Open 0.8B model Jeff answers agents' multiple-choice calls and escalates the unsure ones to a 27B

The post's author argues that sending "which tool?" to a big model costs a prompt, seconds of waiting and a paragraph to parse, with no honest signal of how sure the model was ^[21]. Jeff's route is /v1/systemone on a small FastAPI server ^[10]. Inside, the model scores every option it was given, and a softmax turns those scores into probabilities ^[10]. It writes no text ^[2]. The usage block reports output_tokens: 0 on every call ^[11], the one token count in an agent stack that will never trip a budget alert. Each answer carries a key, a probability and a confidence running from 0, no better than a guess, to 1, certain ^[13]. In the README's client pattern, anything below a threshold goes to the big model ^[14], Qwen3.8-27B in the project's setup ^[3].

The base model takes any option list zero-shot ^[20]. Adapters cover the recurring jobs. According to the README, each one trained in a single epoch on one GPU, in half an hour to four hours ^[5]. All nine together come to roughly 370 MB ^[1].

Setting orders: 2 asks the question again with the options reversed and averages the two runs ^[12]. According to the post, models lean toward options near the top of a list, and the reversal cancels some of that ^[12]. I would turn it on for tool choice. The project's table puts a guard check at about a tenth of a second ^[16].

For the 20x speed on grounding ^[9] to transfer, a team's fallback has to match the configuration in the project's runs ^[7]. The post reports the cascade across eight adapters ^[6], one fewer than the nine that ship ^[4].

The calibration procedure is the part I would copy. Each threshold was fixed on its own calibration rows before any test row was scored ^[8]. Holding those rows apart means no threshold was tuned on the answers it was graded against. On the grounding row, where the small model trails, the post's author wrote that "A results table with a near-loss in it reads like a measurement, not a pitch." ^[22]

Adoption costs show up at concurrency. The server takes its lock without waiting. An overlapping request gets HTTP 529 with a one-second retry hint ^[17]. The loss in issue #7 works out to about 35 percent of the run ^[2]. The Python client retries nothing ^[19]. A guard placed in front of every tool output will hit that lock whenever two checks overlap.

For the five small calls the post traces through one support email ^[1], the evidence supports a cascade with Jeff first and the 27B behind it. It does not show the large model leaving the stack. The author adapted the client example from the README and wrote, "I did not run this" ^[15].