DataCamp: a typed ticket router

Not from the fetch, which was refused, so the source file holds nothing. The entry copy still describes the shape of it, one support ticket judged with Choice, Score and Noul in a single call, split three ways, but the page itself could not be read.

No. The entry copy attributes the 70 to 500 millisecond range, the low per-million input token price and the roughly 68% figure to TypeSafe, not to measurements taken on the tutorial. A number in a walkthrough being vendor-supplied is exactly the case where you rerun it yourself.

The entry copy says the result is split into automate, hand to a human, or fall through based on confidence, but it does not document the thresholds used or any calibration check behind them. Six questions in one parallel call is the documented efficiency claim; the routing policy is not.

The entry copy has the author on preview access and then signups opening, so access widened after he wrote it. Whether his exact SDK call still matches current releases is not something the source could confirm, and the page was never fetched here.

Python 3.10+, pip install typesafe-sdk or uv add typesafe-sdk, and a TypeSafe key exported as TYPESAFE_API_KEY so it never passes through code. New accounts get $5 in credit, which covers the tutorial and then some. TypeSafeClient() is the synchronous client, AsyncTypeSafeClient has the same interface, and both work as context managers.

For the Safari CSV ticket it routes to incident_response at 0.95 confidence, scores goodwill_risk at 1.95 on a four-level patience rubric, and returns is_time_sensitive at 0.97 because a board meeting is named. Usage reports 524 input tokens and 78 output tokens, and output is single digits and free, so the cost lever is how much state you send.

TypeSafe publishes a jaggedness page per model version, linked as docs.typesafe.ai/model-jaggedness/jev-1.13, listing the failure modes it knows about. It is versioned, so read it again on upgrade. The tutorial's method is calibration rather than taste: collect 20 or more real tickets with the answers you would have wanted, run Jev beside current routing without changing behaviour, fix questions before thresholds, then version questions, criteria and thresholds together.