Bayesian API reference

When it was disabled; a disabled account signs in no more.

Lowercased; unique.

a- and 16 hex digits.

The fingerprint of its current API key: what its sessions act as.

u- and 16 hex digits; unique: the account's tenant, which its keys and jobs carry.

ISO date (2023-03-15).

The choice point's id (genocide_definition).

One of its values (broad_war_crimes).

The value's prior.

The value fixed; for a marginalised point, its default.

Where the assumption enters the chain: rule ids, definition '…', indicator <id>, claim #n.

The HMAC-SHA256 key of X-Bayesian-Signature: sha256=<hex> over the raw body.

X-Bayesian-Delivery of every attempt.

The receiver's last status, or why no answer came.

Where the job is read: /api/v1/jobs/{id}.

The municipality's register code (RPE: 190 for Žalec).

Choice points and contexts carried.

What was not carried as it was, one line each.

Definitions carried, by predicate.

Items carried: (rule, id there, id here).

Gaps, sources, aliases and glosses carried (a rebase only).

Rules carried whole (they were not here).

Statuses set on items both hold: the item ids.

The condition: a rule's condition term, or an analysis's condition (its predicate).

The provision that sets it ("ZON, 104. člen, (1)").

In words: the term's gloss, or the condition's description.

How many levels of analysis below a claim (default 3).

The day the check is judged as of and why (the filing date, else the documentation's date).

A context fixing choice points (a named interpretation).

The domain profile: si-building for a permit check.

The caller's facts: the location card, plan acts and rules with their review states, provision texts.

Web research sessions; 0, the default (si-building's knowledge comes from the caller). The service's model backend has no web tools: a value above 0 is answered model_unavailable.

A label for the caller (a project id).

The language of the result's prose (sl, en); default the text's.

The caller's organisation: the job, its verdict and what it derives about the project are scoped to it.

A parse returned by an earlier job (result.parsed of the verbose form), checked again without a parse.

Reason from first principles: concepts analysed into their conditions, a derivation per claim.

The site, when the caller does not supply its card in facts.

The text to check: a DGD's relevant parts, a statement. One of text and parsed.

The value the default context assumes. Absent when proposing a point: the value with the highest prior.

Proposed (by the parser from a text's premise, or by operationalize for a contested predicate) and not yet accepted: listed, never used in a verdict. add_choice_point with the same id accepts it.

The default context fixes the point at its default value (a legal default that applies unless a context says otherwise) instead of marginalising over the values by prior.

What is being chosen, in one line.

snake_case id (genocide_definition); its term is choice:<id>.

Mutually exclusive and exhaustive values, each with a prior.

Where the prior comes from: "scholarly acceptance, SEP 2024", "legal default", "stated".

What the value means, in one line.

snake_case id, unique within the point (convention_intent).

Prior probability that this value is the right one (0 to 1). The priors of a point are normalised to sum to 1 when it is stored.

The values of known choice points (listed under "Known choice points") this claim is made under, when the text attaches one to it ("under the Convention's definition, …"). Omit otherwise. On ingest the claim's evidence counts only under these values; in evaluation they are fixed only in the text's own context.

For an attribution: term id of who said or believes it.

For a quantity in the object: whether it is a bound or an approximation rather than an exact value.

Context ids under which the statement is made, e.g. luxury_goods.

The day the text was read (ISO date), from [ParsedStatement::dated]: a claim without a time of its own is judged as of that day, so "the past 24 hours" in an article keeps its day when the verdict is judged again later.

Other claims of this text this one leans on.

Set only when the wording frames the referent in a misleading way.

The statement reports that no significant effect or difference was found ("no significant reduction in stroke"): absence of evidence of an effect, not evidence that there is none. Encode with positive=false.

Object or value of the predicate when it has one, e.g. sweet for has_property. Omit for intransitive predicates.

Where the claim holds: ISO country code (SI), CC-Name for a place inside a country (SI-Ljubljana), or a region (europe, eu, world). Resolved to a place term and added to conditions.

false when the statement denies the proposition ("mammals do not lay eggs").

snake_case predicate, e.g. lays_eggs, has_property, causes.

The claim is that a public act took place: something was published, posted, said or announced in public, a vote cast, a document signed or filed. The act's own record (the post, the transcript, the voting record) settles it, whatever the act's content says.

The hedge or quantity as written: "most", "rarely", "70%", "always".

The claim is about the text itself or what its author does in it: its structure and sections ("this paper has three sections", "section two addresses …"), its aim, scope or method ("I am concerned with …", "I do not examine …"). The text settles it, not outside sources, so it is not researched. Not for what the text says about the world: "in section two I show that X causes Y" is the claim that X causes Y, without this flag.

Term id of the subject in its resolved sense (an existing KB term id, or the id of a term proposed alongside the claim).

The span of the original text this claim came from.

The period the claim is about; none means the day the text was read (dated), else today (or timeless, for kinds).

Whether the claim's truth is independent of the date it is judged on (an interpretation, a definition, who first did something), so a newer source is another reading, not an update. Omitted: an attribution or evaluation is timeless, any other kind is not.

A number the claim states, with its unit (FZ 0.35 → 0.35 ratio; 9 m).

Under other readings: the verdict with one of those points at one of its values, each with its prior (not equally valid alternatives).

The choice points the verdict above rests on, anywhere in its chain: "genocide_definition = convention_intent (marginalised, prior 0.70)", with where each enters. The verdict above is the headline.

Conditions beyond the document setting.

For an operationalised evaluation: the definition used.

The domain the claim was evaluated under.

For an operationalised evaluation: each indicator's verdict.

Overall label; "misleading" / "possibly misleading" for framed claims.

How true the literal words are, when framing changed the overall label.

The claim in the request language when it is not English: the label, how it is shown, its flags and assumptions through the vocabulary, the subject's gloss and the predicate's reading through the render cache.

For a claimed value at a site: the norm in force it was compared with ([crate::norm::NormResult]), with the assumptions the verdict rests on.

The claim reports that no significant effect was found (absence of evidence).

For a statement about a kind: the conditions of the single case it leaves out, under which it was judged (a universal statement is false for leaving them out).

[belief, disbelief, uncertainty]

Fingerprint of the parameter set that produced the verdict.

Canonical predicate key as evaluated.

The claim is that a public act took place (a post published, a statement made, a vote cast), checkable from the act's own record.

The domain's root claim (si-building: that the permit may be issued), which a first-principles check adds: its conditions are its indicators, and it has no norm block. A client mapping claims to its own lines skips it by this.

Rules that decided the verdict.

The claim is about the text itself (its sections, what its author does in it): judged on no evidence, not researched, not a decided fact.

How to show the verdict to people: "verified" when the evidence is beyond reasonable doubt (with the basis), otherwise the label with a rounded probability. probability stays as computed.

Period the claim is about, e.g. "2026", "1970–"; absent means now.

For a statement about a kind: the exceptions for the single case it leaves out, presumed not to apply.

The context the claims were judged in when not the default ("legal", "text_own", "genocide_definition = broad_war_crimes").

Pairs of the text's claims that cannot both be right as judged (a blanket compliance statement against a value that likely fails the act's norm; a conclusion judged false from a premise that holds), also flagged on both claims ([crate::document::contradictions]).

First principles: one derivation per claim, at claim i's position ([crate::principles]). Empty for other passes.

With fill_gaps: what each fill did, "subject predicate: outcome" (including failures).

Whether the framing review ran and succeeded; without it, framings are the fast pass's guesses.

"subject predicate (reason)" for each gap still open.

The language the request is answered in (ISO 639-1): the one asked for, else the text's own. Labels and flags in local are in it.

What is normal but worth saying, unlike a warning (a corpus without structured rules: the site's norms are read from the plan's articles).

Premises the text states (not claims): the known choice point and value each fixes, or a proposed point.

What the evaluation could not settle and what would settle it: a site to locate, a norm for a factor at a site, a provision or record for a statement ([crate::facts::Question]). A client answers through facts (on evaluate_statement, investigate, supply_facts).

Context terms that scope the whole text; applied to every claim.

Claim count per label, e.g. {"likely true": 12, "insufficient evidence": 30}.

summary with the labels in the request language (not English only).

Coverage problems: parts of the text that could not be parsed, dropped stub claims.

How many framing reviews flagged it, e.g. "1/2".

Verdict on the standard characterisation, when it was encoded.

What it measures and why it bears on the evaluation.

The indicator's verdict rests on evidence about a body the subject belongs to, or inferred, only: the "collective evidence only" flag.

For a measured indicator: the sampled items behind its deciding rule (titles as published, the original first).

What decides it, as cited, and its words.

The condition, in the provision's words.

Whether it holds (none: the reader could not tell).

The norm rule whose condition it is.

The rule holding the decision.

The assumptions this alternative fixes: the context's, with the point in question at the value in question.

Probability minus the headline's.

The choice point varied and its value.

The value's prior: how much the alternative counts for.

How it is shown ("likely true, 0.82", "verified").

Their positions in the text's claims, from 0: the general claim or the premise first.

The pair in words, as both claims' flags give it.

The other claim's position in claims, from 0.

The claim's position in the text's claims.

Levels of analysis below the claim (0: not analysed).

The leaves settled by evidence, in tree order.

Concept-laden conditions left unanalysed, "subject predicate".

The claim denies its predicate ("projektni pogoji niso upravni akti"): the tree analyses the predicate, and the claim holds when it does not.

Leaves nothing decides yet, "subject predicate".

What the derivation presumes or assumes without evidence: each presumption of a provision it rests on (holds unless shown otherwise) and each condition a reader took as holding for want of a fact.

The derivation as indented text.

Draft analyses to review: a necessary condition the statute itself decides does not hold, so the analysis, not the claim, may be wrong.

The choice points the derivation rests on (which definition, which reading).

The sources the definition rests on.

"all conditions must hold", "at least 2 of 4 conditions must hold", …

"'is_fascist' v1 (draft, analysis)".

For an analysed node: the definition and composition used.

For an analysed node that rules on the claim itself also decide: what grounds them. The verdict fuses them with the analysis, so it can stand while the conditions are open.

The node's own flags worth a reader's eye: an undefined concept, a cycle, a draft definition.

For a condition: fact (directly checkable) or evaluation (a concept of its own).

For a condition: the provision it rests on, as the definition names it.

How the node is shown ("verified", "likely true, 0.89").

For a ground, or an analysed node's direct rules: the deciding evidence, "kind: source (date)".

For a condition: whether it counts toward the concept applying (true) or against it (a disqualifier, false).

The act's name inside a citation ("GZ-1", "Odlok o OPN Občine Žalec").

The corpus's key (gz-1, pa-1234): stable, used in term ids (act:<id>).

The municipality of a municipal act (RPE code); none for national acts.

The gazette line ("Uradni list RS, št. 64/13").

A municipal act's role: base (the plan), amendment, other (a detailed plan).

Oldest first.

The day the check is judged as of and why: the filing date, else the documentation's date (an assumption). Every claim about the site is judged as of it; later versions of an act are flagged, never decide.

The acts the norms come from, with their versions (validity, status). A norm whose document is not listed gets a document made from its locus (current version, valid from the norm's version's day unknown).

Luma's own answer to POST /v1/location/resolve, as it came: mapped to the site's card the way Bayesian's Luma adapter maps it (register units over derived ones, shares over the parcels' area, regimes with their basis). Used when site_card is not given.

Luma's answer to GET /v1/plan-acts?ob=<municipality>: the acts as documents with their versions, the base plan the zones belong to, and the acts' corpus keys for the card.

Luma's answer to GET /v1/plan-rules?ob=…&eup=…: the plan's rules, mapped to norms as the adapter maps them (paragraphs are not looked up: the locus names the article).

The plan's rules as structured norms: factor, operator, value, unit, condition, the requirement verbatim, its locus (act, version, article, paragraph, citation, link) and review state.

The organisation the project belongs to: what is derived from its facts (the reader's decisions on the site, the norm comparisons, the rules read from its records) holds under org:<id> only, and the check's claims carry org:<id>, so another organisation's check of the same parcel never sees it. Set from the check's org.

The procedure the check is for (act, instrument, documentation stage, where the caller has it from): a project fact for the reader that applies provisions to the project, an assumption when from a profile.

The site's location card: municipality, parcels, planning units (EUP, PEUP, land use, share), regimes, acts in force, as of a date, with its source. Taken instead of Bayesian's own resolver for the site it names.

Provision texts and records, ingested before the check (one model call each).

Who supplies them ("Luma Intelligence"), for the journal and the notes.

One line per thing stored (a card, the norms of an act, a source read).

One line: why the wording misleads.

How the referent is standardly characterised, in plain words ("the siege and sack of Baghdad by Hülegü's army, 1258").

The standard characterisation encoded as a claim, so the knowledge base can check it (subject mongols, predicate sacked, object baghdad).

Who made it: an account id, key:<fingerprint> or bayesian-admin.

The first 8 hex digits of the token's SHA-256.

When it was accepted, and the account it made.

What became of the callback the job asked for, once it was tried.

Why a failed job failed.

The job whose questions this one answers.

How many of result.questions are still open: the caller answers them with POST /api/v1/jobs/{id}/facts.

The compact evaluation: claims with labels, norm blocks with condition_checks, questions, notes, warnings, derivations when principles.

Seconds the job is expected to take, when known.

Web research sessions for the new job; 0, the default (above 0: model_unavailable, as for a check).

What the subject term means.

The predicate key in words.

The requirement as a reader of the language sees it: verbatim when the corpus writes in that language; for an English reader the English rendering; for anyone else the original (never a translation alone).

What is shown beside it: the original (English reader), the English (a reader of a third language), none (the corpus's language).

The status in a reviewer's words ("izpolnjuje le, če: (a) …").

The citation as a reader expects it ("2. odstavek 75. člena Odloka …"): the original in the corpus language, with its English rendering once made ("Article 75(2) of the … Decree").

The corpus it came from (luma, file).

The document's key (gz-1, pa-1234).

The domain profile the item was ingested under, whose reliabilities override corpus.* ([crate::domain]).

A row inside the unit: an annex row's unit code (SSe, ŽA-1/3).

The paragraph ("(2)" → 2), split by the adapter.

For a rule extracted by a model: its review state (none: the provision's own text, a statute at corpus.statute_reliability).

For an extracted rule: how its words were found in the provision.

The unit as the corpus numbers it (75, 12.a, priloga-1).

The version cited (npb-5, 2).

Text in English as a plain string. Only when it quotes or reproduces a source written in another language: an object with the original in text, its ISO 639-1 code in lang and an English rendering in en.

What became of it, for a machine ([NormDecision]); status says it in words.

meets, violates, within tolerance, needs comparison, …

Text in English as a plain string. Only when it quotes or reproduces a source written in another language: an object with the original in text, its ISO 639-1 code in lang and an English rendering in en.

The corpus's own id of the structured rule it was read from (Luma's rule id).

decisive, overridden (a more specific norm decides), not in force, draft, not applicable.

The zone it was considered for, and the site's share in it.

Why the norm applies or may not: written by the engine, so English is its original (shown in other languages through the vocabulary).

The day the norms were selected as in force.

What the verdict rests on, in words: the version assumed in force, the membership share when below 1, the extraction state when not approved, the norm's condition when the claim does not say whether it holds.

Text in English as a plain string. Only when it quotes or reproduces a source written in another language: an object with the original in text, its ISO 639-1 code in lang and an English rendering in en.

A reader's decision on each condition that shaped the verdict: the deciding norm's, and those of norms set aside because their condition does not hold for the project, each with its basis ([crate::project]).

With status conditional: each way of the norm the claim does not address but could meet, verbatim with its article and paragraph, then the ways not open for the claimed value (marked not applicable).

The stated and the measured value differ beyond the measurement's tolerance.

The requirement verbatim (the original, with its English rendering).

A planner's review of the deciding norm: state, who, when, the note.

The deciding norm rule.

The deciding norm's state: statute (the provision's own text), unvalidated, read, approved, corrected.

The status in a reviewer's words ("meets only if: (a) …; (b) …").

Where the compared value comes from: stated (the applicant's own, in the documentation) or measurement (supplied, with its provenance).

The measurement compared, with its tolerance and provenance.

With corrected: the ways the provision is met, as the planner corrected them.

The norm rules (as a job's norm.rule or considered[].rule names them), or the caller's structured rule ids.

The planner's note; journaled.

summary (a line, in English), summaries (by language: en, sl) and what the kind adds (clashes, targets).

rebase or promotion.

Only when first-principles reasoning is asked for (the request says so under "First-principles reasoning"): the predicate keys of the claims whose truth depends on what a concept means (fascist, democracy, genocide, perjury, British as a status), exactly as used in claims. Otherwise leave it out.

The day the text was read (ISO date), set when it was parsed and kept with the stored parse: claims without a time are judged as of it.

The language the text is written in (ISO 639-1).

Reference resolutions and sense choices made, in plain words ("they = apples"; "apple read as the fruit, not the company").

The place the whole text is about, when it has one; claims without their own place inherit it.

When the request is not in English: a human reading of each claim's predicate key, in English and in the request language.