The concept layer.

The degradation of a model's predictive relationship with its target over time, as the data-generating process moves away from the conditions the model was fitted under. Distinguished from data drift, where inputs shift but the underlying relationship does not.

The property that a dataset reflects only information genuinely available at each historical timestamp — including the values as first reported, before restatement, and the constituent sets as they stood, before survivorship filtering.

Pre-committed, quantified conditions under which an automated strategy or agent is halted without further discretionary judgment — specified before deployment and binding on the operator.

The verifiable record of what a model is: training data and its license terms, base model lineage, fine-tuning corpus, versioned weights or endpoint identity, and the transformations applied between raw data and deployed artefact.

The unbroken record of ownership and license rights running from every input used to build a model through to the model's outputs — the analogue of chain of title in real property, or clearance in music and film.

The persistence layer through which an autonomous agent carries state across invocations — working context, retrieved documents, prior decisions, and learned preferences — distinct from the model weights themselves.

The chain that turns a query into model context — chunking, embedding, indexing, retrieval, reranking and assembly — determining what evidence a model sees before it produces an answer.

Retrieval-augmented generation over a knowledge graph rather than a flat vector index, using typed entities and explicit relationships so that retrieval can traverse connections rather than only match similarity.

Formal specifications of financial concepts and their relationships — instruments, entities, events, exposures — with defined semantics that permit machine reasoning rather than only machine reading.

The set of controls under which a system may take consequential action without prior human authorization for each act — covering the authority grant, its bounds, the oversight interval, and the conditions of revocation.

Artificially generated data that preserves selected statistical properties of real data — used for augmenting sparse regimes, stress testing, and working around licensing or privacy constraints on the original.

The property that a stated result can be regenerated exactly from published inputs and procedure — same data as of the same moment, same code version, same parameters, same output.

An approach to model governance in which the intensity of validation and monitoring is set by a model's materiality — its exposure and its purpose — rather than by a fixed review calendar applied uniformly.

The design principle that an autonomous system operates inside an explicitly declared envelope — permitted actions, value limits, counterparties, time windows — beyond which it must halt or escalate rather than proceed.

The property that an autonomous agent is distinguishable as a specific actor — with a durable identifier, a recorded authority grant, and an attributable action history — rather than operating under a shared service credential.