Skip to content

Areas

Five signals, each with a boundary.

Every area says what it covers, what a reader can learn, and what it will not decide. The studies sit at the end of the page.

SA–11

Models

Covers what a language model is trained to do: predict a continuation from patterns. A reader learns to separate stored weights from a library of facts. The method is to name the task, the output, and the missing check. Concepts: parameters, prompt, output, hallucination.

SA–12

Data

Covers corpora, filters, dates, and languages. A reader learns that a dataset is a selection. The method asks who collected it and what was excluded. Concepts: corpus, cutoff, bias, evaluation set.

SA–13

Workflow

Covers where an assistant can sit in reading: drafting questions, summarizing a text you already have, listing terms. A reader learns that confirmation still belongs to the source. Concepts: draft, source, citation, review.

SA–14

Limits

Covers fluent error, outdated knowledge, and invented detail. A reader learns that tone is not evidence. Concepts: confidence, omission, contradiction, uncertainty.

SA–15

Claims

Covers sentences that put AI next to markets: “the model says,” “AI will,” “the data shows.” A reader learns to ask which model, which data, and which claim is actually in the source. This area does not rate assets.

How to read a model card

A model card is a public note, when one exists. It is not a rating. Read it in this order:

  1. What task the model was built for.
  2. What data the note admits, including the cutoff.
  3. What the authors say it should not be used for.

Continue with the studies

A waveform is not a forecast.

The curves on this site mark a sequence: prompt, pattern, sentence. They are not prices, not returns, and not a chart of any market. A rising stroke here only means “the next step in the explanation.”