Project
Deem
An open-weight System One decision model family: typed, calibrated decisions (Choice, Score, Noul) over long unstructured state, read from a letter slot in one prefill pass. Shipped as deem-9b-v1 and a CPU-native deem-0.8-v1.
Deem does not generate text. You send it a state (text, JSON, or images) plus typed questions, and it returns calibrated probability distributions in a single forward pass: Choice over 2 to 255 options, Score on 2 to 10 ordered levels, Noul for yes/no with abstention.
The category is new; TypeSafe’s Jev proved the market and a wave of clones proved the architecture is a commodity. Deem is the open answer:
- Letter-slot cross-encoder readout. Decisions are read from an answer slot in a single prefill pass, no decode phase, no novel heads, on a fully fine-tuned Qwen3.5 backbone. A short-form decision reads at P50 103ms on a low-power edge GPU; a 3,222-token policy at 788ms.
- Ground-truth-first training. Programmatic generators provide labels by construction; teacher distillation is augmentation, never the reference. An honest teacher study (pilot passed, scale failed) drove the debiasing machinery that the whole distillation pipeline now runs.
- Measured honestly. Two failure modes of decision SFT were caught by pre-registered consistency gates, not accuracy gates, and fixed with one-line data changes. External numbers are on JevBench public, never trained on, with single-pass results published next to composites.
The released checkpoints are deem-9b-v1 (JevBench hard 65.8 composite, 68.9
extended reasoning, 54.1 single-pass, judge-tier 16/17) and deem-0.8-v1
(96.3% long-policy hold-out, 362ms on CPU through our AVX-512 Rust runtime). Both
ship Apache-2.0 with the complete Tare benchmark harness and a
/v1/systemone-compatible serving stack.