Glossary

The rule by which a retrieval or question-answering system declines to answer, for example when the top-1 score is below a threshold chosen on development queries.

The nonlinearity applied to a unit's weighted sum, such as sigmoid, tanh, ReLU or GELU; without it, stacked layers collapse into one linear map.

An optimizer that scales each parameter's step by running averages of its gradient and squared gradient, with bias correction for the early steps.

A small trainable module or low-rank update attached to a frozen pretrained model so that a task can be learned without changing the base weights.

A measure of agreement between two labellings of the same rows that is 1 for identical partitions and about 0 for chance-level agreement.

A control loop in which a language model chooses the next action from the results of earlier tool calls, repeating until it signals completion or a limit stops it.

An effect plot that accumulates the average local change in prediction within narrow windows of a feature, using only combinations of values that occur in the data.

The share of tasks that succeed on every one of k repeated runs, reported next to per-attempt success because single-run success overstates the reliability users experience.

Embedding vectors occupy a narrow cone rather than spreading over the whole space, so even unrelated pairs have high cosine similarity.

A log of the assumptions a design relies on, each with evidence, confidence, owner, impact if wrong, a validation method and a revisit date.

The correlation of a series with itself at lag k, plotted for k = 1, 2, ...; bars outside the white-noise band hint at memory or seasonality.

A model that factorises the data into an ordered sequence and predicts each element from the earlier ones, trained by maximum likelihood.

The area under the precision-recall curve, computed as a precision-weighted average over recall steps; a constant score earns the prevalence.

Applying the inverse of a log or Box-Cox transform to a forecast on the transformed scale returns roughly the median of the distribution, not its mean.

The algorithm that computes the gradient of the loss with respect to every parameter in one backward pass, by applying the chain rule through the computational graph.

A text representation that records which words occur in a message and how often, ignoring their order; one column per vocabulary word.

Fitting many models on bootstrap resamples of the training rows and averaging their predictions, which reduces variance.

How common an event is in the population being scored before any evidence is seen, for example 0.2% of payments are fraud.

The strongest simple method a careful analyst would actually use, computed with only the information available at decision time. A model is worth adopting only if it beats the baseline on unseen data by more than its cost.

The leading tensor axis that indexes examples; it passes unchanged through every layer, and no parameter shape depends on it.

A layer that standardizes each feature over the current mini-batch and then applies a learned scale and shift, using running averages at inference.

Overlap between evaluation items and the text a model was trained, tuned or prompted on, which inflates scores without improving the capability being measured.

A resampling method that redraws whole blocks, such as weeks, instead of single rows so that dependence inside a block is kept when you compute an interval.

A lexical ranking function that scores a document by matching query terms, weighted by term rarity, with saturating term frequency and length normalisation.

The share of treated cases that must be true positives for an action to return its cost, equal to cost divided by (success probability x value kept).

The rule by which numpy and PyTorch combine arrays of different shapes by stretching size-1 or missing leading axes, so a bias vector is added to every row of a batch.

A partial, time-limited deployment of a change to a small share of traffic, evaluated against the rest as a control, with stop criteria agreed in advance.

Observed sales or bookings stop at a capacity or booking limit, so on constrained units they understate true demand.

A directed acyclic graph whose arrows encode assumed causal effects; it shows which paths create bias and what to adjust for.

The difference in an outcome between acting and not acting on the same units, for example how many more seats sell if fares drop 5%. It needs an experiment or explicit causal assumptions; a predictive model alone does not supply it.

A mask that sets the attention scores for future positions to minus infinity before the softmax, so each position reads only earlier ones.

Splitting source documents into the passages that are indexed, retrieved and placed in the context, ideally along the document's own structure with metadata attached.

The ordered set of evidence a comparison claim needs - correctness, fairness, noise, relevance and limits - where the claim is only as strong as its lowest unsupported rung.

How consistently the same groups are recovered when the data are resampled, the algorithm is restarted or the period changes.

A variable caused by two others on a path; conditioning on it (filtering or adjusting) creates a spurious association between its causes.

A scikit-learn component that applies different preprocessing steps to different column groups and concatenates the results, usually placed inside a Pipeline.

The network and loss written as a graph of simple operations (matrix product, add, ReLU, sigmoid, loss) whose local derivatives are combined by the chain rule.

A range computed by a procedure that, over repeated samples, contains the true parameter a stated share of the time (for example 95%).

A procedure that turns a point forecaster and held-out errors into intervals with finite-sample marginal coverage, assuming calibration and future errors are exchangeable.

A common cause of the treatment and the outcome, which creates an association between them even without a causal effect.

The maximum number of tokens, prompt and generated output together, that a model can attend to in one request.

Training an encoder so that an anchor is closer to its positive example than to negatives, using a softmax-style loss over similarities.

A small grid of weights slid over an input; at each position it multiplies the patch under it by the weights and sums, producing one number per position.

The dot product of two vectors divided by the product of their lengths; the cosine of the angle between them, in [-1, 1].

A table of the cost (or value) of each combination of predicted action and true outcome, from which a decision threshold is derived.

What each kind of mistake costs the business. Costs are often asymmetric - a denied boarding usually costs more than an empty seat - so the best decision is rarely the one that minimizes plain error.

Total spend divided by the number of answers that were correct and supported by the evidence, which includes the cost of requests that did not produce one.

The expected product of two variables' deviations from their means; positive when they tend to move together.

A data-quality test that compares the rows present per entity and period with the rows that should exist, so missing records are not mistaken for zero activity.

In Bayesian analysis, a range that contains the parameter with a stated posterior probability, given the prior and the data.

The negative log of the probability the model assigns to the true class, averaged over examples; binary cross-entropy is the two-class case.

Computing a learned transformation for each training row from the other folds only, so no row is encoded with information from its own label.

Repeatedly fitting on part of the data and scoring on the held-out part, so that every row is scored once by a model that did not train on it.

A trend component whose contribution flattens as the horizon grows, governed by a damping parameter between 0 and 1; long-run forecasts converge to a constant.

An explicit agreement between a data producer and its consumers on schema, grain, meaning, timestamps, freshness, update rules and ownership.

A change in the distribution of the model's inputs between training and production; it may or may not reduce performance.

Data leaving the system through a channel the model can use, such as a tool argument, an email recipient, a link or an image the interface fetches automatically.

Information in training or evaluation that would not be available at prediction time, making offline results look better than live ones.

The moment a prediction is made and acted on; only information that existed at that moment may be used as input.

A short document of one significant design choice with its context, the options considered, the decision, its consequences and the conditions for reconsidering it.

The moment a decision is made and acted on. Only information that exists at that moment may feed the prediction, the baseline or the rule that supports the decision.

A defined, tested reduced behaviour, such as showing cited passages without a generated answer, that a feature switches to when a dependency or budget fails.

A layer in which every unit sees every input, computed as H = activation(XW + b) with W of shape (inputs, units) and one bias per unit.

A pass or fail threshold on one constraint of a design (quality, latency, cost, security, reliability), agreed with its owner before candidate results are seen.

A fixed sequence of steps defined in code, in which a language model may be called at particular steps but never decides the order of the steps.

An effect estimate equal to the change in treated units minus the change in comparable untreated units over the same period.

A model trained to reverse a gradual noising of the data; generation starts from noise and denoises over many steps.

Multi-step forecasting with a separate model for each horizon, each trained to predict the value h steps ahead from features available at the cutoff.

An attack in which the user's own input tells the model to ignore its instructions, change its goal or reveal its prompt.

A preference-tuning method that trains the language model directly on chosen and rejected answers with a classification-style loss relative to a reference model, without a separate reward model or RL loop.

A regularizer that zeroes each unit at random with probability p during training and is switched off at evaluation.

Ending iterative training when the score on a validation set has not improved for a set number of rounds, keeping the best round.

How large a difference or relationship is, in units that matter (percentage points of conversion, EUR per booking), independent of sample size.

A dense, fixed-length vector that represents an item (word, document, airport, user) so that similar items lie close together under a chosen similarity.

A safeguard decided by code outside the model, so it still holds if the model fully complies with hostile text; prompt instructions are layers, not controls.

A subset of evaluation rows defined by something the decision depends on (segment, route, season, lead time, volume), with its own count, bias and error.

Computing a number for an unknown quantity from data - an average, a rate, a difference - together with how uncertain it is. "The IST-AYT no-show rate is 7.8%, 95% CI 7.1% to 8.5%" is an estimate.

A versioned, labelled collection of real or realistic cases, written and frozen before any prompt or weight change, against which every candidate system is scored.

A tag attached to a reported number that says what kind of evidence it is - measured retrieval-only, measured generated answers, assumed, or not assessed.

The share of a recorded item that counts toward a dimension estimate - base weight by kind multiplied by 0.5 to the power of age in days divided by 45.

The property that the joint distribution of a set of observations does not change when their order is shuffled; the assumption behind standard conformal coverage.

The probability-weighted average of a random variable, E[Y]; the long-run average over many repetitions.

A retry schedule whose waiting time grows exponentially between attempts and is randomised, so many clients do not retry in lockstep.

Forecasting by weighted averages of past observations whose weights decay geometrically, with separate smoothing for level, trend and season.

The mismatch between training on true history and generating from the model's own outputs, which lets early errors compound in multi-step forecasts or generation.

A small, fixed set of categories (data, model, pipeline, world) used to label each observed failure so that recurring root causes become visible.

Whether every claim in a generated answer is supported by the retrieved context; it measures support, not whether the context itself is true or current.

The output of one filter at every input position, showing how strongly its pattern appears where.

Including a handful of worked input-output examples in the prompt so that the model infers the format and label set without any weight update.

A weighted or simple average of forecasts from different methods, which often has lower error than the best single member when members' errors are not perfectly correlated.

The gap between the best simple baseline's error and the best error any model could achieve given irreducible noise; the most a heavier model can gain.

Adjusting forecasts made separately at different levels of a hierarchy so that parts add up to the totals, using information from every level.

The checklist of problem definition, information gathering, exploratory analysis, model fitting, forecasting and evaluation that frames a forecasting project.

The computation from input to output (and loss) through the layers in order, storing the intermediate values that the backward pass will need.

Layers whose weights are excluded from updates, so they keep their pretrained values and carry no gradients or optimizer states.

Points deliberately left out between the end of the training window and the start of the test block, at least as long as the longest feature window or label span.

A generator network trained against a discriminator that tries to separate real from generated samples; sampling takes one forward pass and there is no likelihood.

A model that learns the distribution of the data itself (or of the data given a condition) so that new samples can be drawn from it.

One model trained on many related series at once, using lags, calendar features and entity information, instead of one model per series.

A frozen, versioned set of questions with labelled evidence passages, expected answers, query-type tags and answerable flags, used to compare system versions.

When a measure becomes a target, optimizing it tends to break its link to the goal it was chosen to represent.

An additive model built stage by stage, each new small tree fitted to the negative gradient of the loss at the current predictions.

A test that compares analytic gradients with central-difference numerical gradients on a few examples, in double precision, to catch backward-pass bugs.

What one row of a table represents, for example one flight departure or one departure x booking snapshot.

An answer whose claims are supported by the evidence supplied to the model, as opposed to one that merely sounds plausible or happens to be correct from memory.

A cross-validation scheme in which all rows of one group (booking reference, customer, aircraft) fall in the same fold, so no group is on both sides of a split.

Generated content that is not supported by the provided context or by reality, usually fluent and confident in tone.

Error variance that changes with the fitted value or an input, so the residual plot fans out instead of forming an even band.

The vector a recurrent network carries from step to step; a learned, fixed-size, lossy summary of the sequence so far.

In high-dimensional nearest-neighbour search, a few points appear in a disproportionate number of other points' neighbour lists while many appear in none.

Searching over settings that are not learned from the training loss (depth, penalty, learning rate, stopping round) by scoring candidates on validation data.

The partial-dependence curve drawn separately for each row, showing heterogeneity and interactions that the averaged PDP hides.

A client-chosen identifier sent with every attempt of one intended action, so the service performs the action once and returns the stored result for repeats.

A data job that produces the same result whether it runs once or several times for the same input range.

A measure of how mixed the targets in a tree node are (variance for numbers, Gini or entropy for classes); splits are chosen to reduce it.

A tree model's built-in importance: the total split gain (impurity reduction) credited to a feature, computed on the training rows.

An attack in which instructions are planted in content the application retrieves or receives, such as a web page, email or document, so a third party attacks through the data.

Drawing conclusions about a population or a data-generating process from a sample, with a statement of uncertainty: which relationships exist, how strong they are, how sure you can be. In ML engineering, "inference" instead means running a trained model.

A score that adds a penalty for the number of fitted parameters to a model's in-sample negative log-likelihood; it ranks models of one class on the same data and does not measure forecast error.

A chart or slide title that states the finding with direction, size and scope, instead of naming the axes.

Supervised fine-tuning on examples that pair an instruction with a desired response, so a pretrained model follows instructions rather than only continuing text.

The state of a competency dimension with fewer than three distinct recorded items or too little recency-weighted weight; it is neither a score of zero nor a judgment of ability.

An assertion that must hold for every row if the data are correct, such as passengers <= capacity or a unique (date, flight_no) key.

The share of realised outcomes that fall inside a stated prediction interval; a calibrated 80% interval covers about 80% of outcomes in the long run.

Row multiplication that happens when the right side of a join has several rows per join key, so measures from the left side are repeated and their sums inflate.

A vocabulary for describing the allowed shape of a JSON document (types, required keys, enumerated values, extra keys) so that code can validate it.

A clustering algorithm that assigns each row to the nearest of k centres and moves the centres to the mean of their rows until assignments stop changing.

Stored attention keys and values of earlier tokens that let a model generate each new token without recomputing the whole context.

A change in the distribution of the target between training and deployment while the feature distribution within each class stays the same.

The period after the cutoff over which the target is measured, for example bookings until departure or churn within 30 days.

The value of the target (or a driver) from a fixed number of steps earlier, used as a model input; for a forecast h steps ahead the shortest legitimate lag is h.

A dense representation obtained by factorising a TF-IDF matrix with truncated SVD, so that words and messages that co-occur share a small number of latent dimensions.

Normalising each token's feature vector to zero mean and unit variance (with learned scale and shift), independent of the batch.

The scale factor on the gradient (or on the optimizer's step) that sets how far parameters move in one update; the most influential training setting.

A rule that changes the learning rate over training, typically a short warmup from a small value followed by a decay such as cosine.

Giving a component only the tools, data and permissions the current task needs, ideally the requesting user's own, so a compromised step reaches little.

A simple linear forecaster over recent history, trained across series, used to measure how much a more complex model adds beyond a linear map.

A simple linear or logistic model trained on frozen pretrained features; both a cheap baseline and a first step before fine-tuning.

A joint test that the first h autocorrelations of a series or of model residuals are zero, avoiding a bar-by-bar reading of the ACF.

A routing pattern in which a cheaper model answers first and a stronger model is called only when an acceptance check fails.

Using a language model to grade or compare other models' answers against a rubric, which scales evaluation but adds the judge's own biases.

The average negative log of the probability the model gave to the outcome that occurred; the objective logistic regression minimizes and a proper scoring rule for probabilities.

The natural logarithm of the odds p / (1 - p); the scale on which logistic regression is linear in its inputs.

A network's raw, unnormalized output score for a class, any real number; a sigmoid or softmax turns logits into probabilities.

A fine-tuning method that freezes a weight matrix W and trains two thin matrices B (d x r) and A (r x k) whose product is added to W, so only r x (d + k) numbers are learned per matrix.

The mean of absolute errors divided by the actual value, times 100; undefined when any actual is zero and unstable near zero.

Missingness that depends only on observed variables; conditioning or imputing on those variables removes the bias.

Mean absolute forecast error divided by the in-sample mean absolute error of a naive (or seasonal naive) forecast; below 1 beats that naive method.

Missingness unrelated to any observed or unobserved value; complete-case analysis stays unbiased and only loses precision.

The average over queries of 1 divided by the rank of the first relevant passage, counted as 0 when none appears within the cutoff.

A decomposition of the change in a metric between two versions or periods into additive steps by slice or driver that sum to the total change.

The specific thing a system fails to do, found by tagging its failures on an evaluation set (knowledge, format, behaviour, or cost and latency), which decides the adaptation lever.

A binary feature that records whether a value was missing before imputation, so a model can use informative missingness.

The process that decides which values are missing; MCAR, MAR and MNAR describe whether it depends on nothing, on observed data, or on the missing value itself.

Training with half-precision (fp16 or bf16) working tensors while keeping sensitive values such as master weights and reductions in float32.

Missingness that depends on the unobserved value itself, so no analysis of the observed rows alone can remove the bias.

An optimizer modification that keeps a running velocity of past gradients and steps along it, speeding up consistent directions and damping oscillation.

Strong correlation among inputs that makes individual coefficients unstable, even though predictions on similar data can remain stable.

Running many tests (segments, metrics, repeated looks) so that at least one false positive becomes likely even when nothing is real.

Presenting probabilities as counts out of one concrete population (190 of 200 fraud cases are flagged) instead of as conditional percentages.

A rank-sensitive score for graded relevance that discounts gains by position and divides by the best possible ordering, giving a value between 0 and 1.

Inflated evaluation scores caused by almost identical messages (resubmissions, forwards, templates) appearing on both sides of a train and test split.

Cross-validation with a second, inner loop that runs the whole tuning search inside each outer training set, so outer folds assess the procedure on rows the search never saw.

One weighted sum of the inputs plus a bias, usually followed by a nonlinearity; a small regression that a layer repeats many times side by side.

The planning quantile c_under / (c_under + c_over) that minimizes expected cost when planning too low and too high cost different amounts per unit.

The extra per-parameter tensors an optimizer keeps between steps, such as Adam's first and second moment estimates.

An error estimate for a bagged ensemble in which each row is predicted only by the trees whose bootstrap sample did not contain it.

Count data whose variance exceeds the mean, more spread than a Poisson model allows.

The probability, computed assuming the null hypothesis and the model are true, of data at least as extreme as those observed.

Comparing two models on the same rows and looking at the rows where only one is right (or at the per-row loss difference), which has less noise than comparing two independent accuracies.

An evaluation format in which a judge or a person is shown two answers to the same prompt and picks the better one, instead of scoring each answer alone.

Reusing the same weights at every position (convolution) or every time step (recurrent networks), so the parameter count does not grow with input length.

The correlation between a series and its lag k after removing the effect of the lags in between; it isolates the direct dependence at each lag.

The average model prediction as one feature is set to each value on a grid, with all other features left at their observed values.

The drop in a model's score on held-out data when one column's values are randomly shuffled; a measure of how much this model relies on that column.

A booking forecast equal to bookings on hand at the cutoff plus the average additional bookings ("pickup") that similar past departures gained from that point to departure.

An asymmetric absolute loss at quantile level q that weights under-forecasts by q and over-forecasts by 1 - q; its minimizer is the q-quantile.

Building each training row only from values as they existed at that row's cutoff, not as they look today after updates.

A join that attaches to each row the version of another table's data that existed at that row's decision time, and no later version.

A weight-free operation that summarises a small window (maximum or mean) to shrink a feature map and add local invariance to small shifts.

A judge's tendency to prefer an answer because of the slot it appears in (first or second) rather than its content, visible as verdicts that flip when the order is swapped.

Information about token position added to or combined with the input vectors, because self-attention alone is blind to order.

The share of the top k results that are relevant; its maximum is below 1 whenever a query has fewer than k relevant passages.

A best guess for an outcome you have not observed yet (or for a unit you have no label for), judged by how close it lands on new data. It says what will happen, not why or what would happen if you acted.

How far ahead of the cutoff the predicted outcome lies, for example 14 days from the D-14 snapshot to departure.

A range intended to contain a single future observation with a stated probability; it includes the observation's own noise and so does not shrink to zero width with more data.

Analytics that recommends an action: it combines a prediction or estimate with the costs of each error and the constraints, and outputs what to do - for example how many seats to oversell on each flight.

The first, large-scale training stage in which a language model learns to predict the next token over a very large text collection, before any task-specific adaptation.

The size of shaded areas in a chart should be proportional to the values they represent, which is why bar and area charts must start at zero.

Rescaling a classifier's odds by the ratio of the deployment class prior to the training prior, so probabilities match the base rate in use.

A one-page record of a production design decision, with context, decision, alternatives, labelled evidence, consequences, revisit triggers and an owner.

A provider feature that reuses the processed form of an identical prompt prefix, billing cached reads at a fraction of the input price and lowering time to first token.

Text placed in a model's input, directly by a user or indirectly through fetched content, that tries to make the model follow the attacker's instructions instead of the application's.

Treating the prompt text, schema, examples, retrieval settings and model identifier as one release that is scored on a labelled evaluation set before it replaces the current version.

LoRA training on top of a frozen base model stored in 4-bit precision, which cuts the memory needed to fine-tune large models while the adapters are trained in 16-bit.

The set of configurations that no other configuration beats on cost, latency and quality at once; constraints then pick one from this set.

Sampling choices that make each output more typical (lower temperature, truncation, stronger guidance) usually reduce how much of the real variety the set of outputs covers.

Storing or computing with weights in fewer bits (for example 8-bit or 4-bit instead of 16-bit) to save memory and bandwidth, at the cost of approximation error.

The three linear projections of each token's vector in attention - the query is matched against keys, and the matching weights mix the values.

The share of a query's labelled relevant passages that appear in the top k results; for RAG, k is the number of passages passed to the model.

The set of input positions that one output unit depends on; it grows with kernel size, depth, dilation and stride.

A way to merge several ranked lists by summing 1/(c + rank) for each document across lists, which ignores the retrievers' score scales.

A network that reads a sequence one step at a time, updating a fixed-size hidden state with the same weights at every step.

Multi-step forecasting with one one-step model whose predictions are fed back in as inputs for the next step, so errors compound with the horizon.

A versioned set of adversarial cases run against every prompt, model or tool change, reporting attack success next to normal task success.

A pass/fail rule on a frozen evaluation set, overall and per slice, that a release candidate must meet before it can receive traffic.

The set of fixed test cases run automatically on every prompt, model or data change to block releases that break behaviour that used to work, including safety and refusal behaviour.

A penalty on model complexity added to the training objective: ridge shrinks coefficients, lasso can set them to zero, elastic net blends both.

The difference between an observed value and the model's fitted or predicted value for that row; the part of the target the model did not explain.

A systematic structure in forecast errors (bias, spread growing with level, clustering in time, a jump at a date) that points to missing information or a changed process.

A pattern in which relevant passages are retrieved from a corpus at request time and supplied to a language model, so answers can use documents the model was not trained on.

A two-way count of labelled questions by whether the gold evidence reached the context and whether the answer was correct and supported, used to locate which pipeline stage fails.

The multiplication of attempts when several layers each retry the same failing call, turning a partial outage into a flood of requests.

A measurable condition - metric, threshold, window and owner - that reopens a recorded decision when it is met.

A post-training pipeline that fits a reward model to human pairwise preferences and then optimises the language model to maximise that reward while staying close to a reference model.

A backtest that moves the forecast origin forward in steps, refits on data before each origin and scores the forecasts for the following horizon.

The distribution of a statistic (a mean, a difference, a slope) over repeated samples from the same process.

The flat tails of an activation such as sigmoid or tanh, where the output barely moves with the input and the gradient passing through is close to zero.

A forecast equal to the observation from the same season in the previous cycle, such as the same weekday last week or the same month last year.

The variation in a metric when the same model and data are retrained with different random seeds; it measures training noise, not evaluation noise.

Distortion that arises when the units in the data were selected by a process related to the outcome or treatment (survivors, responders, sold-out flights).

Attention in which queries, keys and values all come from the same sequence, so every position can read every other position in one step.

Running a candidate model on live inputs beside the incumbent without letting its output affect any decision, to test plumbing, latency and agreement.

A Shapley-value-based attribution that splits the gap between one prediction and the baseline prediction across the features, additively.

For each row, how much closer it is to its own cluster than to the nearest other cluster, scaled to the range -1 to 1 and averaged over rows.

A trend that appears in pooled data reverses or disappears within every subgroup (or the other way round).

A generic narrative order for decision documents - agreed facts, what does not fit, the evidence-backed explanation, what it means, and the next step with an owner.

A dimension table that keeps one row per version of an entity's attributes, with valid-from and valid-to dates, so historical facts join to the attributes valid at the time.

A grid of small charts with the same encoding and a shared axis, one per segment, so the reader compares shapes across segments at a glance.

Synthetic Minority Over-sampling Technique: new minority rows are created by interpolating between a minority row and one of its nearest minority neighbours.

A table that records the state of something at given moments (for example cumulative bookings 30 days out); its values are states, not additive events.

A question rather than an objection that asks you to locate the weakest claim in your own answer, such as "which cutoff drives this?" or "what would change your mind?".

A function that maps K logits to K positive probabilities summing to 1 by exponentiating and normalizing; the output layer for one-of-K classification.

The standard deviation of a statistic's sampling distribution; for a mean of n independent values it is sigma / sqrt(n).

A series is (weakly) stationary when its mean, variance and autocorrelation do not depend on time; trend, level shifts and a fixed seasonal pattern all violate it.

The probability that a test detects an effect of a given size when that effect is real; one minus the Type II error rate.

A hard limit on the number of model-chosen steps (and, alongside it, tokens, cost or time) that one agent run may use before the harness stops it.

A loess-based method that splits a series into trend, seasonal and remainder components; robust options limit the influence of outliers.

A rule enforced by the harness that ends an agent run, such as a verified done state, an exhausted budget, a repeated call, consecutive tool errors or a need for human input.

A date after which the process that generates the series changes, for example a level shift after a competitor adds capacity, so earlier history describes a different regime.

Continuing to train a pretrained model on input-output example pairs so that it reproduces the demonstrated behaviour or format.

Replacing each category with a (usually smoothed) statistic of the target for that category, such as its no-show rate.

Training a sequence decoder by feeding it the true previous output at each step instead of its own prediction.

A positive number that divides the model's scores before the softmax; values below 1 sharpen the next-token distribution and values above 1 flatten it.

An array of numbers with a fixed shape; scalars, vectors and matrices are tensors with zero, one and two dimensions.

Using the same held-out block repeatedly to choose features, models or settings, which turns it into a validation set and makes its score optimistic.

A weighting of bag-of-words counts that multiplies how often a word occurs in a message by how rare the word is across messages, so distinctive words dominate.

The delay between sending a request and receiving the first output token; it includes queueing and processing the whole input, so it grows with prompt length.

A model pretrained on a large corpus of many series (for example Chronos, TimesFM, Moirai, TimeGPT) and used to forecast new series with little or no training on them.

A five-step routine for a timed case answer - clarify, structure, answer with units and uncertainty, take the challenge, revise - with the first answer kept beside the revision.

The unit of text a language model reads and writes: a word, a piece of a word or a short run of digits drawn from a fixed vocabulary, mapped to an integer id.

Splitting raw text into the units a model counts or embeds; word-level for classical NLP, subword units (byte-pair encoding and similar) for language models.

An explicit, per-task or per-route list of the tools an LLM application may call, enforced by the code that executes calls rather than by instructions in the prompt.

A decoding rule that samples only from the smallest set of most likely tokens whose probabilities add up to p, then renormalises.

A trace is the record of one request across components; a span is one timed step in it, carrying attributes such as model, tokens, status and release id.

Three roles for data - training rows fit parameters, validation rows choose between candidates, and a test block, used once, reports the result.

A generative model emitting near-copies of individual training records, which makes generated data a possible leak of the data it was trained on.

A difference between the feature values a model saw in training and those it receives in serving for the same underlying facts, caused by differing code, data or timing.

Starting from a model pretrained on a large task and reusing it, frozen or fine-tuned, for a smaller target task.

The number of configurations or trials (and the compute) a model was allowed to be tuned with; comparisons are fair only when it is comparable across models.

The line, enforced in code, between text the system treats as data (emails, web pages, retrieved passages, tool results) and the instructions and permissions the system itself defines.

Gradients that shrink toward zero as they are backpropagated through many layers, so early layers learn very slowly or not at all.

The gradient from early steps of a sequence is multiplied by a factor below 1 at every step, so it shrinks geometrically and early steps get almost no learning signal.

The expected squared deviation from the mean, Var(Y) = E[(Y - E[Y])^2]; its square root is the standard deviation.

An encoder-decoder model trained to maximise a lower bound on the data likelihood, giving a smooth latent space from which new samples can be decoded.

The failure that occurs when the words in a new message differ from the words the model was trained on, even though the meaning is the same.

Total absolute error divided by total actual volume, times 100; a volume-weighted ratio of sums that stays defined unless every actual is zero.

Shrinking weights toward zero by a small factor at each step; equal to an L2 penalty for plain SGD but different for adaptive optimizers.

The random scale at which weights start, chosen so that activations and gradients keep a stable size through the layers at step 0.

A series of uncorrelated values with constant mean and variance; the benchmark that model residuals should resemble if no linear structure is left.

A forecast from a pretrained model that is applied to a series it was not trained on, using only that series' history as context and no fitting on it.