Install to your first real request

Open Welt, sign in with Google or your verified email/password, and use your workspace's Account page. Google login is accepted; GitHub is currently paused. Create an API key with Allow writes, fits and predictions checked. Copy its secret once; it is not shown again. Read-only keys can browse/download research data but cannot prepare or run your first model.

Use Python 3.11–3.13. Create a fresh environment and install the released wheel, including the existing Parquet reader extra. No Git executable is needed:

python -m venv .venv
# macOS/Linux:
. .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install 'welt-client[parquet] @ https://github.com/ergodic-ai/welt-python/releases/download/v0.6.0/welt_client-0.6.0-py3-none-any.whl'

Download first_request.py from the same reviewed release, then run:

python first_request.py

The script prompts for the key without echoing it or adding it to shell history. It also accepts an existing WELT_API_KEY environment value. Never paste keys into source, notebook cells, command URLs, logs or screenshots. Do not persist an SDK key in browser storage. Browser sessions and SDK keys are separate credentials.

This is a real native request, not a mock or metadata ping. The script downloads exactly one reviewed p10k Iris Parquet asset (7,332 bytes, 150 rows), checks its immutable version/SHA, prints its attribution and uses all four recorded features and the recorded target. Its explicit stratified seed9 recipe uses 120 training rows and 30 query rows. It prepares the qualified tabicl-v2 classification configuration and returns 30 predictions within that task's declared envelope. The recipe is onboarding, not a source-provided split, scientific evaluation or publisher benchmark. Your private fit dataset and predictor persist; the script does not delete them or silently change the full research corpus.

The source asset is asset-d839644b6de36fc5cafdda18c5daeca3, version d839644b6de36fc5cafdda18c5daeca3284a5ce8bb6169aff2fcb3fbb4a12021. Iris is attributed to R. A. Fisher (1936), UCI Iris, DOI10.24432/C56C76, CC BY4.0; the p10k snapshot retains its PMLB delivery and canonical Parquet representation. Dataset license and model rights are separate.

Return to your signed-in workspace. Onboarding checks a successful native prediction attributed by the server to your account/workspace/SDK key. A shared workspace's unrelated success, a metadata GET, queued/failed job or client flag does not qualify. Celebration follows this server acknowledgement, with a reduced-motion alternative. If the operation succeeds but the console is temporarily unavailable, reconnect later; the server evidence remains durable.

If an output file already exists, the script reuses it only after checking the exact size/SHA. It never overwrites a mismatching file: choose a new --output path. If a local wait times out, server work continues; use the supplied job ID and Client.job(id) to reconnect rather than blindly creating another fit. Errors from unavailable workers, capacity or revoked credentials remain actionable and do not produce a success celebration. No latency guarantee is implied.

SDK0.6 contains research methods and the released0.3 interfaces. It excludes the separate unreleased CSV0.4 and batch0.5 candidates. No PyPI release is claimed. See Research datasets, SDK guide and model capabilities. Broader managed signup/recovery and all-model completion gates remain distinct from this bounded walkthrough.