Function 11 · controls and data

Can every other number here be trusted?

The general ledger rebuilt from its postings ties to the reported balances in 786 of 792 account months, to the cent. The rest are one break, planted to prove the control works: account 2010, $4,217.36, first in 2026-03 and carried into 6 month ends because nothing later corrects it. Every sub ledger check ties (216 of 216), every data quality defect that was injected was caught at its injected count, and the warehouse passes 210 of 210 tests.

Account months tied to the cent

786 of 792

Postings with no source transaction

Injected defects caught

3 of 3

16 rules run

Planted identifiers found in staged data

30 tables scanned

Reported balance against the sum of postings, by account and month

Each square is one account at one month end: pale is tied, a solid square with a cross is a break.

786 of 792 account months tie to the cent. The one break is savings deposits, account 2010, first appearing in 2026-03 at $4,217.36 and carried unchanged into every later month: one break, located, not 6. Every posting traces to a sub ledger transaction.

Source: generated, Harborline general ledger, 22 accounts by 36 months, seed 20260831, as of 2026-08-31.

Account 2010: reported balance less the sum of its postings

The difference is zero until 2026-03, then holds at -$4,217.36 in every later month, which is what one unposted or misposted entry looks like: a step, not a drift.

Source: generated, Harborline general ledger, 22 accounts by 36 months, seed 20260831, as of 2026-08-31.

Every data quality rule, with its hits

Three defects were injected into the generated data at known counts; each rule that should find one must find exactly that many. The real datasets' own issues are counted by the same rules and left in view rather than cleaned away.

Rule	Table	Severity	Rows checked	Rows hit	Injected	Finding
application income missing	loan applications	flag	36,210	0
loan opened after as of	loans	quarantine	40,288	0
loan balance negative	loan performance	quarantine	854,389	0
transfer without counterparty	deposit entries	flag	8,229,331	1,721	1,721	✓ caught at the injected count
entry after as of	deposit entries	quarantine	8,229,331	0
balance does not roll	deposit balances	flag	2,127,521	212	212	✓ caught at the injected count
posting duplicate	gl postings	quarantine	17,990,726	7,193	7,193	✓ caught at the injected count
posting zero amount	gl postings	quarantine	17,990,726	0
card txn nonpositive	card transactions	quarantine	1,077,220	0
hmda exempt value	hmda lar	flag	609,413	9,864
hmda income not available	hmda lar	flag	609,413	49,974
hmda ltv not available	hmda lar	flag	609,413	171,322
cfpb response in progress	cfpb narratives	flag	46,084	2
uci undocumented code	uci default	flag	30,000	399
baf missing marker	baf	flag	600,000	564,284
bls month unavailable	macro monthly	flag	116	1

Source: mixed, every staged table, as of 2026-08-31.

FDIC institutions checked for the bank's name

27,834

0 collisions, active or inactive

Warehouse models built

210 of 210 tests pass

FDIC peers that stop reporting in the window

Found by rule and left out of the latest quarter's peer group

The break, as the control reports it

Account	First month	Amount	Months carried	Comparison	Finding
2010	2026-03	$4,217.36	6	reported vs postings	the injected break, located

Method and limitations

The ledger and its sub ledgers are generated, so the control is proved on a known answer: the planted break's account, month and amount are in the truth tables, and the test fails if the control reports anything else.
Reconciliation compares balances at month end. An offsetting pair of errors inside one month would tie; the sub ledger checks and the unsourced posting count are there for that.