Protocol 1.0.0

System One Bench

How it works

A System One decision model answers a typed Choice question in one forward pass: every option is scored at once, a softmax turns the scores into a probability distribution, and the top probability is the confidence. A caller can abstain whenever that confidence falls below a threshold τ. This is what the calibration and selective-risk pages measure.

Collar of the chosen option, and the dial needle (confidence)Four collars: world, sports, business, sci_tech, left to rightA Choice question as an assay bench. The ticket in the rack is the state; one shaft drives all four option rods in the same pass, and each collar rises to its softmax probability on the 0–1 backboard. The dial reads the top probability as confidence; slide τ along the rail and the flag drops to abstain whenever confidence < τ. The four tickets are real Strands Decider 2B outputs from decider__ag_news__20261011T151717Z (AG News test items 00000, 00010, 00006 and 00002). 00006 is a confident miss: 0.92 business, gold sci_tech.

Figure drawn with the Anatomy kit (MIT, wheresryan22). Source: site/figure/; truth model: site/figure/model.mjs.