Ch 46 deliverable A: Transformation language prior art

Ch 46 pre-read — transformation language prior art (2026-08-27)

Produced as a pre-read for the Ch 46 architect window. Recovery note: the surveying agent returned this content in its final message but never wrote the file; reconstructed verbatim from that message by the orchestrator the same day. Treat the tables as the agent’s findings, not the orchestrator’s.

Language × the five measured gaps

Legend: ✓ direct/clean · △ possible but awkward, extension-dependent, or mismatched · ✗ absent

Candidate	Literal key containing .	Per-item transform inside one value	Substring extraction	Row subset	Join two row collections
JSONata 2.2.x	✓ `a.b` or $lookup($, "a.b")	✓ $map, $filter, $split	✓ $substringAfter, $match, $replace	✓ $[predicate] or $filter	△ Variables plus $$; indexed join via $merge/lookup, but no dedicated join operator
jq 1.8	✓ .["a.b"]	✓ map(...)	✓ slices, split, capture, sub	✓ map(select(...))	✓ Native INDEX and JOIN
RML / YARRRML	✓ CSV references are column names; JSONPath can use bracket notation	△ Nested arrays fit iterators; delimited cell values need FNML functions with weak cross-engine cardinality portability	△ FNML/GREL functions, processor-dependent	△ Conditional subject maps or source-query extensions	✓ Referencing object maps and join conditions, but primarily for linking RDF entities rather than producing enriched rows
Morph-KGC	✓ SQL/JSON quoting	✓ DuckDB list/JSON operations or FNML	✓ DuckDB SQL or FNML	✓ RML View WHERE	✓ DuckDB complex joins/RML joins
SPARQL-Generate	✓ Underlying JSONPath/CSV accessor	✓ Nested ITERATOR functions	✓ SPARQL SUBSTR, STRAFTER, regex	✓ FILTER	△ Multiple sources correlate, but examples recommend staged RDF materialization for performance
dbt / SQL models	✓ Quoted identifiers	△ Clean for database arrays/JSON only; dialect-dependent	✓ SQL string functions	✓ WHERE	✓ SQL joins
JMESPath	✓ Quoted identifiers	✓ Projections and map	✗ No substring, split, or regex primitives	✓ [?predicate]	✗ No cross-collection join or root lookup inside projections

Binding constraints

Candidate	Credible runtimes	Embedded/mobile cost	GRC-analyst learnability	What it gets right about shapes
JSONata	Reference JavaScript plus current C++, Go, Java, .NET, and Python ports; conformance still needs Crosswalker-owned fixtures	79,757 B minified / ~24 KB gzip for 2.2.2; pure JS and mobile-friendly	Medium. Concise paths resemble spreadsheet formulas, but sequence flattening, singleton equivalence, silent missing paths, lambdas, and joins are non-obvious	Treats paths as sequence pipelines: nested collections, predicates, construction, and scalar functions belong to one expression model
jq	Reference C, Go gojq, Rust jaq, Python/native bindings, WebAssembly; implementations intentionally diverge at edges	jq-wasm: 928,824 B WASM + ~130 KB glue; npm package ~4.08 MB unpacked. Mobile-capable but materially heavier, and adds WASM asset/loading complexity	Medium-hard. Excellent compact vocabulary, but streams, cardinality, variables, and pipeline precedence are programmer-oriented	Best vocabulary for all five gaps, especially explicit INDEX/JOIN and predictable map(select(...)) composition
RML/YARRRML	Multiple Java, JavaScript, Python processors; core mappings portable, functions substantially less so	No small conformant browser runtime identified. RocketRML is Node-first; RMLMapper is JVM	Hard. YAML surface helps, but triples maps, term maps, function IRIs, and RDF terminology remain specialist concepts	Correctly separates source access, iterator scope, reference formulation, and cross-source identity joins. Nested versus tabular is explicit rather than inferred
Morph-KGC	Python only as an engine; mappings are RML/YARRRML	Not viable inside Capacitor: Python plus pandas/DuckDB and optional native dependencies	Hard for direct recipe authors; reasonable as an external producer maintained by engineers or agents	Shows a practical hybrid: declarative RML envelope, then SQL views for filtering, complex joins, mixed content, and shape normalization
SPARQL-Generate	Apache Jena/JVM implementation	Not mobile-viable; multi-module JVM stack	Very hard unless the author already knows SPARQL	Makes source iteration first-class and distinguishes scalar binding functions from multi-result iterator functions
dbt	SQL across warehouse adapters; runtime not embeddable in Obsidian	Not an embedded/mobile engine	Medium for SQL-literate analysts, otherwise hard	Strongest precedent for where transformation belongs: source-preserving staging, then explicit intermediate models for joins, fan-out, aggregation, pivoting, grain changes
JMESPath	Broad, mature multi-runtime family	Small pure-JS implementations; excellent mobile fit	Medium-low	Demonstrates the ceiling of a deliberately limited query language: excellent projection/filter portability, but failure on substring and joins would recreate Crosswalker’s current problem

Shortlist put to the architect

Option	Strongest argument	Strongest objection
1. Embed JSONata as the transform expression language	Cheapest coherent answer. Already the schema’s declared expression language, ~24 KB gzip, runs on mobile, credible non-JS ports, expresses all five gaps without adding a Crosswalker language.	Cross-runtime parity is not automatic; joins are idiomatic rather than first-class and can become quadratic; silent-missing and sequence-flattening semantics are dangerous for imports. User-authored expressions need 2.2.1+ security fixes, resource guardrails, and conformance fixtures.
2. Adopt jq semantics, embed jq-wasm	Cleanest five-gap vocabulary including genuine INDEX/JOIN; mature implementations beyond C. Strongest if expressiveness outweighs bundle cost.	~1.06 MB WASM plus glue, separate asset loading, mobile memory overhead, implementation divergence. Duplicates the already-declared JSONata slot; least approachable for GRC analysts.
3. Decline a general embedded language; adopt RML/dbt’s staged model	Preserves a small mobile engine: explicit source iteration plus a separate transform stage, complex joins delegated to an external producer. RML supplies semantics; dbt supplies the stage boundary.	A structured Crosswalker transform block risks becoming the “fourth language” Ch 36 rejected. Leaves spreadsheet-only mobile users unable to perform the hardest joins unless the engine later grows equivalent capability.

Sources

JSONata: paths and predicates · higher-order functions · string functions · programming · implementations · releases and security fixes

jq: 1.8 manual · jq-wasm · gojq compatibility · jaq

RML family: RML spec · YARRRML spec · RML-FNML draft · RML Logical Views draft · Morph-KGC · SPARQL-Generate

Other: dbt SQL models · dbt staging · dbt intermediate models · JMESPath spec