What files and environment state could a generated correction release include?

A release could use Parquet for typed tables such as `entries`, `accounts`, `adjustments` and `approval_events`, with JSONL for task requests and message sequences. A schema manifest can define required fields, nullability, decimal precision, currency codes and foreign key relationships. Stable identifiers should connect records across formats.

The environment snapshot can additionally contain role permissions, open work queues, tool-visible documents and pending system events. A `snapshot_id` and a generation seed can identify the starting state. Resetting a run should restore both records and pending events; restoring only balances can leave the next run affected by notifications or jobs from the previous attempt.

MirrorFirm could package tool contracts alongside these assets. Each contract can specify inputs, outputs and documented error responses. A task manifest can link `task_id` to its initial snapshot and evaluator configuration without exposing expected outcomes to the agent.

Buyers should ask for a loading check that validates references and rejects incompatible schema versions before execution. This catches packaging faults separately from agent failures. The package specification should also state whether replay promises identical tool observations or merely the same starting business state, since those are different integration requirements.