One workload. An explicit decision gate.

Test one workload on Jylus.

Keep your existing system serving production. Start with four synthetic order-support events, then adapt the same evaluation to an approved copy of your own workload.

AI retrieval and evidence

1. Create an evaluation workspace

   Start free, without a card. Create a scoped API key with events:write and events:read. Set retention deliberately; the records in this authenticated quickstart remain until expiry or deletion. This differs from the single-use anonymous trial.

   Create free workspace

2. Download, inspect and run the quickstart

   Download platform-evaluation.mjs · Plain Node.js, no packages. It uses the same synthetic dataset and requests as the guided demonstration. Read the file before running it; keep your API key on your own machine.

   # Node.js 22 or later. Use a separate evaluation workspace.
   # Save the linked platform-evaluation.mjs file in this directory.
   # macOS/Linux:
   export JYLUS_API_KEY='your-evaluation-workspace-key'
   node platform-evaluation.mjs ai
   
   # PowerShell:
   $env:JYLUS_API_KEY='your-evaluation-workspace-key'
   node .\platform-evaluation.mjs ai

3. Verify the responses

   The script saves the input, ingest receipt, delivery verification and read responses in a new local folder. It checks all four source IDs and the three order observations. It exits with an error if verification fails.

   This path also checks search results, the earlier delayed observation, the corrected in-transit observation and explicit order/shipment relationships.

   It additionally requires complete evidence, source proof IDs and a ready-to-answer decision. Give the saved Context Pack to your own AI integration; model-provider usage is separate. Generating evidence does not prove the model’s eventual answer is correct.

4. Run alongside your existing system

   Map a representative sample’s IDs, timestamps, entity keys and correction references. Replay writes through your existing application or pipeline. Compare identical query inputs, source completeness, field values, permissions, late events, duplicates and p50/p95 latency at expected concurrency. Measure the applicable search/ranking or evidence checks too.

5. Decide before redirecting traffic

   Write down correctness, latency, capacity, recovery and support gates. Keep the original source and a rollback path. Move only the passing workload; watch real errors and latency, and restore its old read path if the agreed gates regress. Delete evaluation records using workspace controls when finished.

A records workload succeeds when accepted records are verified retrievable and the relevant query/search returns useful data. AI evidence is a separate activation path. A demo run is not a confirmed customer workspace.