Structured decision method index¶
简体中文 · English
This index covers the official System One product contract, related paper methods and independent community implementations. Jev's internal architecture is not public; a callable official model, official paper code and a community reproduction are distinct.
Official product and protocol¶
| Method | Institution | Type | Purpose | Integration status |
|---|---|---|---|---|
| TypeSafe Jev | TypeSafe AI | Hosted System One model | Returns structured probabilistic decisions directly for Choice, Score and Noul questions over a state, without natural-language parsing. |
HTTPS provider and typed contract integrated; no actual calls without credentials. |
Paper methods¶
| Topic | Method | Institution | Date | Main method | Local entry point |
|---|---|---|---|---|---|
| Continual adaptation from typed feedback | Jev-DLC | HKUST (Guangzhou) | 2026-10-22 | Feedback can affect only later requests. A decision must be committed before its feedback is revealed. | CPU state component and independent reference; no JEV loaded. Full 27B experiment not reproduced due to cost. |
| Calibration audits for multiple correct options | When Multiple Choices Are Correct | Independent Researcher | 2026-10-21 | Distinguish normalized categorical shares from independently queried binary correctness probabilities; the former are not calibrated binary confidence. The paper studies Jev, Luna and Gemini. | CPU scoring and isolation contracts; no JEV calls. Full API experiment not reproduced due to cost. |
| Multimodal choice reranking | Multimodal Reranking Is a System One Decision | Aalto University | 2026-10-20 | Instead of generating a ranked list or training a separate pointwise reranker, the model scores each candidate by reading the probability of its letter label at the first answer position. The readout follows setwise and FIRST-style methods; the paper tests how this approach works zero-shot on images and visual documents, and whether text-only decision tuning transfers to visual reranking. | Actual SmolVLM-256M forward pass on A100; L1 interface only, full 27B experiment not reproduced. |
| Retrieval judgments and reranking | Jev-as-a-Judge for Brazilian Portuguese Information Retrieval: Relevance Assessment and Reranking | Câmara dos Deputados | 2026-10-16 | On Quati it separately uses an independent Noul probability that each document directly answers the query, gathers every batch, and sorts the results. Candidates do not compete in a single Choice, probabilities are not mixed with BM25 scores, and there is no trained calibrator. | CPU protocol interface and arithmetic checks of the paper's tables; no Jev evaluation or full reproduction. |
| Option-independent decisions | Prism-1 | NTU | 2026-10-15 | A softmax combines independent scores. Softmax alone cannot guarantee this without attention isolation and position reset. | Standalone small-model component with three-seed A100 mechanism validation; full experiment not reproduced. |
| Simulated lookahead | Code Owns the Simulation | CUHK | 2026-10-01 | Executable code simulates each action's successor state; Jev only compares the progress afforded by actions. In addition to the action and next observation, the evaluator receives newly available actions after execution, addressing the difficulty of inferring long-term reachability from semantic action scores alone. | choose_with_lookahead; L1 protocol diagnostic, no actual Jev evaluation. |
| Dynamic candidate classification | GLiClass | Knowledgator | 2025-08-11 | This avoids unstable parsing of generative LLM outputs and separate cross-encoder inference on each text–label pair. The repository retains the defining mechanisms: candidate content affects scores, and one normalization produces a distribution over all candidates. | system-one:bilinear-ce, system-one:rival-ce |
| Confidence training | RLCR | MIT | 2025-07-22 | RLCR predicts both an answer and its confidence and adds a bounded proper scoring rule so honest probabilities maximize expected reward. The local baseline minimizes multiclass Brier over the full candidate distribution. | system-one:rival-brier, system-one:rival-hybrid |
| Decision probability calibration | Calibration-Aware RL | USC / AWS AI Labs | 2026-01-19 | The paper adjusts decision-token probabilities directly to reduce ECE without sacrificing accuracy. The repository reports accuracy alongside Brier score and ECE, fits temperature only on validation, and forbids test-set tuning. | Validation calibration and ECE/Brier reporting |
Main open implementations¶
| Implementation | Approach | Method summary | Integration status |
|---|---|---|---|
| Bespoke Nimble | Trained model | Qwen3.5-9B LoRA reads enum and Boolean logits directly from a flat typed schema. | P0 candidate; cataloged, awaiting unified execution. |
| NanoJev | Trained decision head | Qwen3-0.6B with set-attention heads for Choice, Boolean and Score. | Pinned-checkpoint provider integrated; A100 OOD smoke test passed. |
| Laya | Trained model and calibration | Typed decisions with validation-set temperature calibration. | P0 candidate; cataloged, awaiting unified execution. |
| SemIf | Parallel constrained decoding | Compares option logits directly after shared-state prefill. | P1 engineering candidate, not a training-based reproduction. |
| DiffusionGemma routes | Diffusion model | Non-autoregressive parallel decision experiments. | Separate approach; calibration and throughput still need verification. |
The selection and integration roadmap explains where each approach fits. The community implementation snapshot records all 48 tracker entries without counting indexes, explainers or unreleased announcements as runnable models.