Skip to content
Bud's Wiki

System One / Decision Models

TL;DR

Section titled “TL;DR”

A “System One Model” or “Decision Model” is a class of model that gives up open-ended token generation and instead emits one or more typed, structured decisions (categorical verdict + probability, scored option sets, structured JSON slots) in a single forward pass, scored at fixed positions in the architecture rather than autoregressively decoded. Across filed papers, the design surface is splitting on backbone choice (decoder-only VLM vs bidirectional encoder vs post-trained general LLM), training objective (RLCD with strictly-proper scoring rules vs standard SFT), and the hardware / cost / latency envelope the models are deployed in. The pole was named by TypeSafe AI’s closed Jev launch and now has at least four distinct open-weights or hosted entries, benchmarked across structured-decision, LLM-judge, and even per-layer diffusion-sparsity routing applications.

Key claims

Section titled “Key claims”

Recent contributions

Section titled “Recent contributions”
  • [2026-10-27] Introducing Clef: our open-source decision models, and new RL fine-tuning platform: Cloudflare releases Clef (Qwen3.8-27B backbone + vision encoder, 64k context) and Clef-flash (Qwen3.5-9B) as open-weights Apache 2.0 decision models — second filed multimodal entry after Open d1, first with a published RLCD training recipe (label-smoothed cross-entropy + Brier calibration → RLCD with partial-credit ordinal reward + reference penalty) and a managed RL fine-tuning product on Workers AI; leads 9/10 Jev Decision Index benchmarks and beats Jev on 3/4 TypeSafe workflow evals, with 2.2s vs 4.7s latency vs gpt-oss-120b on an internal domain-classification workflow.
  • [2026-10-09] Dmytro Hrybov — OpenAI Decisions API (GPT-6 Luna) as a real-time robotics policy in MuJoCo from camera views: First filed robotics-control application of a decision model — Dmytro Hrybov drives OpenAI’s Decisions API (post-trained GPT-6 Luna) as a real-time MuJoCo tabletop-manipulation policy from camera views, factoring each tick into two API calls (next-action decision → arm/gripper decision). Exercises the exact Luna multimodal-input property Ishaan flagged in the launch filing as the pole’s structural advantage over text-only Jev; best run grasps and transports a cube but fails final placement, matching the LLM-as-controller precision-placement failure mode on VLA Models.
  • [2026-10-07] Open d1: Edge decision models for text, vision, and audio: Liquid AI Open d1 adds the first open-weights multimodal entry to the class — d1-3B (text+vision) on LFM2.5-VL-3B tops Decision Index v0.2.1 under 10B, d1-omni-600M (text+image or text+audio) is the first System One model handling three modalities, with 8 ms single-question latency on RTX 4090 and demonstrated hardware-in-the-loop Jetson + Isaac Sim deployment.
  • [2026-09-29] OpenAI Decisions API — real-time typed classification/routing decisions from GPT-6 Luna: OpenAI launches the Decisions API (post-trained GPT-6 Luna, limited preview) — the biggest lab’s entry into the pole, with multimodal input and no latency/cost/benchmark numbers disclosed at launch. First instance of post-training specialization of a general base rather than end-to-end model-class redesign.
  • [2026-09-24] JEV-as-a-Judge: Accept When Confident, Escalate When Unsure: First independent benchmark of Jev — matches GPT-6 Astra within 3 pts on preference/factuality at 0.36% of the fee, drops 9–20 pts on reasoning. A frozen 0.9-confidence cascade keeps 99% of GPT-6 accuracy at 57% of cost. First outside empirical corroboration of the launch’s cost/latency claims, with a sharpened reasoning-ceiling finding.
  • [2026-09-20] Kamimoto — Jev-gated dynamic sparsity for MiniMax H3 video generation (41.7% wall-clock reduction on RTX 4070): Kamimoto repurposes Jev one abstraction layer down — instead of agent tool selection, Jev picks per-layer attention sparsity (1/3/5/10%) inside a MiniMax H3 diffusion transformer at ~30 ms per call. Generalizes the class beyond typed-answer-to-structured-question into any per-step scheduling decision inside a generation pipeline.
  • [2026-09-20] Laya-MLX — Native Apple Silicon port of Laya typed decision models: Independent MLX port of Laya to Apple Silicon — first filed sub-15 ms fully-local instance of the class with no PyTorch/Transformers dependency. Snake demo drives Laya as a real-time decision agent at 75.4 moves/s with cycle-safety intervention.
  • [2026-09-19] Laya — Non-Autoregressive System One Decision Model (open-source RLCD): Convai Innovations ships the first open-weights entry — 421M ModernBERT-large + option-marker head trained with RLCD (strictly proper scoring rule as the reward, REINFORCE with group-mean baseline). Self-reported +3.9 pp accuracy, 3× better ECE, 7.8× lower latency than Jev at 4 GB VRAM. Fine-tuning caveat: base/multilingual checkpoints regress below majority-class baseline on the typed-decisions benchmark.
  • [2026-09-16] Qwen-2.5-1B-RLCD open-sourced as tongue-in-cheek response to System One / Jev launch: Qwen-2.5-1B-RLCD open-sourced as a tongue-in-cheek rebuttal to the Jev launch — the abstract-argument version of the Laya position: any decoder-only LLM can be prompted or lightly fine-tuned to score typed JSON slots in parallel.
  • [2026-09-15] Introducing System One Models & Jev: TypeSafe AI (ex-OpenAI Diogo Almeida) launches Jev and defines the model class. Advertises 20–200× speedup and 40–400× cost reduction vs LLMs on TypeSafe’s own workflow evals; no technical report, no public-benchmark numbers at launch, slack sharer flagged the demo as overhyped. Introduces RLCD as a distinct post-training method and the “mathematical guarantee of 0% type errors” framing.

Open questions

Section titled “Open questions”
  • Architecture vs. supervision: how much of a decision model’s measured advantage over a general LLM is the single-forward-pass + schema-head architecture, and how much is simply fine-tuning on the eval’s own training distribution? Laya’s typed-decisions win only holds for the task-fine-tuned checkpoint (Laya — Non-Autoregressive System One Decision Model (open-source RLCD) §typed-decisions, all three checkpoints); a controlled ablation holding backbone fixed and swapping the training recipe would resolve this. d1-3B’s Decision Index score vs its LFM2.5-VL-3B backbone prompted zero-shot into structured output is the obvious experiment.
  • Does the reasoning ceiling move: Jev-as-a-Judge locates the gap on JudgeBench / RM-Bench hard (JEV-as-a-Judge: Accept When Confident, Escalate When Unsure §Where it fails). Can RLCD-trained decision models close this gap by scaling data/architecture, or is this a fundamental ceiling of the “no reasoning tokens” design? The post-trained GPT-6 Luna behind OpenAI’s Decisions API is a potential counter-example if benchmarks ever appear.
  • Audio decisions: d1-omni-600M handles text+audio but Liquid explicitly calls audio-decision benchmarks an open problem (Open d1: Edge decision models for text, vision, and audio §Benchmarks). The field needs a Decision-Index equivalent that scores audio-conditioned decisions directly rather than demo-only validation.
  • Cardinality limits: Jev hits a 255-option cap on categorical decisions and falls back to a 2-stage pipeline for larger choice sets (Introducing System One Models & Jev §Wikiracing / Nuance); Laya’s choice questions degrade past ~20 options (Laya — Non-Autoregressive System One Decision Model (open-source RLCD) §Limits). Is there a decision-model architecture that scales to open-vocabulary decisions (think dense classification over 10K+ labels) without reverting to autoregressive decoding?
  • Confidence calibration transfer: JEV-as-a-Judge found the 0.9 escalation threshold did not transfer across fallback models (JEV-as-a-Judge: Accept When Confident, Escalate When Unsure §Practical checklist). Does every decision model need per-deployment threshold calibration, or can a universal calibration mechanism (e.g. post-hoc temperature + Platt scaling conditioned on the decoder’s uncertainty estimate) be baked in at training time?
  • Composition and chaining: can decision models be chained — one decision model’s output as another’s input, or as a routing signal for a general LLM — the way SmallVLM → LargeVLM pipelines compose on the generative side? Kamimoto’s per-layer-sparsity use of Jev (Kamimoto — Jev-gated dynamic sparsity for MiniMax H3 video generation (41.7% wall-clock reduction on RTX 4070)) is one data point; no systematic study.

Papers

Section titled “Papers”

2026-10

Section titled “2026-10”

2026-09

Section titled “2026-09”