shifty

A formalism-first SHACL validation and SHACL-AF inference engine written in Rust, grounded in the algebraic treatment of Common Foundations for SHACL, ShEx, and PG-Schema (arXiv:2502.01295). Available as a command-line tool and as Python bindings (pyshifty).

Features

• Full SHACL Core validation — node and property shapes, all standard constraint components
• SHACL-AF inference — forward-chaining sh:rule evaluation (Triple Rules, SPARQL Construct Rules) to a fixed point, with stratification analysis for recursive rulesets
• Algebraic IR — shapes are lowered to a path algebra (π) and shape grammar (φ) before evaluation; the same IR drives both validation and inference
• Native SPARQL execution — a subset of sh:sparql constraints and SPARQL Construct rules runs directly over an indexed dataset without a full SPARQL engine, with automatic fallback to Spareval for unsupported constructs
• Multi-layer pipeline — parsing → algebraic lowering → normalization/CSE → physical planning → execution; each layer is independently inspectable
• pyshifty-compatible Python API — validate() returns (conforms, report_graph, results_text) matching pyshifty's interface

Installation

CLI

cargo install --path crates/shifty-cli

Or build from source:

cargo build --release -p shifty-cli
# binary at target/release/shifty

Python

pip install pyshifty

The package installs as pyshifty but is imported as shifty:

import shifty

To build from source (requires Rust and maturin):

cd python
pip install maturin
maturin develop

CLI usage

Validate

shifty validate --shapes shapes.ttl --data data.ttl

conforms: false
violations: 1
  <http://example.org/bob>  [target: ∃ rdf:type .⊤]
      - (ex:name) 123 → expected datatype xsd:string

Emit a W3C sh:ValidationReport in Turtle:

shifty validate --shapes shapes.ttl --data data.ttl --report

JSON output:

shifty validate --shapes shapes.ttl --data data.ttl --format json

Graph mode controls which triples are visible to path traversal and SPARQL evaluation:

# default: focus nodes from data; paths/SPARQL use data ∪ shapes
shifty validate --shapes shapes.ttl --data data.ttl --graph-mode union

# focus nodes and evaluation use data only
shifty validate --shapes shapes.ttl --data data.ttl --graph-mode data

# focus nodes and evaluation both use data ∪ shapes
shifty validate --shapes shapes.ttl --data data.ttl --graph-mode union-all

# validate only selected named shapes as top-level entry points
shifty validate \
  --shapes shapes.ttl \
  --data data.ttl \
  --shape-name http://example.org/PersonShape

--shape-name is repeatable and has the alias --entry-shape. Selected shapes are used only as entry points; referenced helper shapes are still evaluated normally.

Infer

Run SHACL-AF rules to a fixed point, then print the derived triples:

shifty infer --shapes rules.ttl --data data.ttl

inferred 3 triple(s):
  <http://example.org/r1> <http://example.org/area> "6"^^<http://www.w3.org/2001/XMLSchema#integer>
  ...

Inspect

Inspect how a shapes graph looks at each stage of the pipeline:

# Raw triples after parsing
shifty inspect --stage rdf shapes.ttl

# Lowered algebraic IR (φ/π notation)
shifty inspect --stage algebra shapes.ttl

# After normalization and common-subexpression elimination
shifty inspect --stage normalized shapes.ttl

# Stratification analysis (recursion detection)
shifty inspect --stage strata shapes.ttl

# Physical plan: focus sources + cost-ordered shape checks
shifty inspect --stage plan shapes.ttl

# SPARQL constraint capability: which queries run native vs. Spareval
shifty inspect --stage capability shapes.ttl

All stages support --format text (default), --format json; the algebra and normalized stages also support --format dot for Graphviz output.

Shapes files and data files may be local paths or HTTP/HTTPS URLs. Both --shapes and --data are repeatable to merge multiple files.

Python usage

import shifty

Validate (pyshifty-compatible)

shapes = """
@prefix sh:  <http://www.w3.org/ns/shifty#> .
@prefix ex:  <http://example.org/> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

ex:PersonShape a sh:NodeShape ;
    sh:targetClass ex:Person ;
    sh:property [
        sh:path ex:name ;
        sh:minCount 1 ;
        sh:datatype xsd:string ;
    ] ;
    sh:property [
        sh:path ex:age ;
        sh:maxCount 1 ;
        sh:datatype xsd:integer ;
    ] .
"""

data = """
@prefix ex: <http://example.org/> .

ex:Alice a ex:Person ; ex:name "Alice" ; ex:age 30 .
ex:Bob   a ex:Person .
"""

conforms, report_graph, results_text = shifty.validate(data, shapes)
# conforms → False
# report_graph → rdflib.Graph with sh:ValidationReport
# results_text → human-readable summary

Graph inputs can be a string, bytes, pathlib.Path, or rdflib.Graph. If shacl_graph is omitted or passed as None, shapes are expected to be embedded in the data graph. Do not pass an empty rdflib.Graph() for embedded shapes; that is treated as an explicit empty shapes graph.

Where shapes are read from. Pass a single graph (omit shacl_graph or pass None) and shifty reads both the shape definitions and the data from that one graph. Pass a separate shapes graph and the schema is compiled only from it — SHACL vocabulary that happens to sit in the data graph is ignored, never turned into constraints. This keeps validation predictable and matches SHACL's separation of the shapes graph from the data graph. To validate against shapes that live in the data graph, union that graph into the shacl_graph argument yourself (it accepts a list, unioned before evaluation); shifty will not read shapes from the data side automatically.

Any data or shapes argument also accepts a list (or tuple) of the above; the members are unioned (merged at the RDF triple level, the same way the CLI's repeatable --shapes / --data merge) before being passed to the engine. A single input keeps its native fast path.

# Union two shapes files and two data files before validating.
shifty.validate(["data1.ttl", "data2.ttl"], ["shapes1.ttl", "shapes2.ttl"])

# Works everywhere graphs are accepted:
validator = shifty.PreparedValidator(["shapes1.ttl", "shapes2.ttl"])
validator.validate([rdflib.Graph(), extra_data])

To validate a shapes graph against itself, pass it once. The embedded path parses and plans one graph without constructing separate data and shapes graphs:

result = shifty.validate_algebra("shapes.ttl", infer=False)
conforms, report_graph, results_text = shifty.validate("shapes.ttl", infer=False)

pathlib.Path inputs are parsed directly in Rust. rdflib.Graph inputs are transferred as N-Triples.

Reuse prepared shapes

For multiple data graphs using the same shapes, cache parsing, normalization, and planning with PreparedValidator:

validator = shifty.PreparedValidator(shapes)

result = validator.validate_algebra(data, infer=False)
conforms, report_graph, results_text = validator.validate(data)

Validate with structured result

validate_algebra returns an AlgebraResult with typed Violation objects instead of an RDF report graph:

result = shifty.validate_algebra(data, shapes)
print(result.conforms)        # False
print(result.results_text)    # human-readable summary (built and cached on first access)
for v in result.violations:
    print(v.focus_node)       # IRI of the failing focus node
    print(v.shape_name)       # IRI of the violated shape, or None
    for r in v.reasons:
        print(r.message)          # engine-generated failure description (always set)
        print(r.author_message)   # the shape's sh:message if it declared one, else None
        print(r.path)             # path that was checked, if applicable
        print(r.value)            # the offending value node

        # Prefer the author's message when present, fall back to the generated one:
        print(r.author_message or r.message)

Set infer=False when validation should not first run embedded SHACL-AF rules to a fixed point.

Infer

Run SHACL-AF rules to a fixed point:

rules = """
@prefix sh: <http://www.w3.org/ns/shifty#> .
@prefix ex: <http://example.org/> .

ex:RectangleShape a sh:NodeShape ;
    sh:targetClass ex:Rectangle ;
    sh:rule [
        a sh:TripleRule ;
        sh:subject sh:this ;
        sh:predicate ex:area ;
        sh:object [ sh:path ex:width ] ;
    ] .
"""

data = """
@prefix ex: <http://example.org/> .
ex:r1 a ex:Rectangle ; ex:width 3 ; ex:height 2 .
"""

result = shifty.infer(data, rules)
print(result.inferred_count)    # number of newly derived triples
g = result.graph()              # rdflib.Graph with original + inferred data

If rules are embedded in the data graph, omit the second argument or pass None:

result = shifty.infer(combined_data_and_rules)
result = shifty.infer(combined_data_and_rules, None)

Passing rdflib.Graph() as the second argument means “run with an explicit empty rules graph,” so no embedded rules will be parsed.

graph_mode

validate() and validate_algebra() accept a graph_mode keyword argument:

shifty.validate(data, shapes, graph_mode="union")      # default
shifty.validate(data, shapes, graph_mode="data")
shifty.validate(data, shapes, graph_mode="union-all")

When shacl_graph is omitted, all three modes are equivalent because focus discovery and evaluation use the same embedded graph. infer() does not accept graph_mode.

shape_names

validate() and validate_algebra() accept shape_names=[...] to validate only the named SHACL shapes in that list as top-level entry points:

result = shifty.validate_algebra(
    data,
    shapes,
    shape_names=["http://example.org/PersonShape"],
)

Only target-bearing statements owned by those named shapes are used as entry points. Any helper shapes referenced from them through sh:node, sh:property, qualified value shapes, boolean shape expressions, and similar dependencies are still evaluated normally. Shape names may be passed as bare IRIs or wrapped in angle brackets.

File inputs

import pathlib

conforms, report, text = shifty.validate(
    pathlib.Path("data.ttl"),
    pathlib.Path("shapes.ttl"),
)

Property witnesses (conformance bindings)

validate/validate_algebra report violations. PreparedValidator.witnesses() is their inverse: for every focus node that conforms to a target/profile node shape, it returns the values each sh:property shape's sh:path resolved to. Useful when a SHACL profile doubles as an extraction schema — e.g. disambiguating several same-typed sensors on a piece of equipment via sh:qualifiedValueShape.

(This is a different concept from the RepairSession "witnesses" below, which records why a focus node fails a statement for repair synthesis. This one only reports observed, passing bindings — no repair suggestions.)

shapes = """
@prefix sh:  <http://www.w3.org/ns/shacl#> .
@prefix zea: <http://example.org/zea#> .
@prefix ex:  <http://example.org/> .

ex:VavProfile a sh:NodeShape ;
    sh:targetClass ex:Vav ;
    sh:property [
        zea:role ex:OutsideAirTempRole ;
        sh:path ex:hasPoint ;
        sh:qualifiedValueShape [ sh:hasValue ex:oat ] ;
        sh:qualifiedMinCount 1 ;
        sh:qualifiedMaxCount 1 ;
    ] ;
    sh:property [
        zea:role ex:ReturnAirTempRole ;
        sh:path ex:hasPoint ;
        sh:qualifiedValueShape [ sh:hasValue ex:rat ] ;
        sh:qualifiedMinCount 1 ;
        sh:qualifiedMaxCount 1 ;
    ] .
ex:OutsideAirTempRole zea:roleName "outsideAirTemp" .
ex:ReturnAirTempRole zea:roleName "returnAirTemp" .
"""
data = """
@prefix ex: <http://example.org/> .
ex:vav1 a ex:Vav ; ex:hasPoint ex:oat, ex:rat, ex:sat, ex:mat .
"""

validator = shifty.PreparedValidator(shapes)
for w in validator.witnesses(data, key_path="zea:role/zea:roleName"):
    print(w.focus, w.key, w.values)
# <http://example.org/vav1> outsideAirTemp ['<http://example.org/oat>']
# <http://example.org/vav1> returnAirTemp  ['<http://example.org/rat>']

key_path is a SPARQL 1.1 property path expression (sequence /, alternation |, inverse ^, and the Kleene forms */+/? are all supported), evaluated from each sh:property shape's own node, over the shapes graph, to produce a stable key. The key here isn't a direct annotation on the property shape — it lives one hop further away, through an intermediate role-descriptor node — which is exactly what a bare predicate lookup couldn't reach but a path can. A direct annotation (zea:roleName "outsideAirTemp" right on the sh:property shape) would just be key_path="zea:roleName"; a descriptor that points at the property shape instead of the other way around would use an inverse hop, key_path="^zea:describes/zea:roleName". Prefixes resolve against the shapes document's declared @prefixes. Property shapes where the path resolves to no value fall back to their own IRI/blank-node id as .key. .values entries are rendered in full (<iri>, "lit", "lit"@lang, "lit"^^<datatype>) so IRI and literal bindings stay distinguishable.

Witnesses (symbolic repair)

RepairSession exposes the witnessing layer: for each statement it reports why a focus node fails (a FocusWitness) or why it holds (a FocusSatisfaction), the structured input to repair synthesis. The session is immutable; it computes and gates but decides nothing.

shapes = """
@prefix sh:  <http://www.w3.org/ns/shacl#> .
@prefix ex:  <http://example.org/> .

ex:PersonShape a sh:NodeShape ;
    sh:targetClass ex:Person ;
    sh:property [ sh:path ex:name ; sh:minCount 1 ] .
"""
data = """
@prefix ex: <http://example.org/> .
ex:carol a ex:Person ; ex:name "Carol" .   # passes ex:PersonShape
ex:dan   a ex:Person .                      # fails: no ex:name
"""

session = shifty.RepairSession(shapes, data, infer=False)

The whole horizon

witnesses() returns one FocusWitness per (focus node, failed statement) across the entire schema. Empty ⟺ the graph conforms.

for w in session.witnesses():
    print(w.focus)        # '<http://example.org/dan>'
    print(w.statement)    # 0 — index into the schema's statements
    print(w.target)       # 'class(<http://example.org/Person>)' — rendered selector

Structured access (strings and objects)

Everything that has a readable string also has a structured, inspectable form, so you can branch and process externally instead of parsing text. w.target is the rendered selector; w.selector is the same thing decomposed:

sel = w.selector
print(sel.kind)      # TargetKind.Class — an enumerated discriminant
print(sel.value)     # '<http://example.org/Person>' — N-Triples, round-trips
print(sel.render)    # 'class(<http://example.org/Person>)' == w.target
print(str(sel))      # same rendered string

if sel.kind == shifty.TargetKind.Class:
    ...              # dispatch on the kind, not on a substring

kind fields are real enums, not bare strings — so the valid set is discoverable at runtime and usable in match/comparisons:

shifty.TargetKind   # Class | SubjectsOf | ObjectsOf | Node | Path | Sparql
shifty.WitnessKind  # Atom | Relational | Closed | CountLow | CountHigh | Not | Opaque
shifty.SatKind      # Atom | Match | Not | Blocked | Coinductive
shifty.ChoiceKind   # Any | Repeat

Scope to one shape

witnesses_for(shape_iri) narrows the horizon to the statements that target a single shape, matched against the schema's shape IRIs (angle brackets optional). It raises ValueError if no shape is named shape_iri.

for w in session.witnesses_for("http://example.org/PersonShape"):
    # flat bag of failing leaves (AND/OR structure dropped)
    for a in w.summary():       # a is a WitnessAtom
        print(a.kind, a.path, a.detail)   # WitnessKind.CountLow <…/name> have 0, need 1
        if a.kind == shifty.WitnessKind.CountLow:
            ...

    print(w.explain())          # indented witness tree:
                                # CountLow along <…/name>: have 0, need 1

    tree = w.repair_tree()      # synthesize the repair space for this violation
    print(tree.is_blocked)      # False — a data repair exists in scope

Passing nodes and the values that satisfied them

satisfactions_for(shape_iri) is the dual: one FocusSatisfaction per passing focus node for that shape. Each records why the node conforms, including the values matched along every checked path — the satisfaction-side mirror of witnesses_for.

for fs in session.satisfactions_for("http://example.org/PersonShape"):
    print(fs.focus)             # '<http://example.org/carol>'
    print(fs.statement)         # 0
    print(fs.target)            # same rendered selector as the witness side
    print(fs.selector.kind)     # TargetKind.Class — same structured selector too

    for a in fs.summary():      # a is a SatAtom
        # one Match leaf per value that satisfied a checked path
        if a.kind == shifty.SatKind.Match:
            print(a.path, a.value)        # <…/name> "Carol"

    print(fs.explain())         # CountHeld: 1 match(es)

witnesses_for and satisfactions_for partition the targeted focus nodes: every node that fails appears in one, every node that holds in the other. For closed, relational (sh:equals/sh:lessThan/…), and opaque-SPARQL constraints a satisfaction leaf is reported as SatKind.Blocked — the node holds, but no enumerable value set is exposed.

Crate structure

Design docs

The docs/ directory contains the full design:

• docs/00-formalism.md — path algebra π, shape grammar φ, selectors, reference semantics
• docs/01-gap-analysis.md — W3C SHACL/SHACL-AF coverage and known gaps
• docs/02-roadmap.md — layered build plan (Layer 0 → 7)

License

BSD-3-Clause