Imagine you're at a multiple-choice exam, but instead of circling an answer and handing in the whole sheet, you just glance at the question and your hand twitches toward one bubble harder than the others. The proctor reads your hand tension, not your written answer. That's what this technique does to an LLM: it never writes a response. It reads the probability distribution the model assigns to a handful of pre-numbered option tokens at a single position, then picks the winner. One forward pass, no chain-of-thought, no JSON generation. The committed claim: an off-the-shelf LLM (GLM-5.3-Flash), with no fine-tuning, can be turned into a typed decision engine that matches the accuracy and latency of purpose-built "System One" models like TypeSafe's Jev. The technique numbers the options in the prompt, prefills the assistant turn with "choiceindex:", restricts the vocabulary to option-index tokens via vLLM's allowedtokenids, and reads the logits at that single position using logprobtokenids. Normalized over the allowed tokens, you get a calibrated probability distribution over options — not just a pick, but a confidence score for each. The ladder comparison is the strongest part of this work. They benchmark against Jev (a commercial, purpose-trained decision model) and Laya (a 421M-parameter lightweight) across 29 public labeled datasets spanning 2–151 options, covering intent routing, sentiment, topic classification, moderation, entailment, QA, legal text, and scanned documents in English and German. GLM-5.3-Flash wins on 10 datasets, Jev wins on 10, and 8 are within one percentage point. The median gap is 0.7pp in Jev's favor — not statistically significant (Wilcoxon p = 0.64). Laya trails both by 13–15pp (p < 0.001). Crucially, both GLM and Jev show up to 3.5% answer variation between identical runs at temperature 0, which the authors attribute to batching and floating-point non-determinism — an honest disclosure that sets the noise floor. Architecturally, this is not a new model. It's a prompting and inference-time technique layered on top of any autoregressive LLM served by vLLM. The key engineering dependencies are: (1) vLLM's continuefinalmessage and addgenerationprompt: false flags, which allow prefilling the assistant turn; (2) logprobtokenids, which returns exact log probabilities for specified tokens (toplogprobs is insufficient because formatting tokens dominate before vocabulary restriction); (3) allowedtokenids for vocabulary masking. The method is model-agnostic in principle, though tokenizer quirks matter — GLM-5.3-Flash has a single token for "12", for instance, and the library resolves token IDs dynamically via the /completions echo endpoint. The integrity regime is better than most applied ML posts. The benchmark is public and reproducible (GitHub repo provided). All 29 datasets are public and labeled. The same state, option names, option order, and instructions are sent to all three systems. They ran each dataset twice to establish a noise floor and used a Wilcoxon signed-rank test for statistical significance. They did not tune their prompt on the benchmark datasets. The main weakness: this is a company blog post from Edgeless Systems (which sells Privatemode), not a peer-reviewed paper. Jev and Laya were run with default settings — whether those defaults are optimal for these specific tasks is an open question. The bonus capability — typed decisions on images — is undersold. Because vLLM's chat completions endpoint accepts images alongside text, the same single-forward-pass technique works on multimodal inputs (e.g., scanned invoices), which Jev cannot do. This is not benchmarked as rigorously as the text results, but the playground demonstrates it live. The obvious next experiments not run: scaling to more than 151 options (where does the single-token-index scheme break?), calibration analysis (are the normalized logit probabilities well-calibrated as actual confidence scores?), and comparison against structured-output APIs like OpenAI's or Anthropic's constrained decoding, which also avoid full JSON generation. The honest read: calibration analysis would either validate or undermine the confidence-score selling point, and they likely haven't done the work yet. The option-count ceiling is probably a tokenizer engineering problem they'll address in a future release.
