A 0.6B Qwen model and 25 lines of Python reproduce the interface TypeSafe sells as Jev

The NobodyWho post is labeled a parody, and the code runs anyway ^[10]. Run it and no text is generated ^[6]. The script loads the Q8_0 GGUF build of Qwen3-0.6B through llama.cpp, with a 512-token context and `logits_all=True` ^[5]. Its prompt lists the options as A, B and C and closes on an empty `<think></think>` block ^[8]. A single `model.eval` call processes the prompt. The script then reads `model.scores` at the last position, the logits for whatever token would come next ^[6]. It keeps three of those logits, one per label, subtracts their log-sum-exp and exponentiates ^[6]. What comes out is a softmax over three choices. Any weight the model put on other tokens is discarded ^[6].

According to the dev.to post, the sample email, a payroll request for a password on a non-company sign-in page, scored 0.031 Legitimate, 0.084 Spam and 0.885 Phishing ^[11]. Those add to exactly 1.000 ^[13]. They have to, because the normalization runs over the three label tokens only ^[6]. The top label gets a confident-looking number even when the model's own first choice of next token lies outside the list ^[6]. The code also keeps only the first token of each label, the `[0]` in the source, so labels must differ in their first token ^[6]. Single letters do.

TypeSafe's calibration claim rests on training. It says Jev was trained with reinforcement learning for calibrated decisions, a signal that rewards honest probabilities as well as right answers ^[15]. On that basis TypeSafe claims a 0.7 from Jev can be treated as a real 70 percent chance ^[15]. The parody pulls a stock checkpoint from the Qwen/Qwen3-0.6B-GGUF repository and trains nothing ^[5]. Checking either model takes many labeled predictions: of all answers scored near 0.7, about seven in ten should be right ^[15]. NobodyWho's example scores a single email ^[11].

TypeSafe's speed and cost ratios are its own numbers, measured against comparable LLMs on classification tasks ^[14]. They transfer to a team whose current path looks like the baseline the dev.to author describes, a full frontier model writing a paragraph of reasoning for every small decision ^[4]. A team already reading logprobs from a small model is not on that baseline, so the ratio does not describe its savings ^[14][4].

In my view the local copy is the right first test for routing and for bulk scoring across thousands of rows ^[7]. Those are the jobs the dev.to author calls absurd uses of a frontier model's cost and latency ^[7].