Survey · Preprint

System One Decision ModelsA Survey of the Typed Decision Contract, Its Ecosystem and Its Evidence

Amir Rafe  ·  Subasish Das
Texas State University, San Marcos, Texas, USA
Version of 29 October 2026 · manuscript under review
The typed decision contract. A state and a typed question with declared options go through one forward pass of a decision head, which returns a calibrated probability for every option. A policy threshold on that probability decides whether to act, send the case to review or escalate it.

Abstract

Typed probabilistic decision models, which their first vendor calls System One models, return a probability distribution over a caller-declared answer space and promise that it is calibrated. They form a contract assembled from five research lines. The survey provides a genealogy, taxonomy, formal core, ecosystem census, graded evidence ledger, benchmark summary, pattern catalog with failure atlas, and reporting standard. The evidence supports three regularities. The contract travels, while robustness to option names belongs to each checkpoint. Calibration belongs to the model and the task together. Escalation earns the saving when confidence ranks errors and the price gap is wide.

At a glance

5
research lines that converge on the contract
379
original open implementations in a dated census, each coded by decision head
248
studies graded for risk of bias on five domains, with independent second ratings
3,519
extracted values, each tied to a verbatim excerpt of its source

Three regularities

R1 · Transfer

The contract travels, robustness does not

Open models on public backbones reproduce the typed contract, while robustness to option names belongs to each checkpoint and must be tested per model.

R2 · Calibration

Calibration belongs to model and task together

A calibration result for one schema does not carry over to another, so calibration is audited per schema rather than assumed from one report.

R3 · Escalation

Escalation earns its saving conditionally

Sending uncertain cases to a stronger model saves money only when confidence ranks errors and the price gap to that model is wide.

Figures from the survey

Click a figure to enlarge it.

Genealogy. Five research lines, on an axis compressed before 2017, converge on the hosted release of 15 September 2026. The right panel shows the open and hosted decision models that followed within weeks.
Decision heads in the open census. The five ways an open model reads a distribution over the declared options, with the number of original implementations and families of each type.

Explore the graded studies

Every study in the evidence ledger with its overall risk-of-bias grade. Search by title, author, model or task, filter by grade, and sort by any column.

Study Date Evidence role Task Risk of bias

What the repository holds

Ecosystem census

A card for every open implementation and hosted model up to the cutoff of 28 October 2026, with its pinned revision and decision-head code.

Evidence ledger

Every extracted value with its verbatim excerpt and location, the study notes and the full-text excerpt checks.

Risk-of-bias grades and second rating

Five-domain grades for every study, the blind second-rater workbooks of both sweeps, agreement and the adjudication logs.

Synthesis and scripts

The selection of every pooled number and the scripts that rebuild the ledger, the grades, the synthesis and the tables.

github.com/pozapas/system-one-models-survey, release v1.2.3. A curated index of papers, models, datasets and tools is maintained at awesome-system-one-models.

Citation

@misc{rafe2026systemone,
  title  = {System One Decision Models: A Survey of the Typed Decision Contract, Its Ecosystem and Its Evidence},
  author = {Rafe, Amir and Das, Subasish},
  year   = {2026},
  note   = {Preprint, manuscript under review},
  url    = {https://pozapas.github.io/system-one-models-survey/}
}