Inside Open System One Models: How Single-Pass Readout Heads Produce Calibrated Probabilities
Technical breakdowns of open decision architectures explain how non-causal readout layers replace token-by-token decoding with one parallel forward pass.
While TypeSafe AI keeps Jev’s weights proprietary, the surge of Apache 2.0 open-weight System One models in October and early November has given ML engineers a clear look at how non-generative decision architectures work under the hood.
Skipping autoregressive decoding altogether
In a standard LLM, classifying an input into one of fifty options or answering ten structured questions requires decoding dozens or hundreds of JSON tokens sequentially, each conditioned on the previous token. Open decision models—ranging from compact 0.6B–4B models to 27B and 105B mixture-of-experts readouts—short-circuit that loop. They encode the shared input state once, attend across the declared question and option definitions, and feed final representations into a lightweight scoring head that emits normalized probabilities for every option in a single forward pass.
Why calibration takes dedicated training
Simply reading raw token log-probabilities off an off-the-shelf chat model rarely yields trustworthy confidence scores: RLHF-trained models are notoriously overconfident and sensitive to the order in which choices are listed. Open System One training recipes counter this by combining label-smoothed cross-entropy with proper scoring rules such as Brier loss, alongside option-order permutation during training—mirroring the calibration goals TypeSafe AI described in its RLCD launch post.
The result is an architecture whose compute cost scales primarily with input prefill rather than output length, explaining why vendors across the market can waive output-token fees altogether.