Inference
31 terms described. A term is a word of the business with its definition, its aliases and the words it relates to.
| Term | aliases | First line | Links |
|---|---|---|---|
| Operation | endpoint, operation note | One callable unit of an API: a method and a path in an OpenAPI contract, a port and an operation name in a WSDL. The build imports every operation of a declared contract as an endpoint entity with the properties the contract states. An author who has business to explain writes an operation note; the build attaches it to the imported operation on its declared identifier, then on its method and path, then on its title, and the note becomes the single page of the operation, showing its own markdown, the properties of the contract and its own links. | 82 |
| Provenance | evidence | The evidence behind a link: the method, the file, the line, the section, the words as written and the context around them. Every link keeps every provenance it received; none is discarded. Provenance is what lets a reader check why the tool claims two entities are related. | 64 |
| Publication threshold | threshold, keyword page threshold | The count a candidate expression must reach to get a keyword page: three occurrences in at least two files, inference.keyword_pages in the configuration. Below the threshold the expression stays in the search index but has no page. The discovery has thresholds of its own, inference.ngrams, which decide which expressions are kept at all; the publication threshold decides which of those become pages, and the build summary reports the pages generated and the expressions under the threshold. | 96 |
| Query | model query | A reading of the model from the command line, without the site: concordance query resolves an expression the way the recognition reads a note, then prints what the model knows about it, where it is used, what it is linked to and the decisions that touched it, or answers a question of the whole corpus, in a compact text made for a person at a terminal or for the context of an agent, every fact with its file and line. The answer is as fresh as the last build, and says so. | 72 |
| Recognised word | recognised mention, marked word | A word of the text of a note that the occurrence scan matched against the dictionary, a title or an alias of an entity, or a recurring expression that the discovery gave a keyword page, read in a paragraph, a list item, a table cell or a quote. On the entity page the build turns it into a link to the page it names, drawn so that it is told apart from a link the author wrote: a dotted underline in the accent when the entity has a note, grey dashes when the expression only has a keyword page and no note; a legend under the text says which is which, and every mark tells on hover which note it leads to or how many passages the expression has. A recognised word carries less weight than a written link: it is what the tool found, not what someone asserted, which is why the three are never drawn alike. A page is marked once per note, on the first occurrence of its word; the later occurrences stay plain text, and the counts of the model do not change. | 64 |
| Relation | relation type, relation name | The name a link carries, taken from the profile: accesses, constrains, serves, related. Each relation lists the pairs of types it may join and a label for each direction, so that a screen reads "accesses" an object and the object reads "is accessed by" the screen. A relation is decided on four rungs, the first that applies winning: a mapped section, a typed frontmatter attribute, a type pair that admits a single relation, then related, the fallback capped at 0.60 and reported as ambiguous when a written link or a located mention carries it, never when co-occurrence alone does. A relation the profile does not allow between two types never enters the model. | 91 |
| Section mention | mapped section mention | A recognised word read in a section whose heading the profile maps to a relation: an object named under ## Objects of a screen is accessed by it, a rule under ## Rules constrains it, a note under ## Affects of a decision is affected by it. The section decides the relation, so the link is typed on the first rung and carries the section_mention confidence of 0.70, above a word read in prose. Headings are matched without regard to case or accents, in the language of the source. | 17 |
| Source | origin of a link | The file, line and method from which a link was produced: one element of a link's provenance. A link can have several sources when several methods found it. | 176 |
| Stopword | stop word, stopwords | A word the recognition sets aside: it never enters the dictionary as a title or an alias, a phrase made only of stopwords is never a term, and a candidate expression never starts or ends with one. The defaults come from the language pack of the source, stopwords.txt; inference.stopwords adds files of the project, one word per line, # opening a comment. Stopwords are compared in the same comparison form as terms. A default stopword names nothing on its own in any corpus: an article, a pronoun, a form of an auxiliary or light verb, an adverb of prose, a numeral, a noun that only counts or places; a word that could be a business term somewhere, or that ends a technical compound, is left to the project's own files. | 65 |
| Term | glossary term | A note of a glossary source: the H1 is the term, the first paragraph its definition, aliases its other names and broader the term it specialises. Every title and alias of a term enters the dictionary first, so that its occurrences in every other note link to it; a term never cited yields W-TERM-UNUSED, and a form it shares with another entity makes them homonyms. A term is short and exact; a ## Not to be confused with section settles a homonym. | 85 |
| Word confidence | keyword confidence, candidate confidence, suspected noise | A value between 0 and 1 that says how much a candidate expression looks like a term of the corpus rather than a word of its own prose. Counts alone do not tell them apart, an adverb and a business object both reaching eighty occurrences; the shape of the distribution does. The confidence is the product of five signals read from the corpus alone, in the same pass as the score, without learning: the spread, the share of the files holding the expression, penalised beyond one half; the burst, its occurrences per file, penalised under one and a half from five files; the position, a bonus for the appearances in a heading, a written link or the frontmatter; the neighbourhood, a bonus when a defined term accompanies it in a frequent expression; the morphology, a soft penalty when a word ends like an inflected form of the language pack, suffixes.txt. A plain term of the subject scores 0.72, a word met once in every file 0.36, a word present in every file of the corpus 0. The publication threshold reads it as inference.keyword_pages.min_confidence, one half by default: under it the expression is suspected noise, gets no keyword page and no mark in the text, stays in the search index and is listed, with its reason, on the to-do page of the site, where a call invites to add it to the project's stopwords. The model carries the confidence, its signals and its penalties on every candidate, and the confidence among the attributes of every keyword page. | 43 |
11 inference of 31 — pagination by twenty. The Links column counts the related pages, which brings the most central inference of the journey to the top.