Reference

The terms this work defines.

Sixteen concepts the rest of the library relies on, each written to stand on its own and each linked to the piece that develops it.

Model to metal#model-to-metal: Owning a production system across every layer it actually spans, from the machine learning model at the top down through pipelines, protocols, drivers, firmware and silicon. The phrase exists because the hard failures are rarely inside one layer and almost always between two.
The seam#the-seam: The region between competent specialists that none of their maps include. A seam is not a gap in anyone's ability; it is the place responsibility was never assigned, because responsibilities are given out in terms of components and a relationship between components is not a component.
Load-bearing wrong#load-bearing-wrong: A wrong output that becomes the foundation for actions and decisions rather than being read and discarded. A system that is wrong costs you one answer; a system that is load-bearing and wrong costs you everything standing on it.
Computable residual#computable-residual: A number your code calculates deterministically, without asking the model, that shrinks as a task approaches done and has a floor it cannot pass. Rows unmigrated, tests still failing, records unmatched. Every field that ever built a reliable automatic stopping rule needed one.
The third state#the-third-state: The terminal state most agent systems are missing: gave up, as distinct from done and from killed. When a run exhausts its budget it still has to return something, and a system with only two names files it under the first one, which every downstream layer then believes.
Fuel gauge, not progress bar#fuel-gauge-not-progress-bar: An iteration cap measures how much budget has been spent, not how much work remains. It is accurate and honest, and like a fuel gauge it cannot distinguish arriving from running out, which is why a run that gave up and a run that finished come back identical.
Ghost state#ghost-state: State added purely to make a proof go through, which changes no behaviour if you delete it. Adding a counter to an infinite loop produces a genuine termination proof and an unchanged program, which is how a method that proves everything proves nothing.
Group split#group-split: Holding out entire units, every row belonging to a given speaker, site, device, tenant or time period, so none of that unit appears in training. A random split scatters the dominant source across both sides and is structurally incapable of detecting a group-level shortcut, however much test data you add.
The placement gap#the-placement-gap: The distance between the anatomical spot where a signal is strongest and the spot a device actually puts its sensor, measured as signal lost. It is decided in month two by people who do not work on signals, and it determines whether you are building a tracker or a model.
The return path#the-return-path: Where an injected current actually re-enters your circuit, and what is sitting on that node when it arrives. A ground symbol is not a sink; it is a node with real impedance, and the voltage that develops across it is the failure most closed-loop biosignal devices die of.
Desired state#desired-state: Expressing an action as make the world be in this configuration rather than change the world by this amount. Absolute writes are safe to repeat and increments double-count, which is how fleet systems made duplicate execution structurally impossible without ever discussing idempotency.
The abstain path#the-abstain-path: A state that is neither success nor failure, letting a step return no answer and escalate. Without it, a system's only vocabulary is answers, so every uncertain step becomes an action.
Intended versus executed#intended-versus-executed: Counting the tool calls a model asked for against the tool calls that actually ran, at every boundary between a model and code. One emitted and zero executed is an entire class of failure expressed as a pair of integers, and it is the check whose absence lets four other causes hide.
The grader and the graded#the-grader-and-the-graded: The rule that a check cannot be the same thing it is checking. A model grading its own output keeps the identical blind spots twice and shows measurable self-preference for length and for its own style, so the agreement it reports is one computation run twice.
The broken corner#the-broken-corner: Wrong answers do not spread evenly, they clump. A system is rarely wrong ten percent across the board and is usually badly wrong in one specific slice and fine everywhere else, so the fix is breaking one score into many rather than finding a better one.
Near perfect is not good#near-perfect-is-not-good: For a multi-step system, reliability multiplies, so ninety-five percent per step and ninety-nine point seven percent per step are two orders of magnitude apart on the error rate. Near perfect is not the neighbour of good; it is the neighbour of failure, and closing that distance is the actual engineering.