Fused

PASS consensus from 2 PASS, 1 PARTIAL.

The Workbench path is complete, but a caching page shows udf.fused.io instead of the current canvas-token udf.ai form. One page carries the whole path from decorator to live URL, and the canvas-token step is stated identically on first-udf-basics, tokens-endpoints, and run-udfs-as-api, giving the concrete call form https://udf.ai/fc_<CANVAS_TOKEN>/my_udf.json. From zero to running UDF to HTTP endpoint: the quickstart gives a complete @fused.udf example, and the 'UDFs as API' page documents the exact URL pattern (https://udf.ai/fc_<CANVAS_TOKEN>/udf_name.json). The only extra hop is learning how to share a canvas to get the token, which is linked.

PARTIAL consensus from 2 PARTIAL, 1 FAIL.

The docs state 120s and ~10GB, but scaling guidance uses ~4GB, and free-tier credits have no numeric quota. Realtime timeout (120s) and parallel ceilings (max_workers default 32, max 1000; worker_concurrency default 1; max_retry default 2) are stated precisely, but the realtime RAM ceiling is given as ~10GB on the realtime and fused-run pages and as ~4GB twice on the scaling-out page, and no plan or account quota carries a number anywhere in the docs. Realtime limits (120s, ~10GB) and parallel max_workers default (32) are documented with exact values. But free tier 'daily compute credits' is unspecified, dedicated-instance specs are absent, and plan quotas link to an external pricing page rather than being documented inline.

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The playbook documents status, errors, and failed-only retry, but it never shows a code correction before retrying. A dedicated Debugging playbook page walks the entire recovery loop against a deliberately flaky example UDF: start small, pool.status(), any_failed(), first_error(), errors(), times(), results(return_exceptions=True), then pool.retry() followed by pool.wait(). Inspecting errors is comprehensively documented (pool.status(), first_error(), errors()). The max_retry parameter enables automatic retries. But a manual 'retry only the failed subset' pattern is not explicitly shown in the docs.

PASS consensus from 3 PASS.

One page covers Python 3.10+, installation, browser authentication, credential storage, bearer-token retrieval, and a callable UDF test. A single page covers Python 3.10+, pip install "fused[all]", the NotebookCredentials browser flow, the credential file location ~/.fused/credentials, logout, Bearer token retrieval, and a runnable quick test. The full install-authenticate-test cycle is on one page: pip install, NotebookCredentials for browser auth, access_token for programmatic use, and a quick test UDF. Every step is documented with code examples.

PASS consensus from 2 PASS, 1 PARTIAL.

The official SDK's direct-call example matches Workbench's automatic execution and the separate canvas-token HTTP path. The SDK and the minimum example (fused.load("my_udf") then calling the returned object) are unambiguous, but reaching them requires reconciling live and legacy APIs that sit side by side: the API Reference index lists fused.run and fused.submit under UDF Execution, the JobPool class page still attributes pools to the deprecated fused.submit(), the current Writing UDFs page instructs the reader using fused.run, and the FAQ names a run_batch entry point that appears nowhere in the SDK reference. The 'fused' package on PyPI is the SDK. A minimum runnable example appears in the install guide, and the API Reference provides full class/method docs for Udf, JobPool, and all decorators. All pages use the same @fused.udf pattern consistently.