The words on this site, defined.

Finding records that don't look like the rest. Powerful when you can't label what you're hunting for, and treacherous for the same reason: the model tells you something is unusual, never that it's wrong, and in most real data the unusual rows are a new supplier or a changed form rather than a problem.

The breakdown of which inputs pushed a single prediction up or down, and by how much. Distinct from feature importance, which describes the model overall: attribution is per-row, and it's what lets somebody argue with a specific answer rather than accept or reject the whole model.

Used here: Explainability in practice →

People leaving — an organization, a program, a customer base. The modeling trap is that leavers are labeled only after they go, so a model trained on last year learns last year’s reasons. Useful attrition work predicts early enough that somebody can still act, which is a much harder problem than predicting accurately at the exit interview.

Area under the ROC curve: the chance a model scores a randomly chosen positive case above a randomly chosen negative one. It measures ranking, not correctness — a model with excellent AUC can still be badly calibrated and produce probabilities no one should budget against.

The formal decision that a system may run on a government network, granted after a security review. It's a schedule item, not a formality, and a delivery plan that doesn't name who sponsors it has a hole in it.

Position broadcasts from ships, public and continuous. It supports real maritime analysis and has a known blind spot: a vessel can switch its transponder off, so absence in AIS isn't absence at sea.

Used here: Government & defense →

How often the thing happens at all, before any model looks at it. A model that predicts a 2%-prevalence event can be 98% accurate by predicting "no" every time, which is why accuracy on rare events is close to meaningless and why we quote precision and recall instead.

An approach that starts from what you already believed, weighs the new evidence, and returns an updated belief with its uncertainty attached. Its practical value in our work is that it answers the question people ask — how likely is this, given what we know — rather than the one classical tests answer.

Whether a stated probability means what it says: of everything the model called 70%, does roughly 70% happen? A model can rank cases perfectly and still be badly calibrated, and calibration is what matters the moment somebody uses the number to decide how much to spend.

When you know an event has not happened yet but not when it will. Every asset still running at the end of the study is censored, and dropping those rows because they have no failure date is the single most common way remaining-life estimates come out far too pessimistic.

Used here: Remaining useful life →

A map that shades regions by a value. It inherits the boundaries it's drawn on, so shading a whole country for work done in one province overstates it — which is a live constraint on our own reach map, not a hypothetical.

Used here: Where the work has been →

When the outcome you care about is rare — failures, fraud, readmissions. It makes ordinary accuracy misleading, distorts training unless handled, and is the normal condition in almost every problem worth modeling.

Used here: Healthcare →

A recommender has nothing to go on because the user or the item is new. It's the defining problem of the field rather than an edge case: the customers you most want to convert are exactly the ones with no history.

Used here: A 99.35% empty matrix →

Recommending by similarity of behavior rather than by properties of the thing: people who did what you did also did this. It needs no description of the item at all, which is its strength, and it can't say anything about an item no one has touched yet, which is its weakness.

When the relationship a model learned stops holding because the world moved: new equipment, a changed process, a different population. The model doesn't announce it. Accuracy decays quietly, which is why monitoring is part of delivery rather than a later phase.

A range that expresses how much the estimate would move if you ran the study again. A point estimate published without one is a claim about precision that no one has checked, and it's the first thing we ask for when reviewing somebody else’s number.

The four-cell table of predicted against actual: hits, misses, false alarms and correct rejections. Worth insisting on, because a single headline accuracy figure is one number summarizing four, and the four rarely cost the same.

The government’s own record of how a contractor performed, written by the customer. It's the past-performance evidence that matters most in a federal evaluation, because unlike a case study the vendor didn't write it.

Repeatedly holding back part of the data, training on the rest and scoring on the held-back part. It exists to stop a model being graded on its own homework. Where rows are related — same site, same patient, same time window — the split has to respect that grouping or it quietly leaks and flatters the result.

Removing the fields that identify a person. It's weaker than it sounds: enough ordinary attributes together — ZIP code, date of birth, sex — re-identify most people, so removing names is the beginning of the work rather than the end of it.

Collapsing several records that describe the same thing into one. The mirror image of entity resolution and just as consequential: merge too eagerly and you understate exposure, too cautiously and you count the same asset twice.

Linking records by rules that must hold exactly — same identifier, same date of birth and ZIP code. Fast, auditable, and it misses everything with a typo. Usually the first pass before probabilistic and spatial methods handle the remainder.

Used here: Multiple sources, no shared key →

A model of a physical thing, kept current with data from the real one, used to test what would happen before doing it. The word is applied to everything from a live simulation to a dashboard, so it's worth asking which is meant before anyone signs anything.

Combining several models so their errors partly cancel. Reliably better than any single member, and reliably harder to explain — which is why we pair one with per-row attribution rather than asking anyone to accept the output on trust.

Deciding which records across different systems refer to the same real thing when nothing shares an identifier. The most consistently underestimated part of applied analytics: every downstream model inherits its errors, and neither a missed match nor a wrong one shows up in the output as an error.

Used here: Entity resolution at national scale →

Extract, transform, load — moving data out of source systems, reshaping it, and putting it somewhere it can be analyzed. A substantial part of most data projects.

Models whose outputs can be accounted for in terms a domain expert can check. In our work it's a contract deliverable — a prediction that can't be defended in review doesn't get acted on, however accurate it is.

The entity-matching engine we built for infrastructure records that share no key. It combines probabilistic linkage, spatial reasoning and deterministic rules, because none of the three is sufficient alone: geography resolves what string similarity cannot, and rules catch what both would guess at.

Used here: Multiple sources, no shared key →

A single measured input a model reads. Most of the effort in a real project goes into constructing features from raw records rather than into choosing an algorithm, which is the reverse of how projects are usually described.

A ranking of which inputs matter to a model overall. Often confused with attribution, which is per-prediction. Importance tells you what the model relies on in general; it won't tell you why this particular case was flagged.

The Government Accountability Office and the Congressional Research Service — two sources of rigorous, public and rarely-read reporting on federal programs. Several of our federal tools are built on them, which is why those tools contain nothing privileged.

The recorded answer a model is trained and scored against. It's a record of what somebody wrote down, not of what happened — and where the two differ, the model learns the recording process. Worth auditing before the model is.

Distance between two points on a sphere, given latitude and longitude. Doing it with flat geometry is fine across a city and wrong across a continent, and the error grows exactly where infrastructure analysis needs it not to.

A contract that sets terms without committing to an amount; work is ordered against it later. Winning one is permission to compete for task orders rather than an award of work, a distinction worth keeping straight when reading anyone’s past performance.

An index that makes "what is near this point" fast. Without one, matching two sets of locations means comparing every pair, which stops being possible somewhere around the low hundreds of thousands of records.

A model trained to predict text, which turns out to be enough to summarize, classify and draft. It has no notion of whether what it produces is true, so anywhere the answer must be checkable it belongs behind retrieval and citation rather than in front of them.

How much better a model does than choosing at random. Quoted because it survives class imbalance: on a rare event, accuracy flatters and lift does not.

The record of where a value came from and everything that happened to it on the way. It's what makes an answer defensible a year later, when the person who built the pipeline has moved on and somebody senior is asking why the number changed.

The engineering around a model in production: deployment, versioning, monitoring, retraining. It's the difference between a model that worked once and a model that still works, and it's the part most often left out of a proposal.

Answering a question by simulating it thousands of times and looking at the spread of outcomes. Useful precisely when the maths is intractable, and it returns a distribution rather than a single number, which is usually the more useful answer.

The North American Industry Classification System code identifying what a business does. Federal solicitations name one, and it decides both who is eligible and what size standard applies to them.

Used here: Our NAICS codes →

The US National Health and Nutrition Examination Survey — a public dataset of examinations, laboratory results and questionnaires. It underpins our health demonstrations, and because it's a survey of a population rather than a clinical cohort, anything built on it is a demonstration and not a diagnostic tool.

Used here: Healthcare →

A formal statement of what entities exist in a system and how they relate. It sounds academic and decides something concrete: whether a pumping station and a substation stay distinct through ingestion, or get flattened into one class and lose the distinction the operator cares about.

When a model learns the noise in its training data and so performs far better there than on anything new. The reason a number quoted without saying which data it was measured on tells you nothing.

The window a contract covers. Worth distinguishing from the period the data covers — our Alberta engagement analyzed records from 1975 to 2013, which says nothing about when we did the work.

Used here: Timeline →

Precision: of the cases flagged, how many were real. Recall: of the real cases, how many were flagged. They trade against each other, and which one matters is a business decision — a false alarm and a missed failure rarely cost the same.

The prime holds the contract with the government and carries the delivery risk; a subcontractor delivers through the prime. Worth stating which you were, because past performance means different things in each case.

Used here: Past performance →

Re-describing many correlated measurements as a few uncorrelated ones that carry most of the variation. Excellent for seeing structure, awkward afterwards: the new axes are combinations, so an explanation in terms of them isn't an explanation anyone outside the analysis can act on.

Matching records by scoring how much their fields agree and setting a threshold, rather than demanding exact equality. It handles typos and naming variants, and it produces a match probability you can tune — which means somebody has to choose where a maybe becomes a yes.

The modelled probability that a specific person or account does a specific thing. It's a ranking tool, and its usefulness collapses the moment somebody reads a propensity of 0.8 as a prediction rather than as eight times in ten.

Where a figure came from, stated so a reader can go and check it. On this site, a number without a traceable source doesn't get published.

How much service a piece of equipment has left before failure. Useful output is a window with an uncertainty attached rather than a date, because a maintenance planner needs to know how wrong the estimate could be to schedule around it.

Used here: Preventive maintenance →

Whether running the same analysis again gives the same answer. Harder than it sounds — an unset random seed, a library upgrade or a source table that has since been corrected is enough to break it — and it's the property that makes a result evidence rather than an anecdote.

Fetching relevant documents first and having a language model answer from those, with citations. It's the difference between a system that can be checked and one that cannot — but it inherits the quality of the retrieval, so a confident answer over the wrong three documents still reads as authoritative.

The declared shape of a dataset: what fields exist, of what type, and what is required. Most integration pain is two systems whose schemas disagree about something small, like whether a missing value is null, an empty string or zero.

A contract or portion of one reserved for a category of small business. It determines which vendors may compete at all, so it's usually the first thing a contracting officer establishes and the first thing a vendor checks.

Used here: Capabilities statement →

A method that assigns each input a contribution to one prediction, derived from cooperative game theory so the contributions add up to the prediction itself. What sits behind most per-row attribution you will see, including ours.

A table that's almost entirely empty — every buyer against every property, where almost no pair has an interaction. Recommenders live here, and the emptiness is the problem to solve rather than a data-quality defect to fix.

Used here: A 99.35% empty matrix →

Vehicle data recorded in motion — position, speed, acceleration, braking, often many times a second. It arrives enormous and unlabeled, so the work is turning a stream of physics into behavior somebody can name.

Used here: Driver behavior profiling →

The score above which a model says yes. Moving it trades false alarms against missed cases, and whoever sets it is deciding which of those two costs more — a decision that shouldn't be left at the default.

The two identifiers a federal buyer uses to confirm a vendor is real and registered: the Unique Entity ID from SAM.gov, and the Commercial and Government Entity code. Both of ours are published rather than offered on request.

Used here: Capabilities statement →

How much the answer could reasonably differ from the one given. Publishing an estimate without it transfers the risk quietly to whoever acts on it, and a planner told a part fails in 40 days needs to know whether that means 38 to 42 or 10 to 200.