SafeDecide

A decision model that knows when not to decide.

Ask a typed question about a ticket, a log or an agent step. SafeDecide answers, then marks the answer automate or escalate against an error budget it was certified for. The 140M model runs in this tab: your text never leaves your browser.

Downloads 563 MB once; your browser keeps it for next time, except in private windows. These full-precision weights are the ones the guarantee covers.

Your decision

These are held-out cases from typed-decisions, the benchmark the line was certified on: agent traces, customer service, invoices and security alerts, with JSON states and described options. Inputs unlike these are outside the guarantee, so re-calibrate on your own labelled data before relying on the flag.

Answer type

Decision

WaitingRun a decision to see where it lands.

Error budget

Where the line comes from

The model's confidence is not trusted as-is. The threshold is set by a statistical test on labelled decisions the model never trained on.

  1. Hold out labelled decisions

    600 labelled typed-decisions questions are kept out of training, split by case so questions about the same ticket stay together.

  2. Certify a threshold per budget

    Exact Clopper–Pearson bounds with fixed-sequence testing find the lowest confidence where the error among automated decisions is at most the budget, with 90% confidence.

  3. Act or hand off

    Above the line, act on the answer. Below it, send the decision to a person or a bigger model. On your own traffic, re-run step 2 with your labels.