The words this industry never defined

Four words do most of the work in a data vendor's pitch, and none of them have an agreed definition. Real time. Fresh. Verified. Coverage. Each is load-bearing enough that a purchase turns on it, and vague enough that two vendors using the same word can differ by three orders of magnitude without either of them being wrong. The glossary exists because of this paragraph.

Take the first. Real time, measured from when - the moment the event occurred, or the moment your system noticed it? Those differ by the polling interval, which is where the half-interval tax hides, and the distinction between event time and processing time is the whole ballgame. A vendor quoting 400 milliseconds from ingestion and a vendor quoting four minutes from occurrence may be describing the same pipeline.

Fresh is the same trick in a different tense. It usually means the dataset was updated recently, which says nothing about whether the world it describes moved in between. The precise version is a standing property with a number attached, which is why data freshness and a published watermark are worth more than an adjective: one tells you the point up to which the data is believed complete, and the other tells you a vendor updated something.

Verified is the word most worth interrogating, because it sounds like a guarantee and is usually a synonym for processed. Verified by what evidence, against what, and what happens when the check fails - is the record withheld, flagged, or scored? We split it deliberately into authenticity (is the event real and is the source who it claims to be) and data quality (is the record complete, deduplicated, and correctly resolved), because those fail independently and a single word hides which one you are getting.

Coverage is the one that quietly decides renewals. A vendor listing four thousand sources has told you about its supply, not about whether the specific things you care about arrive. The useful unit is subject areas and the events inside them, which is why our coverage is organized that way rather than by source: the sources behind a subject area change, and a client's integration should not have to.

The commercially convenient thing about all four words is that vagueness is not disprovable. A number can be checked; an adjective cannot. This is not a conspiracy, it is a gradient - every vendor is nudged toward the word that survives diligence rather than the number that might not, and the gradient has been running for a decade. The result is a market where the specification and the marketing have drifted so far apart that a buyer's first job is translation - and where a definition you can be held to is itself a differentiator, which is a strange thing to be able to say about an infrastructure industry.

Our response is a glossary written to be used against us. Sixty-odd terms, each defined the way a practitioner would define it rather than the way it would sell, with the awkward distinctions kept in: delivery guarantees that name at-least-once for what it is, point-in-time history described as an append-only obligation, tail latency as the number that matters rather than the median. If we later ship something that contradicts one of those pages, the contradiction is public and it is ours.

The test we hold each definition to is simple: could a reader look the term up elsewhere and find our meaning, rather than a meaning we invented for the occasion? That rules out coining vocabulary to describe ordinary things, and it rules out borrowing a technical term for a weaker claim. It does not rule out precision that costs a sale - ML-KEM and FIPS 203 stay on the post-quantum page, because the reader who checks those is the reader we want.

Every record we deliver carries its own provenance, which is the same principle applied to data rather than to language: a claim that arrives with its evidence attached can be checked by somebody who does not trust us yet. That is the standard, for the numbers and for the words.