The Decision Model Glossary: Every Term This Site Uses, Defined

New categories mint vocabulary faster than anyone defines it, and week-one vocabulary hardens into jargon by month three. This page is the antidote: every recurring term across the directory and this blog, defined in plain language, with the deep-dive link where one exists. Alphabetical, opinionated where honesty requires it, and updated as the field invents more words.

Blocking. In entity resolution, the cheap candidate-generation step that shrinks "compare all pairs" to "compare plausible pairs" before any judge runs. The reason million-record dedup doesn't cost a million dollars.

Boundary clause. The parenthetical legislation inside a judge question that settles edge cases in advance ("cosmetic damage counts; disliking the color does not"). Where most accuracy actually comes from.

Calibrated probability. A probability output meant to be taken literally (0.9 should be right about 9 times in 10). Ecosystem descriptions of Jev's Choice API use this term; your own calibration checks (see LLM evals) verify it holds on your data.

Cascade. The architecture: cheap model rules on everything, uncertain cases escalate to expensive models or humans. The routing page owns it; the fraud build is its reference receipt.

Choice API. The native question shape ecosystem clients describe: you provide the question and its candidate answers; Jev returns a probability per choice (defaults YES / NO / UNCLEAR, editable). The probabilities are the confidence signal the cascade gates on.

Choice cap. The ~255-option ceiling per question reported by ecosystem tooling; larger answer sets are handled by staged or recursive questions (see jev-tree in the community).

Confidence gate. The threshold deciding whether a verdict stands or escalates. Set empirically per route, never guessed; the dial that trades cost against accuracy.

Decision model. A model built to answer constrained questions (labels, scores, yes/no, choices) fast and cheap, rather than to generate text. Jev is the current reference example; the comparison pages map the species.

Escalation. What happens below the confidence gate: routing to a bigger model or a human. A cascade without escalation policy is a log file with opinions.

Eval suite. The fixed test set plus graders that tells you whether today's system beats yesterday's. The evals guide builds one; per-commit is the point.

Grounding / faithfulness. Whether a generated answer's claims are supported by its provided sources. The central RAG-evaluation metric and the central guardrail in production.

Guardrail. A verdict on your own model's output before it ships: policy, grounding, tone, leakage. Sibling of moderation (user content) and injection screening (input).

Instruction bleed. The judge failure where text that mentions your categories gets judged as being them (an email quoting your spam policy flagged as spam). Test for it deliberately.

Judge question. A single closed-set question posed to a model about a piece of evidence. The atomic unit of everything on this site; the craft guide is its owner.

LLM as a judge. The pattern of models grading outputs against criteria; the flagship guide covers uses and biases. Decision models are the pattern productized.

Materialization. Running judgments once and storing them as columns, rather than re-judging per query. The database piece demonstrates; caching's respectable older sibling.

Nondeterminism. The same input occasionally drawing different verdicts. Designed around (logs, caches, gates), never denied; see limitations and production discipline.

Operational question. A question a new hire could answer consistently on day one: observable evidence, closed answers, edges legislated. The opposite of "is this good?"

Receipts. This site's house currency: author-reported numbers with sources linked, "as reported" always attached. The pricing page is the ledger.

Self-preference bias. Models grading their own family kindly. The reason judges, verifiers, and labelers should cross model families.

Straight-through processing. Automation completing without human touch, gated by confidence and policy ceilings; the invoice page shows the finance-grade version.

Triage. Multiple judgments per item driving routing: tickets, email, leads, alerts. Classification's employable form.

Verdict. One structured answer to one judge question. The unit everything above is denominated in, priced per the cost table.

Verification. Judging an agent's claims against evidence before downstream systems act on them; the agent-verification page owns it. "Trust, but judge."

Frequently asked questions

What's the difference between a verdict, a judgment, and a classification?

Interchangeable in casual use; on this site, verdict is the unit (one answer to one question), judgment the act, and classification the special case where the answer set is category labels.

Why does this site keep saying "as reported"?

Because every number here is author-reported with receipts rather than lab-audited, and the phrase keeps that honest; the review explains the epistemics.

Is a decision model just a small LLM?

It's a differently shaped one: closed answers only, no generative channel, optimized for verdict speed and price. The small-model comparison draws the line precisely.

A term is missing. Where does it go?

Submit it with the build or thread that coined it; this page is maintained, and vocabulary earns entry the same way builds do: by being used.