DigitalOcean has published the most rigorous piece on LLM structured output I have read, and it includes a result against its own product. In Structured Output Reliability at Scale: When JSON Schema Validity Breaks Down Under Concurrency, Manikandan Kurup starts from a failure every team shipping agents will recognize: “Structured output is 100% valid in development. Then production starts showing a creeping invalid-output rate nobody can reproduce, because development tests one request at a time and production doesn’t.” The study runs four schema families, a flat classification, a nested tool call, an enum-heavy triage decision, and an array-of-objects extraction, on a single H100 GPU Droplet with vLLM and xgrammar in strict mode, ramping concurrency from 1 to 400 sustained, and then runs the same harness against DigitalOcean Serverless Inference.
The numbers are the tutorial’s own, measured on disclosed infrastructure with the harness scripts named, which is as close to independent as a vendor post gets. On the pinned self-hosted stack, schema validity among returned responses stayed at 1.00 from concurrency 1 to 400, even with 270 requests queued. On the managed endpoint, a black box whose “backend, engine version and batching policy are provider-internal,” validity fell from 1.00 to roughly 0.85 under load, because generations ended early. That is DigitalOcean reporting that its own managed service truncates under pressure where a stack you can inspect did not, and it deserves credit for publishing it. The rest of the findings travel. Schema validity is not correctness: a strict response returned every required field with the right type and still reported an invoice total of 38.58 for line items worth 34.65. XGrammar silently ignored three untyped constraints another backend enforced, “so ‘strict’ quietly isn’t.” And retries must branch on the failure category, because “repeating the same prompt at the same budget reproduces truncation and deterministic semantic errors while adding load.” The prescribed order of checks is termination, parse, schema, then semantic.
The catalog can place half of this story and misses the other half. The Digital Ocean provider page lists 37 API pages, and the self-hosted arm lives on them: the Digital Ocean Droplets API and the Digital Ocean Droplet Actions API are how you provision and manage the GPU Droplet the study ran on. The managed arm does not. Serverless Inference, the endpoint that slipped to 0.85, has no API page on the record, and neither does the rest of DigitalOcean’s AI platform: managed agents, knowledge bases, the model library. That is our gap, and it is filed. A catalog that cannot show the inference surface cannot tell a buyer which of the two arms they are signing up for.
The Kin Score is 45.6, developing band. Access clarity carries it at 77.6, discoverability at 61.7, contract quality at 55.7. Developer ergonomics is 16.7 and contract governance is 0.0. The Agent Readiness score is 52.2, agent-ready, with the MCP server and the whole identity layer lit, delegated identity, protected resource metadata, and dynamic client registration. Error semantics and idempotency are unlit, and those are exactly the two things the tutorial’s retry design depends on: you cannot branch on a failure category the contract does not describe, and you cannot safely resend what the API does not promise to deduplicate. DigitalOcean has written the best guide to handling structured-output failure. Its own API contract does not yet describe its failures.