What would a JSONL or Parquet schema for receipts and allocations look like?

A release can be delivered as linked tables. This is one schema a buyer can ask for, to be adjusted in the pilot.

`receipts.jsonl` - `receipt_id` (`RCPT-7731`), `account_id`, `payer_id`, `amount`, `currency`, `value_date`, `channel` (bank credit, card, cheque), `payer_reference`, `reference_quality` (exact, abbreviated, wrong, missing)

`obligations.jsonl` - `obligation_id` (`INV-3301`), `account_id`, `type`, `issued_date`, `due_date`, `original_amount`, `balance`, `status` (open, disputed, closed), `terms`

`policies.json` - `rule_id` (`POL-TOL-01`), `applies_to`, `text`, `parameters` such as tolerance or approval limit, `version`

`allocations.jsonl` - `allocation_id`, `receipt_id`, `obligation_id`, `amount`, `rule_id`, `reason_code`, `decided_by_role`, `decided_at`, `reverses` (nullable `allocation_id`)

`events.jsonl` - `event_id`, `case_id`, `timestamp`, `role`, `type` (note, query, reply, escalation, approval), `text`

`cases.jsonl` - `case_id`, `receipt_ids`, `accepted_outcomes` (a list, since several can be valid), `exception_tags`, `split`

Foreign keys: allocations refer to receipts, obligations and rules; events refer to cases; cases refer to receipts. Dates use ISO 8601 and amounts are decimal strings to avoid float drift. The same tables can be supplied as Parquet with typed columns.

Add a `generated: true` marker and a `manifest.json` describing how each field was produced, so reviewers can separate rule-derived values from free text and look at grounding.