatformMomentum frame
The five business movements executive surfaces are told in: Protect, Grow, Operate, Value, Allocate. Deliberately not risk language.
Platform
O
Ontology
The business’s own vocabulary: entities, state dimensions and their bounds, relationships, dynamics, interventions and policies.
CapabilitiesLearn moreShow less
Owned, versioned and reviewable by the business. It is the reason the platform works without historical data.
Related: Domain · State dimension
P
Persistence baseline (the null)
The prediction that the next state looks like the current state. The honest null to beat.
EvaluationLearn moreShow less
Beating your own previous model proves very little; beating “nothing changes” proves something.
Related: Per-horizon error
Per-horizon error
Forward error reported at each evaluated horizon, never averaged into a single number.
EvaluationLearn moreShow less
A model that is good at one tick and useless at 20 is reported as exactly that.
Related: Decision horizon
Promotion gate
The control between a trained model and serving. It refuses on, in order: corpus provenance, divergence, failure to beat the persistence baseline, and an unmeasured declared horizon.
EvaluationLearn moreShow less
A human may override, and the override is recorded.
Related: Corpus provenance · Divergence
Provenance
The trace from an alert or conclusion back to the rule, state dimensions and events that produced it.
Evidence
R
Reference world
A complete, runnable, fictional domain used for demonstration and evaluation, for example OgMart.
Core conceptsLearn moreShow less
Always labeled as fictional wherever it appears.
Related: Domain
Residual weight (α)
The configured amount of trust placed in the learned residual relative to the baseline.
The engineLearn moreShow less
A per-domain, inspectable number, not a hidden internal.
Related: Learned residual
Receipt
The record of what was decided about an alert, and by whom.
EvidenceLearn moreShow less
Makes “we looked and did nothing” a recorded decision rather than an absence.
Related: Provenance
Reasoner
Decisioning logic that answers one specific question about a proposed action and returns a reasoned result rather than a score.
CapabilitiesLearn moreShow less
Three ship and run on every decision: the safety envelope, action-consequence diffusion and guard arbitration. The product calls this surface Impact; the engineering term is deliberately not user-facing.
Related: Impact analysis
S
State
The position of an entity in the dimensions that matter for a decision. Not a status field, and not a score.
Core conceptsLearn moreShow less
State has a value per dimension, a direction of travel, and an accuracy that we measure.
Related: Entity · State dimension
State dimension
One axis of an entity’s state, for example payment integrity, capacity pressure or degradation.
Core conceptsLearn moreShow less
A domain declares its dimensions and their bounds. One dimension is designated primary: the one that carries business value and drives ranking.
Related: State · Ontology
T
Tick
One advancement of the whole world.
Core conceptsLearn moreShow less
On every tick each entity’s state is advanced and the model’s prediction is compared with what actually happened.
Related: Elapsed time
Teacher
The source that produced a training label: a model, a rule, or a real observed outcome. Attributed per row.
EvaluationTime at risk
How long an entity has been in an elevated band. Separates a spike from a slide, which a current value cannot.
DecisionsTenant
An isolated customer organization. Isolation is enforced by row-level security in the database, not by application filtering.
Platform
V
Value at stake
Exposure × probability × impact for an entity: the quantity used to rank the queue.
DecisionsLearn moreShow less
Chosen over “deviation from normal” because the business question is what it costs, not how unusual it is.
Related: Time at risk
W
What-if (forward)
Forking the current state and rolling it forward on the horizon, with the option to change the intervention.
DecisionsLearn moreShow less
Runs entirely in the browser, on the same engine as the server.
Related: Counterfactual replay
A term missing, or defined badly?
Tell us. A glossary that is wrong in one place is wrong in all of them.