# Persistence acceptance sitting — preparation

Status: **historical fresh-source protocol; superseded as a release prerequisite by the
[owner's existing-corpus regression amendment](persistence-regression-decision-260909.md)**.
The requirements below describe independent acceptance and remain unfulfilled; they must not be
claimed by a reused-corpus run.
Use an isolated Preview tenant and preserve the first measured outcomes. Production data and the
spent extraction-v1 reserve are not an acceptance corpus. See the
[owner decision](persistence-acceptance-decision-260908.md) and
[coverage/criteria](argus-persistence-slice-260909.md#acceptance-matrix-to-freeze-before-opening-the-new-set).

## Required case manifest

Supply 24 records: eight fiscal straight-through, eight fiscal review and eight statements.
For each case record these fields before opening the source for the measured run:

| Field | Required content |
|---|---|
| Case ID and group | Unique ID; fiscal-straight, fiscal-review or statement |
| Source path and SHA256 | Fresh original document; no development copy |
| Independent verifier | Person who checked the expected result against the source |
| Coverage tags | Every condition from the coverage matrix must be represented |
| Initial expected outcome | Exact entry, hold or refusal, with its reason |
| Planned human actions | Only specified corrections/rulings; blank when none |
| Final expected outcome | Exact entry, hold or refusal after those actions |
| Expected identity | Company, counterparty and bank account where relevant |
| Expected register facts | Document key, flow, dates, accounting periods, currencies, exact amounts/signs and ordered lines/movements |
| Expected absence of writes | Which records must not exist for a held/refused case |

Include source declarations/rate policies separately if matching is exercised. Identify the source
of fees, opening/closing balances and control references; do not derive them from matcher output.

## Freeze and run order

1. Check the 8/8/8 counts and coverage, independent verification, source hashes and expected facts.
   Agree the manifest and the existing all-cases-correct criteria before opening the set.
2. Freeze a clean commit, published host hashes, both migration histories, non-secret worker
   configuration and pinned skill/model tuple, tenant catalogue/policy snapshot, this protocol and
   the manifest digest. Keep credentials outside the evidence package.
3. Intake each original once through the normal path. Record the first result, time, intake ID,
   extraction revision, register links and every expected hold before any operator correction.
4. Apply only the manifest's planned human actions, then record final outcomes. Read stored facts
   through the tenant-scoped application path and compare exact identities, counts, amounts, signs,
   dates and periods. Assert no partial or duplicate records and no unauthorized writes.
5. Preserve the first results as immutable evidence. Report failures against the frozen denominator;
   corrections to a key and development reruns remain separately labelled. Do not tune on this set
   and then call the rerun independent acceptance.
6. After acceptance, run the governed deployment WhatIf/Execute path against the exact candidate.
   The existing deployment runner retains its renewal guard, interactive privilege proof and
   module confirmation. A separate C6 ruling is required to activate NP Group native persistence.

The current candidate can be packaged and its dry-run plan prepared while the corpus is pending.
Neither a successful build nor Main/Preview parity substitutes for steps 1–5.