JSON-LD 1.2 and Beyond

Extensions for machine learning data exchange.

§ 1The mismatch

JSON-LD was designed for human-curated, open-world linked data. Modern AI pipelines need the opposite: closed-world inputs at runtime, strict schemas, probabilistic outputs, and sub-millisecond budgets. Three motivating gaps drive the work:

• Epistemic uncertainty — low-confidence model outputs are treated as established facts.
• Provenance & temporal validity — required for auditing AI-generated assertions.
• Security — remote context loading enables DNS poisoning, SSRF, and parser differentials.

§ 2Twelve identified gaps

§ 3Proposed extensions

Backward-compatible additions that downgrade cleanly to JSON-LD 1.1:

• A1 @integrity — SRI-style hashes for contexts
• A2/A3 Context allowlists and resource limits
• E1 @confidence, @source, @method — provenance vocabulary
• E2 @validFrom / @validUntil — temporal extensions
• F1 @vector container — first-class embeddings

§ 4Validating experiments

Four experiments using a healthcare wearable (posture classification at the edge):

1. Backward compatibility — extended docs parse cleanly in stock JSON-LD 1.1 processors after stripping.
2. Security — context-poisoning attempts blocked by @integrity at 100%.
3. Payload efficiency — annotated payloads stay within wearable bandwidth budgets.
4. Uncertainty-aware processing — 28% reduction in low-confidence false positives reaching downstream models.

§ 5Try it

The companion pages in this site exercise each extension live. Or run the package yourself:

pip install jsonld-ex          # python
npm install @jsonld-ex/core    # javascript

§ 6Contact

Muntaser Syed · msyed2011@my.fit.edu · Florida Institute of Technology