You are given feature definitions and user records as JSON. Each feature flag has a `key`, an `enabled` switch, a `default` Boolean, and an ordered list of rules. Each rule has one or more conditions, a Boolean value, and may have a rollout percentage. A user has a stable `user_id` plus attributes such as country, plan, email, and support tier. Design an evaluation engine with two operations: evaluate one flag for one user, and evaluate every flag for one user. Conditions within a rule use AND. Support exact equality, membership, anchored regular-expression matching, and a documented approximate-string operator. Treat a missing or null attribute as a non-match. A disabled flag returns its default without evaluating rules. If no rule matches, return the default. If several rules match, use the reported true-wins policy rather than sorting by rule name. Percentage rollout must be deterministic for the same user and flag. Treat the bucketing function as an injected dependency that maps `(user_id, flag_key)` to an integer from 0 through 99; a rule with percentage `p` is eligible when its bucket is less than `p`. Explain your data model, control flow, error handling, and test strategy. Also explain how you would distinguish an evaluator defect from a corrupted input record when a supplied expected-decisions file was generated before some user attributes were erased. ### Constraints & Assumptions - Flag and user files fit in memory, but there may be thousands of users and multiple flags. - Regular expressions are evaluated from the beginning of the attribute value. - Rule names have no precedence semantics. - Invalid operators or malformed rule shapes must produce an explicit validation error rather than silently changing a decision. - The evaluator must not mutate the parsed flags or users. ### Clarifying Questions to Ask - What exact behavior should the approximate-string operator provide? - Is true-wins applied only among matched rules, and can a false rule ever override a true rule? - Is the rollout bucket implementation supplied, or must its cross-language hash contract be specified? - Should malformed data fail the whole file, one flag, or one user evaluation? ### What a Strong Answer Covers - Separates parsing and schema validation from pure evaluation logic. - Defines each operator and handles null or missing attributes consistently. - Applies disabled flags, rule matching, rollout eligibility, true-wins, and defaults in the correct order. - Makes rollout decisions stable by injecting or versioning the bucketing contract. - Uses focused tests for conflicting rules, anchored regex behavior, dirty records, repeated evaluation, and boundary percentages. - Diagnoses fixture drift independently from evaluator correctness. ### Follow-up Questions 1. How would you cache decisions without serving stale results after a flag update? 2. How would you add rule priorities while remaining backward compatible with true-wins flags? 3. What telemetry would help explain why a user received a particular decision without exposing sensitive attributes?