Conformance

prov implements the W3C PROV Data Model and five serialisations of it: PROV-JSON, PROV-XML, PROV-O (RDF), PROV-JSONLD and PROV-N. This page states what the current release conforms to; it was revised for 3.2.3.

The component tables answer two questions: is a PROV-DM concept implemented and how is it written, and does each format preserve it on a round trip. Every cell that is not a plain tick points at a numbered caveat. The values table does the same for literals. “Reading other producers’ output” lists what each reader accepts beyond prov’s own output, with the reason for each allowance. “Corpus evidence” is the test record behind the PROV-N claims. Every cell is checked against the source and the shared test suite, not against the specification text alone; The PROV Data Model gives the background for the six components.

Reading the tables

• A round trip is deserialize(serialize(doc)) == doc for every shared test case. ✓ means it holds; ✓ C1 means it holds except in the case caveat C1 describes; ✗ C2 means the format cannot represent the concept.

• The PROV-N column shows the keyword get_provn() writes. Every row round-trips through PROV-N under the default profile, and under strict when written with strict=True.

• prov enforces no structural constraint at assertion time (#257). Merge-time conformance, PROV-CONSTRAINTS unification, is unified()’s job and is described in Unification and flattening.

Component 1: Entities and Activities

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Entity §5.1.1	ProvEntity	entity()	entity	✓	✓	✓	✓
Activity §5.1.2	ProvActivity	activity()	activity	✓	✓	✓	✓
Generation §5.1.3	ProvGeneration	generation() / wasGeneratedBy()	wasGeneratedBy	✓	✓	✓ C1	✓
Usage §5.1.4	ProvUsage	usage() / used()	used	✓	✓	✓ C1	✓
Communication §5.1.5	ProvCommunication	communication() / wasInformedBy()	wasInformedBy	✓	✓	✓	✓
Start §5.1.6	ProvStart	start() / wasStartedBy()	wasStartedBy	✓	✓	✓ C1	✓
End §5.1.7	ProvEnd	end() / wasEndedBy()	wasEndedBy	✓	✓	✓ C1	✓
Invalidation §5.1.8	ProvInvalidation	invalidation() / wasInvalidatedBy()	wasInvalidatedBy	✓	✓	✓ C1	✓

The factories are ProvBundle methods. ProvEntity also has wasGeneratedBy() and wasInvalidatedBy(), and ProvActivity has used(), wasInformedBy(), wasStartedBy() and wasEndedBy(), as chaining methods with the record itself as subject.

Component 2: Derivations

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Derivation §5.2.1	ProvDerivation	derivation() / wasDerivedFrom()	wasDerivedFrom	✓	✓	✓	✓
Revision §5.2.2	ProvDerivation + prov:Revision type	revision() / wasRevisionOf()	wasDerivedFrom with [prov:type='prov:Revision']	✓	✓	✓	✓
Quotation §5.2.3	ProvDerivation + prov:Quotation type	quotation() / wasQuotedFrom()	wasDerivedFrom with [prov:type='prov:Quotation']	✓	✓	✓	✓
Primary Source §5.2.4	ProvDerivation + prov:PrimarySource type	primary_source() / hadPrimarySource()	wasDerivedFrom with [prov:type='prov:PrimarySource']	✓	✓	✓	✓

The three subtypes share one class and differ only in prov:type, which is how PROV-N expresses them: the grammar has no wasRevisionOf keyword.

Component 3: Agents, Responsibility and Influence

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Agent §5.3.1	ProvAgent	agent()	agent	✓	✓	✓	✓
Person / Organization / SoftwareAgent §5.3.1	prov:type on ProvAgent	none, see “Not implemented”	agent with [prov:type='prov:Person'] and so on	✓	✓	✓	✓
Attribution §5.3.2	ProvAttribution	attribution() / wasAttributedTo()	wasAttributedTo	✓	✓	✓	✓
Association §5.3.3	ProvAssociation	association() / wasAssociatedWith()	wasAssociatedWith	✓	✓	✓	✓
Plan §5.3.3	the entity named by association(plan=...)	none	none; the plan is the association’s third argument	✓	✓	✓	✓
Delegation §5.3.4	ProvDelegation	delegation() / actedOnBehalfOf()	actedOnBehalfOf	✓	✓	✓	✓
Influence §5.3.5	ProvInfluence	influence() / wasInfluencedBy()	wasInfluencedBy	✓	✓	✓	✓

Component 4: Bundles

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Bundle constructor §5.4.1	ProvBundle	ProvDocument.bundle() / add_bundle()	bundle <id> ... endBundle	✓	✓	✓	✓
Bundle as an entity §5.4.2	not implemented, see below

Only a ProvDocument may contain named bundles. A bundle that declares a default namespace different from the one its identifier lives in is written with a prefixed identifier, so the identifier’s IRI survives re-reading.

Component 5: Alternate Entities

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Specialization §5.5.1	ProvSpecialization	specialization() / specializationOf()	specializationOf	✓	✓	✓	✓
Alternate §5.5.2	ProvAlternate	alternate() / alternateOf()	alternateOf	✓	✓	✓	✓
Mention (PROV-Links)	ProvMention, a ProvSpecialization	mention() / mentionOf()	mentionOf, or prov:mentionOf with strict=True	✓	✓	✓	✗ C2

alternate(a, b) is written as the triple a prov:alternateOf b, PROV-DM’s argument order.

Component 6: Collections

Concept (PROV-DM §)	Model class	Factory / alias	PROV-N keyword	JSON	XML	RDF	JSON-LD
Collection §5.6	ProvEntity + prov:Collection type	collection()	entity with [prov:type='prov:Collection']	✓	✓	✓	✓
EmptyCollection §5.6	ProvEntity + prov:EmptyCollection type	none, see “Not implemented”	entity with [prov:type='prov:EmptyCollection']	✓	✓	✓	✓
Membership §5.6	ProvMembership	membership() / hadMember()	hadMember	✓	✓	✓	✓

Values and datatypes (PROV-DM §5.7)

The shared attribute corpus exercises every XSD type below in every format, on prov:label, prov:location, prov:role, prov:type, prov:value and on ordinary attributes.

Value	Written as	Round trip
String, with or without an explicit xsd:string	a string	✓ in every format
Language-tagged string	the tag as given; tags compare case-insensitively	✓ C5 C6
Boolean	xsd:boolean, spelled true and false; the readers also accept 1 and 0	✓
Python int with no datatype	xsd:int when it fits 32 bits, else xsd:long when it fits 64, else xsd:integer; PROV-N writes xsd:int as a bare number	✓
Integer with an asserted datatype (xsd:long, xsd:short, xsd:unsignedByte, …)	the asserted datatype, never collapsed to a narrower one	✓
xsd:decimal	as given; compared in value space, so 10, 10.0 and "10.00" are one value	✓
Python float	xsd:double at full precision	✓
Qualified name as a value	xsd:QName in PROV-JSON, native elsewhere	✓
xsd:dateTime	ISO 8601 text; a naive datetime is written without a zone	✓ C7
xsd:anyURI, xsd:base64Binary and the other XSD types in the corpus	lexical form and datatype as given; datatypes other than xsd:decimal compare lexically, so "1" and "01" differ	✓

Caveats

• C1. Two relations with the same identifier and different prov:time (PROV-O, permanent). PROV-O names a relation by its identifier, so two asserted relations that share one and differ only in time land on one node. Only that construct is affected; decoding third-party RDF shaped this way raises prov.model.ProvException. Minting a synthetic IRI and guessing values on decode were both rejected (#217). unified() reports the conflict.

• C2. No Mention term in PROV-JSONLD (permanent). The submission defines none, so serialising or deserialising a document with a ProvMention raises prov.serializers.provjsonld.ProvJSONLDException (#248).

• C3. PROV-XML escapes attribute names that are not NCNames. A name may start with a digit or contain ' ( ) , : ; [ ] =; each illegal character is written as _xHHHH_, the OpenXML convention, and reversed on read (#289). A third-party name that already looks like an escape is therefore unescaped.

• C4. PROV-N local parts. A character PN_LOCAL cannot express, such as a space, is percent-encoded with a ProvWarning, and the identifier reads back as the encoded IRI, so ex:a b becomes ex:a%20b. A % that starts no escape is written as %25, with the same warning. From 3.2.3 such a name also warns when it is built, and 3.3.0 rejects it (#524). PROV-O rejects an IRI containing a space. A namespace prefix must be a valid PN_PREFIX (a letter first, then name characters or ., not ending in .); prov does not validate or rename one that is not.

• C5. Empty language tag. Literal("hi", langtag="") survives PROV-XML only, as xml:lang="". PROV-N, PROV-JSON and PROV-JSONLD read it back as a plain string and PROV-O raises. From 3.2.3 an empty tag warns when the literal is built, and 3.3.0 rejects it (#472).

• C6. Language tag that is not BCP 47. PROV-N writes an underscore separator as a hyphen, so en_US reads back as en-US; PROV-O raises rdflib’s ValueError. PROV-JSON, PROV-XML and PROV-JSONLD pass the tag through unchanged. From 3.2.3 such a tag warns when the literal is built, and 3.3.0 rejects it (#472).

• C7. Zone offsets that are not whole minutes. Every format writes such an xsd:dateTime in UTC, because the datatype allows no seconds in an offset. The instant is unchanged and the round trip compares equal; the original offset is not preserved.

• C8. RDF’s json-ld syntax is not PROV-JSONLD. rdf_format="json-ld" is rdflib’s generic JSON-LD writer over the PROV-O graph, with every PROV-O limitation. Use format="jsonld" for PROV-JSONLD.

Reading other producers’ output

A round trip covers prov’s own output. Reading what other tools write is a separate claim. Each allowance below states its reason; prov does not adopt another implementation’s departure from the Recommendations without one.

PROV-N profiles

Profile	Accepts
strict	The Recommendation grammar only. Mention is prov:mentionOf, as PROV-Links spells it.
default	The grammar plus the bare mentionOf keyword. The Recommendation has no Mention production, and every existing producer, prov and ProvToolbox included, writes the bare keyword, so refusing it would refuse every file in circulation.
lenient	As default. A statement that fails to parse, or that the model rejects, is skipped with a ProvWarning naming its line and column; a bundle whose identifier duplicates another is skipped whole. Extensibility expressions (prefix:name(...)) are skipped this way.

No profile accepts the keywords person, organization, softwareAgent, collection, emptyCollection, plan, wasRevisionOf, wasQuotedFrom or hadPrimarySource. They are not PROV-N. The grammar expresses subtypes through prov:type, no producer writes them and ProvToolbox’s reader rejects them; they raise ProvNSyntaxError under strict and default and are skipped under lenient. The PROV-Dictionary set-of-pairs literal ({("k1", e1), ...}) is not supported.

The reader also accepts a leading byte order mark, a carriage return as a line end, an all-digit local name as an attribute name or datatype, and a bytearray stream. A prefix declared twice in one scope and a duplicate bundle identifier are errors.

PROV-XML

• A reference element with no prov:ref whose child is itself a reference, the shape ProvToolbox writes for a hadMember member, is resolved with a ProvWarning; the intent is unambiguous. A reference with neither prov:ref nor text raises ProvXMLException.

• <prov:other> elements and attributes the model cannot represent are skipped with a ProvWarning.

• Names escaped as in C3 are unescaped.

PROV-JSONLD

• The decoder accepts the submission’s compacted shape only. It is not a JSON-LD processor: no expansion, no remote contexts.

• ProvToolbox’s prov:-prefixed spellings of the type and special terms are accepted.

• An array-valued entity on a Membership statement, which the submission (section 4.18) allows and ProvToolbox always writes, decodes to one hadMember per member, each with the statement’s @id and attributes; an empty array raises ProvJSONLDException.

• Attribute values follow the 2026-10-14 revision of the PROV-JSONLD schema (openprov/prov-jsonld#7). A node reference, such as {"@id": "ex:shed"}, is a qualified name under any attribute. A bare string is a qualified name under type, location and role, which the context coerces to @id, and a string under any other attribute. A node reference whose IRI no declared namespace compacts becomes an Identifier. A node reference with a key other than @id, or with an empty @id, raises ProvJSONLDException. An xsd:QName or prov:QUALIFIED_NAME value object, the form prov writes for a qualified name under any other attribute, is also read as a qualified name, although a JSON-LD processor reads it as a typed literal.

PROV-O

• An attribute predicate under a namespace declared neither in the document nor in the graph gets a minted prefix, with a ProvWarning.

• Two relations sharing an identifier raise, as C1 describes.

• rdf_format selects the RDF syntax to read.

prov.read() detects the format from the content and forwards each reader’s own options (profile for PROV-N; rdf_format, relation_mapper and predicate_mapper for PROV-O; the json.load hooks for PROV-JSON and PROV-JSONLD). An option a format does not accept raises TypeError naming the format.

Corpus evidence

The PROV-N conformance corpus runs at every test run.

Set	Profile	Files	Result
Every example in the PROV-N and PROV-DM Recommendations	strict	127	100 parse and round-trip through the writer. 27 are excluded because the Recommendation’s prose presents them as PROV-N but the grammar rejects them: fragments missing a paired argument, bare literals, ellipses, a copy-paste dateTime error, and the PROV-Dictionary and extensibility syntax.
ProvToolbox’s hand-written PROV-N test documents	default	9	4 parse. 5 are excluded: bare local names with no default namespace, the PROV-Dictionary set-of-pairs literal, and a bare IRI as a datatype.
PROV-N that ProvToolbox’s writer produced from the shared test corpus, compared with the PROV-JSON fixture of the same name	default	388	229 equal. 159 differ for reasons in ProvToolbox’s 2023 generator, not in prov: 80 embed a wall-clock timestamp, 79 carry a datatype list that no longer matches the fixture.

The corpus README names every excluded or differing file with its reason.

Not implemented

• A bundle as a first-class entity of type prov:Bundle (PROV-DM §5.4.2), so the provenance of a bundle itself cannot be asserted (#261).

• Factories for the agent subtypes and for EmptyCollection; add the prov:type through other_attributes (#260).

• PROV-Dictionary (#129).

• PROV-CONSTRAINTS validation beyond what unified() checks (#62).

Related pages

Unification and flattening covers merge-time conformance, which the tables above do not. The PROV Data Model gives the background for each component, prov.model is the class and method reference, and the how-to pages cover each format: Work with PROV-N, Work with PROV-JSON, Work with PROV-XML, Work with PROV-O (RDF), Work with PROV-JSONLD.