Typesafe Evaluate

TypeSafe evaluations

Use typesafe_evaluate for semantic decisions over explicit supplied state. It returns typed decisions, not generated explanations or permission to act. If the tool is unavailable, report that; do not substitute shell/HTTP calls or ask for a credential in chat. Calls send the supplied data to TypeSafe and may incur charges.

Supply {state, questions}; see mixed request example. State and descriptions accept text, JSON objects/arrays, or null. Each question sees the same state and is independent: it cannot read another answer in the batch. Put meaning in instructions and criteria, not just the question ID.

• Choice: choose one of 2–255 named alternatives. Include a no-match alternative if appropriate. Preserve the selected label, full distribution, and confidence.
• Score: rate against 2–10 ordered descriptions. The answer is the fractional probability-weighted zero-based position, not a normalized score or integer.
• Noul: probability of yes, from 0 to 1, with optional criteria.true and criteria.false descriptions. This is not intensity and has no confidence field.

Only send necessary evidence authorized for sharing. Never include credentials. Do not silently truncate evidence or split competing Choice options to fit a request. The plugin bounds JSON to 4 MiB, 262144 nodes, and depth 64; vendor token limits are separate. A model override is optional; pin a version when comparing runs.

Reported probabilities may be rounded and need not sum exactly to one. Preserve the vendor-selected label and score; normalization or selecting the largest reported probability is an explicit consumer policy, not an adapter repair.

Batch independent questions. Use a later call when its state or options depend on an earlier result. Validate judgment quality on representative cases before using probabilities as thresholds; confidence is not the probability that a whole workflow is correct. Preserve uncertainty and verify deterministic rules, current evidence, and authorization separately. A service/validation error is not a negative judgment; do not retry automatically or fabricate an answer.

The optional tool is distinct from the background decision-provider API. Enabling the plugin/tool does not itself select a decision model or schedule consumer work. Native consumers use the global or per-agent decisionModel role. Plugin model configuration is only the optional tool’s default.