Where the model fails
Design around uncertainty, incomplete categories, and changing data.
ren-zero returns structured judgments. It does not establish truth, prove a policy was followed, or authorize an action. We make no benchmark or accuracy claim on this page.
Incomplete or overlapping options
Probability is relative to the offered answers. Missing categories and overlapping descriptions can produce a confident but unhelpful selection. Add a meaningful fallback, use a companion fit check, and test category boundaries.
Arithmetic, dates, and exact rules
Calculate totals, compare timestamps, and apply eligibility rules in deterministic code. A model judgment should not override a database fact or a policy formula.
Missing evidence and ambiguous intent
A yes/no question can conflate false with unknown. An ordered score can conceal disagreement about the rubric. Ask for the missing facts or route the case to review instead of increasing confidence by changing wording repeatedly.
Long, noisy, or adversarial text
Relevant evidence can be buried in unrelated text. User-controlled text can contain instructions that conflict with your question. Include only necessary evidence, keep policy separate, validate outputs, and test adversarial examples. The server rejects input beyond 2048 formatted tokens; it does not silently shorten it.
Option order and wording
Equivalent descriptions or reordered options can change a judgment. Keep IDs stable and test plausible variants on development examples before fixing your workflow policy.
Calibration and distribution shift
Calibration on one collection does not guarantee calibration on your users, language, or future traffic. The current confidence field is the chosen-answer probability, not an independent correctness check. Reevaluate on representative cases when the model or workflow changes, and monitor human overrides.
Consequential actions
Use explicit authorization, deterministic eligibility checks, and appropriate human review for financial, medical, legal, or similarly consequential decisions. The server’s act field is a routing signal, not professional advice or proof that an action is safe.
Use human escalation when a model cannot resolve a case responsibly. Keep a path for review, correction, and appeal.