Import saved before/after replies with your expected fields. Run Decision Audit and Reply Compare, plus optional conversation or bill checks, then download one local review bundle.
No model calls, upload or storage after closing. This reviews supplied records, not platform reach, execution, permission or measured improvement. Expected fields and model metadata are not authenticated. The output includes all inputs; inspect before sharing.
No file loaded.
Plain JSON object, schema paired-review-input-v1, cases array (1 to 100). Start with the two-case sample, replace its replies and expected values, then save that JSON as your input file. For bill/conversation-only cases, omit expected, before and after entirely: no paired score is calculated. If any paired field is present, all three are required. Each case needs unique case_id (outer whitespace trimmed; case-sensitive, so A and a are distinct), either a paired review (nonempty expected object with no null/empty-string values, before and after replies) or optional bill/conversation inputs. Optional row-top-level owner_approved/source_verified must be booleans and are supplied restrictions only. Optional metadata records your provenance labels, not proof. Optional conversation_log uses Assistant Audit's log format; conversation policy uses the existing default approved labels openai/anthropic/groq and blocked families qwen/deepseek; this is a review configuration, not provider approval or authorship verification. Optional bill has csv and options (tax, fees, discount, total, currency). Input limit 2 MB. No automatic answer key or provider connection.
Counts use exact supplied fields and existing core limits. Extra output fields and risks are separate from a pass. Invalid/unparseable replies remain in the bundle; comparison may be unavailable. Decimal values use JavaScript floating point; Bill Check uses exact cents. No significance test or learning claim.
Open an input file or try the sample.
Pass means the supplied expected fields matched; risk flags are separate and can occur on a passing row. Counts do not prove improvement or safety. The export repeats some values in differences/findings and may be larger than the input; repeated full input payloads are removed.
No review yet.