Component class
The top-level division of the market. A class is a kind of thing you buy, with one unit it is ultimately bought in.
Classes are registry rows in witness.component_class, seeded from platform/witness/seed-data.mjs.
| id | Label | Unit | Wired |
|---|---|---|---|
model | Language models | One successful task | yes |
tooling | Agent tools | One tool call | yes |
retrieval | Embedding and rerank | One answered query | yes |
compute | GPU compute | One GPU-hour delivered | yes |
image | Image generation | One accepted image | yes |
extraction | Document extraction | One correctly extracted document | yes |
translation | Machine translation | One accepted translation | yes |
speech | Speech recognition and synthesis | One usable transcript; one accepted utterance | yes |
A class can hold two units
Most classes name one unit. speech names two, because it holds two things bought together and sold by overlapping sellers: a transcript, billed by the minute of audio, and an utterance, billed by the million characters of text.
What the two halves do not share is a quality story, and the class says so rather than leaving a reader to assume. Transcription quality is word error rate over audio the customer already has, which a laptop computes. Synthesis quality is a listening panel, which is people in a room. Both are measurable, which is what admits the class at all: a class whose quality cannot be measured has nothing to weigh a price against. Only one of them is cheap.
The unit is not the meter
The unit column is deliberately not what providers bill in. A language model is billed per million tokens; what you actually buy is a completed task. A GPU is billed by the hour; what you buy is an hour that delivered something.
The gap between those two is the whole point of the record. The unit names what you are buying. The dimension names what you are billed by.
A class a language model is eating is still a class
translation is the clearest case of the gap this record exists to price. Dedicated translators sell by the million characters. General models sell by the million tokens and do the same job. A buyer choosing between them is asking the stay, test or switch question in its plainest form, and almost nobody publishes the comparison.
The record holds both and refuses to convert between them. See dimension for the arithmetic that does not exist, and note that the absorption is visible inside the class as well as across it: two of its five products are large language models billing input and output separately, in the translator unit.
Image generation is the class where that gap is widest, and it is the class that made the gap visible in the data rather than only in this paragraph. What you buy is an image somebody accepts. What you are billed for is an image at a particular output size and a particular quality setting, and the same model can cost six tenths of a cent or twenty-one cents depending on nothing else. So an image sku carries the size and the quality its figure is for, and a seller who charges one price whatever comes back says that too, with a null and a sentence rather than by leaving the question unanswered.
Quality is what makes the class admissible at all, and it follows the same rule every other class follows: it is measured on the customer's own work or not at all. For an image that is the acceptance rate on the customer's own briefs, judged by the people who already review them. Nothing in this record scores a picture.
Wired
wired says whether Vitna has live sources for a class yet. An unwired class is registered, appears in /api/v1/markets, and answers its tape endpoint with an empty array and a 200.
That is deliberate. A 404 would say the class does not exist. An empty tape says the class exists and Vitna is not yet reading it, which is true.
